Seminar 1 Michael Samudra Contents The name of the game Problem: Patients are served late 5 Task 9 Improvement strategies Simulation results 12 The Gasthuisberg setting 22 Model 40 Implementation 45 References 51 Literature Review: OR planning 52 Literature Review: Patient scheduling 87 Perspectives 109 Conclusion 111 Appendix 114 1 1. Introduction In most hospitals there are patients who receive surgery later than medically advised. In Belgium’s largest hospital, Gasthuisberg, this is the case for approximately every third patient. At hospitals such as Gasthuisberg this could be solved by simply adding to their capacities, i.e., opening a new OR (operating room) and hiring the necessary additional personnel. Unfortunately, in today’s economic environment this is rarely an option. Serving patients late is a problem because their health condition can potentially quickly worsen which exposes them to an increased health risk. This needs to be avoided, primarily from a humanitarian standpoint, but there is also a hidden cost perspective as a patient in a worsened health state is likely to require larger amounts of resources and thus to cost more money. In order to improve the current situation, the lateness of patients has to be, firstly, quantified and, secondly, the responsible mechanism, the patient scheduling process, has to be understood. We analyzed the percentage of patients being served in time at Gasthuisberg. At the hospital, an elective patient is associated with one of five due time (DT) intervals within which the patient has to be served. We also analyzed the lateness of patients across disciplines, using all data from 2012 and 22 ORs. We tried to understand many of the different aspects related to the scheduling process, which knowledge we then included into a simulation model. We investigated from the data: patient arrival patterns, the relation between estimated and realized surgery durations, rescheduling mechanisms and the allocation patterns of emergencies. 3 We also used the model to investigate the effects of switching from the current scheduling practice of assigning surgeries directly to slots (OR and day) to a two-step procedure, where patients are scheduled to a surgery week first and only in a second step to slots. Our results suggest that in case of the two-step procedure it is very important to allow patients with shorter DTs to break into the already fixed weekly schedule. Additionally, it is important that in the second step of the scheduling procedure, in the within week scheduling, the DT is considered. We conclude that improving the patient scheduling process can help to decrease the amount of patients served too late. As a next step, we try to develop a sound scheduling schema, which allows to further decrease the number of patients served too late. 4 2. The name of the game 2.1. Problem: Patients are served late We are focusing on one particular problem: every third patient receives surgery later than the medically advised time limit. This is a serious problem as the health condition of patients that are not served on time can become worse. This is undesirable, primarily for humanitarian reasons, but it is also important to keep in mind that a patient in a worsened health state is likely to require more hospital resources and therefore ultimately will cost more money. Figure 1: The first axis shows the offset of the number of days patients were served respective to their DT. Most patients (11%) are served 6-7 days before their DT. Whether a patient received surgery too late or is still on time is dependent on the urgency status, that is, the DT category of the patient. The DT is assigned to the patients by the physician in charge. If the patient is an elective patient, then the DT category will define a time interval measured in weeks within which surgery is advised. For instance, it is a Monday and the surgeon decides that a patient requires surgery with urgency category DT 4 (1 week), then it means that the patient should undergo surgery the latest next week Monday. 5 Non-elective Elective GYN Tx ABD CAH NCH ONC RHK THO TRH URO VAT MKA NKO Category 1 2 3 4 5 6 7 8 Meaning Now! Up to 6 hours Today 1 week 1 – 2 weeks 2 – 4 weeks 4 – 8 weeks 8 weeks - Closed set [0, 0] [0, 5h] [0, 23h] [1d, 7d] [8d, 14d] [15d, 28d] [29d, 56d] [57d, 112d] Color GYNAECOLOGIE EN VERLOSKUNDE Gynecology and obstetrics HK ABD TRANSPLANTATIECHIRURGIE Abdominal transplant surgery HK ABDOMINALE HEELKUNDE Abdominal surgery HK CARDIALE HEELKUNDE Cardiac surgery HK NEUROCHIRURGIE Neurosurgery HK ONCOLOGISCHE HEELKUNDE General medical oncological HK PLAST RECONSTR ESTH CHIR Plastic, reconstructive and cosmetic surgery The DT categories can vary from 1-8 where category 1-3 are assigned to non-elective patients and category 4-8 to electives. The first three non-elective categories have to be scheduled within a time interval less than or equal to 1 day. The shortness of this interval leaves no room for planning and thus all three non-elective categories are excluded from the scheduling process. In other words, we only schedule elective patients (DT category 4-8). Nonelectives are nevertheless considered in the simulation model as we in detail model their arrival to the ORs and thus implicitly test for their impact on the elective schedule. HK THORAXHEELKUNDE Thoracic surgery HK TRAUMATOLOGIE Traumatology HK UROLOGIE Urology HK VAATHEELKUNDE Vascular surgery MOND, KAAK, AANGEZ.CHIRURGIE Oral and maxillofacial surgery NKO, GELAATS- EN HALSCHIR Head and neck surgery Another property of the DT is that it is defined as an interval, suggesting that it is best for the patient to get surgery only after a certain reference period. It might seem unreasonable to let patients wait unnecessarily, but it can be the case that the patient or surgeon needs some time to prepare for the surgery. For example, a patient with DT category 5 (1-2 weeks) should be able to wait without complications for 1 week, but it is highly advisable to serve the patient in the 2nd week the latest. From a scheduling algorithmic perspective the end time of the interval is of greater importance. Table 1: http://www.uzleuven.be/en/departments 6 Figure 2: For each discipline a different distribution of DT categories is observed. Interestingly, the distribution or mix of DT categories varies largely across different medical disciplines. Some disciplines contain patients from only 1 or 2 DT categories and some from 3 or 4 DT categories. For Urology, for example, the vast majority of patients is associated to DT 8, i.e., none of the urology patients is urgent. A patient category with a more balanced spectrum is Oral and maxillofacial surgery (MKA). One could wonder if Urology and MKA patients should be handled by the same scheduling mechanism: for urology a first-come, first-served (FCFS) approach might suffice, whereas the same for MKA might result in many late treatments. Surprisingly, looking at the ratio of late treatments, it appears that DT category 4 is served the most efficiently (Figure 3). Most of DT category 4 patients are served within their DT (1 week). This is less true for DT category 5 patients who have to wait relatively long and are in almost half the cases served late, that is, after 2 weeks. DT category 6 patients have variable waiting times, but as the peak between -28 and -21 indicates are often served immediately. Figure 3: The amount of patients served per DT category. DT category 4 patients are in 84.16% of the cases served within their DT. So far the lateness of patients was investigated for the whole inpatient population irrespective of the disciplines. But, as Figure 4 shows, two patients with the same DT but different discipline will have different probabilities of being served in time. Whether patients of a certain discipline get surgery within their DT does not only depend on the discipline’s DT distribution (Figure 2), but also on the way how the patient scheduling process is handled within the discipline. A discipline that performs very well is RHK, where most patients are served in time despite the large number of high urgencies (DT 4). 7 Figure 4: The amount of patients served late for each DT category and discipline. Care has to be taken when interpreting the results: for example, the graph shows that 100% of DT 4 gynecology (GYN) patients are served within their DT, but as shown by Figure 2, there are only a few patients from that DT category and thus it is of little importance that those patients are served in time. 8 2.2. Task We are pursuing two targets. Firstly, we want to develop a scheduling procedure that allows patients to be served in time, i.e., within their DT. Secondly, keeping in mind the first objective, we want to investigate the implications of switching from the current one-step scheduling practice of scheduling patients directly to slots to a new two-step strategy where patients are first scheduled to a week and only in the second step to a slot. Our first goal is to increase the amount of patients that get timely access to the OR, i.e., serve more patients within their DT. This goal can be achieved in three ways. Firstly, the simplest way is to increase the capacity on the supply side, for example, by opening a new OR and hiring the necessary personnel. The problem is that increasing existing OR capacities requires additional financial resources which in the current economic environment are not necessarily available. Moreover, in case the financial resources are available, there are other departments of the hospital that could need the resources as well. The second way how to schedule more patients within their DT is to allow for more flexibility, for example, by use of open scheduling. Open scheduling allows disciplines to occupy any OR at any given time of the weekday, i.e., there is no Master Surgery Schedule (MSS). As a consequence, open scheduling, using the pooling effect, would allow to deal more efficiently with occasional short term capacity shortages affecting single disciplines. Despite some of the benefits of open scheduling it is often avoided in practice, to one part, as it requires much more effort to create a good OR plan and, on the other part, as the OR plan is less repetitive surgeons cannot not select fixed weekly days for consultation or teaching. The third and most cost effective way to improve patient access to the OR is the application of operations research methods. 9 Our second goal is to investigate the implications of switching from the current one-step scheduling strategy, where patients are scheduled directly to slots, to a two-step strategy where patients are first assigned to a future week (to-week scheduling) and, in a second step, to the actual surgery weekday and OR (within-week scheduling). It fits within the future plans of Gasthuisberg to do this switch from the one-step strategy to the two-step strategy. One of the main benefits of changing to the second, two-step strategy is that many of the scheduling decisions that had to be done in advance using the one-step strategy, can be done later in the second step using the two-step strategy. Patients arrive on a Tuesday. If a week is selected that is only partially within the DT then in the within-week scheduling step it has to be made sure that the DT is obeyed. At Gasthuisberg patients are normally told about their surgery date 3 weeks in advance. In other words, in case patients do not fit the schedule for the next 3 weeks they are put on a waiting list. This should normally only happen to patients with DT 6, 7 or 8. Patients with DT 4 and 5 should be scheduled within 1 and 2 weeks respectively and thus should preferable not be put on the waiting list. Aligning the goal to schedule patients within their DT with the goal of switching to the two-step procedure results in a problem: in the two-step procedure, the second step will be largely dependent on the first step. In case in the first step (to-week scheduling) a week is selected that is only partly within the patients DT, then in the second step (within-week scheduling) it has to be made sure that the DT criterion is obeyed. In other words, depending on the week selected in the first step, it is often the case that in the second step only the early days of that week are valid. Surgeons schedule their patients individually, i.e., surgeons are not necessarily coordinating amongst each other or with other departments. This means that a surgeon will find for a patient an appropriate slot according to the criteria that for the individual surgeon are important. This also means that different surgeons, based on their past experience, might be differently efficient at scheduling their patients. Moreover, surgeons at Gasthuisberg as well as generally in Flanders are typically creating their patient schedules by hand [4]. As schedules are not created centrally nor by algorithms, it is likely that different disciplines create patient schedules of varying quality. The current situation can be improved in two ways. Firstly, by learning, i.e., showing surgeons the implications of certain scheduling decisions, and secondly, by algorithmic life support: an algorithm could either suggest several schedules to the surgeons from which they can select the one they prefer, or in case surgeons themselves create their schedules, the 10 algorithm could compute some of the implications of the manually created schedule. In order to be able to use a uniform scheduling framework, we focus our attention on Gasthuisberg’s 22 inpatient ORs. Those inpatient ORs are part of one of 5 kerns each denoted by a letter from A to E. Kerns A-D are located closer to one another, whereas kern E is positioned a little further away. Each kern is generally only used by a restricted set of disciplines. ORT (orthopedics) serves some of its patients in kern C and the rest of it in other facilities. Inpatient department [8] Kerns are fairly similar: kerns A-D contain 4 ORs each and kern E contains 6 ORs. In a kern there is a floor in the middle and in the center there is a room with a computer. It also contains small elevators connecting it with the sterilization department. The floor itself contains the facilities for washing hands. Layout of a kern [8]. Each corner is occupied by one of the four ORs. 11 3. Improvement strategies 3.1. Simulation results We use DT driven scheduling decisions both to test one-step and two-step scheduling strategies. The current practice at Gasthuisberg is to schedule patients in one step, which means that patients are scheduled directly to slots. Using the DT we will test four different one-step scheduling strategies. Following that section, we will introduce the two-step scheduling system and, in that context, test three different basic scheduling mechanisms. 3.1.1. One-step strategy In this section we test four different one-step assignment policies. The first policy is FCFS which does not differentiate between patients. The other three policies are based on the DT. We will show that FCFS outperforms the three DT based policies. At Gasthuisberg the planner (typically the surgeon) schedules patients in one step. Finding an appropriate slot happens manually, that is, the planner finds a slot without algorithmic support. There are several criteria the planner can and often will consider [2]. Firstly, the planner will only consider slots that have enough free capacity to accommodate the new surgery. Unlike for example in the Erasmus Medical enter in the Netherlands where slack time is used to ensure that the probability of overtime is 30% [5], in 12 Gasthuisberg the slots can be fully utilized. Therefore, the sum of expected surgery durations assigned to a slot can sum up to the total capacity of the slot (9 hours). As surgery durations are highly variable, ORs will regularly run into overtime. Slots can also be selected based on the preferences of surgeons. One such preference is to have only one difficult surgery (e.g., hip replacement) in a slot. Similarly, also the number of children can be restricted. This is done as patients before their surgery are not allowed to eat for a certain amount of time, which is more difficult to do for children. It is therefore best to serve one child in a slot first in the morning. Figure 5: Different one-step policies result in remarkably simlar OR related performanc measures. The only exception is undertime, which using FCFS is lower than for all other policies. Performances are in percentages, for example, an averege of 1 hour overtime for a 9 hour slot equals 11.1 % (1/9). One of the most important criteria driving the patient scheduling process is the patient’s urgency status. The urgency status translates to the DT. In case the surgery is not urgent, patients can request and propose surgery dates that would suit them the best, e.g., when they have holiday. The first policy we test is FCFS which assigns surgeries to the earliest slot that has enough free capacity left. A slot has enough free capacity to accommodate an additional surgery if adding the surgery does not yield the slot overbooked, that is, the sum of the estimated surgery durations of all patients served in the slot remains less than the slot’s total capacity. The other three assignment policies are based on the DT. The idea behind the policies is to postpone less urgent surgeries, thereby creating short-term buffer capacity usable by more urgent patients. The first policy assigns patients into the closer end of the patients DT interval. This is similar to the FCFS strategy with the restriction that patients can only be served after a certain reference period. The second policy pushes patients into the end of their DT, that is, the patients are served as late as possible while still within their DT. If there is no such slot, then a date after the patients’ DT is chosen, in which case the patient is served late. The third policy schedules patients as close to the middle of the DT as possible, that is, it minimizes the temporal distance between the selected slot and (DT end – DT start) / 2. 13 As Figure 6 shows, the four policies largely affect patient waiting time. Using FCFS (most left figure) the average waiting time will generally be low for patients from all DT categories. The figure also shows that when applying any DT based policy many patients will experience a different waiting time than what was targeted by the policy, i.e., some dots are between the layers. More results can be found on www.econ.kuleuven.be/healthcare/seminar_1. Figure 6: Waiting time of neurosurgery patients. Each data point is a patient where the color represents the patient’s DT. First axis: day the patient arrived. Second axis: the amount of time the patient waited to get surgery. The colored horizontal lines denote time limits of DT categories 4 to 7. The most right figure shows the historical data for neurosurgery patients from the year 2012. 14 The results on the left suggest that FCFS outperforms all three DT based policies. The least effective strategy is to schedule patients towards the end of their DT. Interesting to note is the fact that the shape of the distribution of the center policy shows the greatest resemblance to the one experienced in reality (Figure 1). The different policies were compared using both patient and OR related performance criteria. Patient related performances are, for example, expected patient waiting time and percentage of patients being scheduled within the DT. Measures were taken jointly for all disciplines together (in order to show how the OR complex as a unit performs), but also for each discipline separately (in order to confirm that the results are valid for all disciplines). OR related performance measures are: utilization, overtime and undertime (Figure 5). FCFS also benefits DT category 4 patients. One might think that FCFS performs well because less urgent patients profit on cost of more urgent ones. This would make sense as FCFS assigns patients to slots regardless of their DT. Figure 49 (Appendix) shows that this is not the case and that FCFS benefits all DT categories. It is surprising that FCFS outperforms the three DT based policies. The main reason is that the DT based polices fail their goal to provide the necessary buffer capacity for DT 4 (and to some extent DT 5). This can be seen in Figure 6. In the figure the large peaks are the bottlenecks, that is, temporal capacity shortages leading to excessive patient waiting times. As can be seen from the figure, in case there is a bottleneck using FCFS, it will also appear in any of the other policies. Difficult to spot on the images is that the DT based policies shift the appearance of the bottlenecks some weeks into the future. To summarize, DT based polices will on the one hand occasionally waste OR capacity as patients are postponed. On the other hand, DT based polices do not provide the necessary buffer capacity required by urgent DT categories. The results do not suggest that the DT cannot be used to guide patient scheduling decisions, but it is probably counterproductive to use them in the rigid way we define them. In the next section, we will incorporate the DT into the scheduling process in a more flexible way. 15 3.1.2. Two-step strategy In Gasthuisberg, as in most hospitals, patients are assigned to slots directly. This means that a patient will be directly planned for surgery to an OR and a date. At Gasthuisberg, they intend to switch from this one-step procedure to a two-step procedure. Instead of assigning patients directly to slots, they will be assigned to a week first. This means that a second step is required, where for all the patients assigned to a given week a suitable OR and weekday is found. The advantage of the two-step procedure comes from the fact that the second part of the procedure, the within week scheduling, can be done just before the start of the week (e.g., Thursday or Friday before the scheduled week). In other words, a large part of scheduling related decisions can be postponed to the second step. We use a simulation model to test the two-step procedure, focusing primarily on the effect it has on the amount of patients scheduled after their respective DT, i.e., too late. Our results suggest that in case of the two-step procedure it is very important to allow higher urgency patients to break into the already fixed weekly schedules. Additionally, it is important that the second step, the within-week scheduling, is guided by the patient’s DT category. Interestingly, we found that reserving a constant amount of capacity for high urgency patients is from a whole system perspective less beneficial. We tested three mechanisms: push, protection levels and DT driven within-week scheduling. Push allows patients to be assigned to a slot that is in the same week as they arrived. Using the push mechanism, for example, a patient that arrives Wednesday can be directly assigned to a slot on Thursday or Friday in the current week. Using the two-step scheduling strategy this would normally not be possible as the first step of the procedure, the to-week step, requires patients to be assigned to a complete week. Earliest such week follows the week in that the patient arrived, i.e., is registered for surgery. 16 If patients are pushed into the current week, then they skip both steps of the scheduling procedure. In reality, if the health condition of the patient allows, it’s better to give them some time to prepare for their surgery. The ‘Push (4, 5)’ mechanism excludes DT 6, 7 and 8 patients from the push mechanism as it only allows to push urgent DT categories 4 and 5. Protection Levels are used to reserve a certain amount of weekly capacity for each DT category. A general patient planning related problem is that the exact time when a future urgent patient arrives is unknown in advance. Consequently, there is a problem if a larger number of DT 4 and 5 patients arrive while all ORs are fully booked for the future days or weeks. In order to mitigate this problem, some OR capacity is reserved for urgent patients, i.e., protected from less urgent ones. Protection levels are not defined for DT category 8 as it contains the least urgent patients. The amount of weekly capacity protected for each DT category corresponds to the amount of weekly capacity needed for that DT category. This is the multiplication of the expected number of weekly patients (Figure 2) and their estimated average surgery duration. Protection levels are not defined for non-electives, but as will be shown in the next section, are implicitly accounted for by the MSS. Learning from the results in the previous section, we decided to generally avoid postponing surgeries. We assume therefore that the results that were valid for the one-step strategy are to some extent also valid for the two-step strategy. Consequently, we keep ourselves mostly to the FCFS ideology. Therefore patients are only postponed if we have reason to believe that there is a shortage of OR capacity which is expected to affect more urgent patients. Patient can only be scheduled for a certain week if the weeks usable OR capacity for the patient’s DT category allows it. This is the case if the free OR capacity minus the sum of protected capacities for more urgent DT categories is larger than the surgery’s estimated duration. Protection levels protect some amount of capacity for each DT class. Consequently, a patient is only postponed if the expected capacity requirements for a more urgent patient classes make that necessary. Protection levels are nested and therefore capacity reserved for a given DT category can always be used by a more urgent DT category (e.g., capacity reserved for patients of DT category 6 can always be used by a DT 4 or 5 patient. No capacity is protected for DT category 8. 17 We define protection levels in two ways: FCFS (PL basic) and FCFS (PL DT). Using the FCFS (PL basic) approach scheduled patients use capacity that is reserved for their DT category. Only if this capacity is insufficient, then unprotected capacity or, in case this is still not enough, capacity protected for less urgent DT categories is used. The FCFS (PL DT) approach differs from the FCFS (PL basic) approach in that patients can even if there is capacity protected for their DT category use unprotected capacity or capacity allocated to less urgent DT categories. This happens in case they are scheduled to a date that is later than their DT. In other words, capacity protected for a DT category can only be claimed by patients that are scheduled within their DT. Consequently the FCFS (PL DT) strategy is more restrictive to low urgency patients than the FCFS (PL basic) strategy. Wfit creates a balanced within week schedule. wfit Lastly, we also tried to get a preliminary understanding of the importance of the within week scheduling step. We tested two heuristics: ‘wfit’ and ‘wfit (DT)’. The former evens out slot occupancy whereas the latter additionally considers the DT of patients. The name wfit is an abbreviation of the phrase worst fit from the memory management literature and is the opposite of best-fit. The aim of the heuristic is to create a schedule with slots that have a similar utilization. Practically this translates into a strategy where a surgery is always assigned to the slot with the most leftover capacity. As a consequence, patients are first always assigned to empty slots. After each slot is occupied by one patient the slot is chosen with the most remaining capacity, that is, the one that was previously assigned the shortest surgery. The ‘wfit’ algorithm starts with the longest surgery and ends with the shortest. Once all surgeries are scheduled, they are randomly permutated slot wise. 18 wfit (DT) Figure 7: The colors represent the last day patients may get surgery without exceeding their DT. An extension of ‘wfit’ is ‘wfit (DT)’ where also the DT of patients is considered. The algorithm starts by assigning patients to one of 6 groups. The 1th group contains those patients that are already late (dark red). The 2nd, 3rd, 4th and 5th group contain patients that have to be served the latest by Monday, Tuesday, Wednesday and Thursday respectively. The 6th group contains those patients who are within their DT even if scheduled for Friday. Within each group surgeries are processed from long to short. The algorithm first schedules group 1. The patients in the 1th group are scheduled for Monday slots first. A patient that is assigned to a slot is removed from its group. At this stage, slots are not allowed to be overbooked, that is, the sum of the estimated surgery durations assigned to a slot has to be smaller than or equal to 9 hours. If all patients in the 1th group are scheduled or, alternatively, all Monday slots are fully booked then the algorithm enters the second stage. In the second stage the remaining patients from the 1th group are merged with the patients from the 2nd group. The patients in the newly formed group are scheduled for both Monday and Tuesday slots. As before, slots are not allowed to be overbooked. The same procedure continues until patients from all 6 groups are included. At this stage patients are scheduled to days Monday until Friday. In case there are any unscheduled patients left the slot overbooking becomes allowed and the remaining patients are distributed using the basic ‘wfit’ algorithm. 19 Figure 8 shows how the three different mechanisms affect patient waiting time. As the ‘Push’ mechanism schedules patients into the week of their arrival, evidently some patients will have a very short waiting time. This can be seen by the missing strip on the bottom of the graphs in rows 2 and 3 of the figure. In the 3rd row the ‘Push’ mechanism is only applied to DT category 4 (yellow) and 5 (green). It is also easy to spot the effect of protection levels (columns 2, 3 and 5, 6) which produces “layered” patient waiting times. It is more difficult to eyeball the effect of DT driven within scheduling. What would be seen on a higher resolution image is that for each line representing a DT category, the points with corresponding color would show a higher density under than above the line. wfit No PL PL DT wfit (DT) PL basic No PL PL DT PL basic No Push Push all Push (4, 5) Figure 8: Neurosurgery. First axis: patient arrival, second axis: patient waiting time. Color of the dots (patients) and lines matches DT. 20 The results suggest that in order to maximize the number of patients served within their DT it is essential to use both ‘Push’ and ‘wfit (DT)’ mechanisms while it does not bring any benefit to use protection levels. The results were obtained using a full factorial design, therefore all 18 possible combinations of different mechanisms were tested. Detailed results can be found on www.econ.kuleuven.be/healthcare/seminar_1. Figure 9 The results for the one-step and two-step methods can be compared against each other as the same environment (e.g., same seed) is used in the simulation model. In the one-step strategy the estimated surgery duration of patients assigned to a slot needs to be smaller than the slot capacity, i.e., patients are not planned into overtime. Therefore, if the capacities provided by the MSS are too little, the system might turn unstable, i.e., patient waiting times start to continuously increase. It is easy to see that this phenomenon is less a problem using the two-step procedure where patients have to fit the cumulative weekly capacity of slots and where in the within-week scheduling step planning into overtime is admissible. Testing how the system reacts if capacities on the supply side are decreased or, equally, on the demand side increased is therefore only possible using the twostep method. It is surprising that applying protection levels does not bring any real benefits (Figure 9). This can be seen as, one the one hand, the average improvement is not significant (below 1%) and, on the other hand, as Figure 9 shows, batch means (dots) are largely overlapping. Positive is that at least patients from DT category 4 seem at least minimally to profit from protection levels, though batch means are also this time strongly overlapping (Appendix, Figure 51). This does not mean that protection levels as a mechanism are inadequate, but it merely points to the fact that they might need to be used differently (e.g., time dynamic protection levels, see Chapter 7). Contrary to protection levels, large improvements can be achieved using DT driven within week scheduling. In Figure 9 right graph, ‘wfit’ scenarios are consistently outperformed by ‘wfit (DT)’ scenarios (triples above them). The improvements can mostly be attributed to better scheduled DT 4 and DT 5 patients (Appendix, Figure 51). Besides DT driven within week scheduling it is also important to apply the ‘Push’ mechanism. This is shown by Figure 9 where the 12 upper scenarios using ‘Push’ outperform the respective lower 6 scenarios. Interestingly, the performance gains can almost entirely be realized by using ‘Push (4, 5)’ and therefore only applying the mechanism to patients with DT 4 and 5. Using ‘Push (4, 5)’, additionally, the number of patients who are pushed into the schedule will reduce from a daily average of 15.6 to 5.4 (Appendix, Figure 52). 21 3.2. The Gasthuisberg setting The people at Gasthuisberg we cooperate with are: Frank Rademakers Nancy Vansteenkiste Pierre Luysmans Christian Lamote Philip Monnens Jo Vandersmissen Guido De Voldere This section contains our understanding of the patient scheduling process at Gasthuisberg. The information to build up this understanding comes from two sources. Firstly, from a very detailed data set covering patient and scheduling related information. Secondly, from the experts at Gasthuisberg who explained to us those aspects of the scheduling mechanism that are not in the data. 3.2.1. Patient arrival pattern The arrival time of a patient is defined as the time when the surgeon in charge makes the decision that a patient needs surgery. For elective patients this will usually happen during consultation, i.e., working hours on weekdays. For non-electives it can happen at any hour of the day. We tried two approaches to understand the patient arrival process. The first approach involves the statistical analysis of inter-arrival times, i.e., the mean length and the variance of time intervals measured between two consecutive arrivals. In the second approach we are only interested in the expected number of arrivals for each weekday. In order to model inter-arrival times we used a Maximum Likelihood Estimator (MLE) and fitted different distributions to the data. We found that the exponential distribution generally fits arrival patterns better than other distributions. Nevertheless as Figure 10 (red continuous line) shows, the results are not satisfactory. One way to improve the quality of the fit is to remove outliers from the data set. Outliers are very large values, i.e., very long inter-arrival times. The reason why some of the inter-arrival times are longer than what you would find in reality has its roots in the property of the data set. The data covers only those arrivals from 2010 to 2012 that received 22 surgery in the year 2012. The patients that received surgery before 2012 or after 2012 are excluded. In other words, our data according to surgery dates covers a 1 year period while the data according to arrival times covers a 2-3 year period. As a result, inter-arrival times will tend to be longer and especially at the end part of the spectrum we will find values that are unrealistically large. Those entries should ideally be identified and removed. Since it is impossible to identify the outliers, we chose to remove the largest 5% (red dashed line). Figure 10: The histogram is based on 3 hour bins where each column shows how often patients arrive within 3, 6, 9 …60 hours. In all elective graphs there is a peak at around 24 hours, representing those patients that arrived on consecutive days. This peak is one reason why it is difficult to achieve a good fit with any standard probability distribution function. 23 As we chose to use an exponential distribution, we can make use of the fact that its parameter can easily be calculated directly (inverse of the mean). Calculating the parameter directly, we are still left with the problem of outliers. The quality of fit improves but is for electives still not satisfactory. As non-elective patients are generally served quickly, the mismatch between the length of the interval covered by surgery dates and the length of the interval covered by arrival dates is much smaller than for electives. Consequently, non-elective inter-arrival times will not contain outliers. The inter-arrival time of non-electives can be modeled with an exponential distribution. The parameter of the distribution can either be approximated with a MLE or calculated directly. In case a MLE is used, it is advisable to remove outliers from the dataset first. An alternative way to understand patient arrival patterns are arrival rates. Using arrival rates, only the total number of patients arriving on a given weekday (Mon-Sun) is important whereas the exact arrival hour is not. As elective patients are registered during consulting sessions, it is not a surprise that most patients arrive on weekdays (Figure 11). A benefit of using arrival rates in comparison to inter-arrival times is that there are no outliers, i.e., the mismatch between the intervals covered by the surgery dates and by arrival dates does not influence the collected statistics. Customer arrival rates are generally modeled using a Poisson distribution, that is, it is assumed that the mean arrival rate equals the variance. Investigating the patient arrival rate for different weekdays, we noticed that for most disciplines the mean arrival rate is significantly lower than the variance and thus they are not equal (Figure 12). As the arrival rate does not follow a Poisson distribution, it can be modeled, for example, by a discrete distribution. So far we analyzed the arrival patterns for each discipline separately, but we did not distinguish between urgency categories. Unfortunately, this is difficult to do. As Figure 2 shows, the DT categories are for most disciplines not well balanced, i.e., at least some of the DT categories will have a very few patients. Consequently, the sample size for those discipline - DT category pairs will be small. This is even the case for ABD which is the discipline with the largest total sample size and with a fairly well balanced distribution of DT categories (Appendix, Figure 53). 24 Figure 11: The height of the column represents the amount of patients expected to arrive on that given weekday. For example, twice as many gynecology (GYN) patients arrive on a usual Monday than on a Wednesday. Figure 12: Each column represents one day of the year 2012 (52 columns for each weekday). The height of the columns represents the number of ABD arrivals for the given weekday. The average arrival rate is the highest for Wednesday (9) but the maximum number of patients arrived on a Friday (21). On the graph there are some zero entries which means that there was no consulting on that day, e.g., because it was a holiday. The statistics for other disciplines can be found on www.econ.kuleuven.be/healthcare/seminar_1. 25 3.2.2. Master Surgery Schedule Week 1 / Week 2 Room MON TUE WED THU FRI Kern A A1 A2 A3 A4 URO NKO GYN GYN URO NKO GYN GYN URO NKO GYN URO URO NKO GYN URO URO NKO GYN NKO/GYN Kern B B1 B2 B3 B4 ABD ABD Tx ONC ABD ABD Tx Tx/ABD ONC ABD Tx ONC ABD ABD Tx ABD ABD ABD Tx ABD Kern C C1 C2 C3 C4 THO TRH TRH TRH THO THO TRH ORT/[] THO THO TRH TRH THO THO TRH []/ORT THO TRH TRH ONC NCH RHK MKA NCH NCH RHK RHK/MKA NCH RHK MKA/[] NCH NCH RHK CAH CAH CAH CAH VAT VAT CAH CAH VAT CAH CAH VAT VAT CAH CAH VAT VAT CAH CAH Kern D D1 NCH D2 RHK D3 D4 RHK Kern E E1 E2 E3 E4 E5 E6 VAT CAH CAH CAH Table 2: MSS used at Gasthuisberg. NCH An easy but admittedly oversimplified way of structuring the OR planning process is to distinguish between the following three levels: strategic, tactical and operational. At the strategic level, a certain amount of OR capacity is allocated to each discipline (e.g., urology). This is the same as creating the case mix as the hospital effectively decides on the number of future patients it wants to serve from each discipline. On the tactical level, the MSS is created (Table 2), i.e., a 1 or 2 week cyclic plan is constructed that allocates to each OR slot (weekday and OR) a corresponding discipline. On the operational level, patients are assigned to OR slots. A more detailed schema is described in [3] and in Chapter 5. The basic OR to discipline allocation schema is defined by the MSS template (Table 2). In reality, smaller changes can occur to the weekly MSS. This is, even though often on a short notice, known in advance and can happen if a given discipline either has not enough patients to fill up a slot or if a needed surgeon is unavailable (e.g. goes to a conference). Consequently, a slot can remain empty or it can be occupied by a discipline other than predefined in the template. The MSS including those prearranged changes forms the planned MSS. It can happen that some of the prearranged changes are not logged. Unlogged changes are difficult to track. One way of doing it is to label slots based on the discipline of the first elective patient that is scheduled and served. This way, each slot that is open can be associated to a discipline. A slot is regarded to be open if any elective patient gets surgery within the slot time (7:45-16:45), i.e., starts or ends within slot time or started before and ended after the slot. This relabeled MSS forms the realized MSS. The weekly capacities allocated to disciplines by the MSS template, the planned MSS and the realized MSS are visualized in Figure 13. Interesting to note is the fact that for most disciplines the mismatch between the MSS template and the planned MSS is only minimal (Figure 13). There are some exceptions. One such discipline is 26 Patients need less OR capacity if estimated surgery durations are taken as basis. More details on surgery durations can be found in the next section. thoracic surgery (THO) where relative to the template +7.2 hours are planned but -7.6 hours seems to get realized. Other disciplines where less is used than planned are: transplantation (TX), cardiology (CAH) and vascular surgery (VAT). The transplantation OR is generally regarded to be the OR that is used to accommodate non-electives, thus it is no surprise that it is often empty or actually occupied by non-electives. For cardiology another phenomenon plays a role: the length of surgeries is usually long and difficult to predict, thus some extra buffer is allocated to them. From Figure 13 it seems that both the template and the planned MSS is over-capacitated and therefore, at least in theory, less OR capacity would be enough to serve the patients. In other words, the hospital plans in some amount of buffer capacity. There are two reasons for this. Firstly, it is used to compensate for uncertainty in surgery durations. This is especially important in a setting where surgery durations are systematically underestimated. A second reason are the emergency patients [1]: more details on this topic can be found in Section 3.2.4. Figure 13: The amount of weekly OR capacity required by patients is shown in color. The MSS template, the planned MSS and the realized MSS are usually similar. 27 3.2.3. Surgery duration As Figure 13 shows, surgery durations are generally underestimated. A more detail picture of the link between estimated surgery duration, realized surgery durations and the DT of patients will be given in this chapter. Realized and estimated surgery durations are dependent variables. The surgery duration of a patient represents the time between the moments the patient is rolled into the OR and the time when the patient leaves the OR. Cleaning time is not included whereas setup time might or might not be included, depending on whether the patient is already present in the OR. Normally patient specific setup time is included. The estimated surgery duration is suggested to the planner (e.g., surgeon) based on the mean of realized surgery durations of previous similar OR sessions. The planner can accept this value or overrule it. Surgery durations are for most disciplines systematically underestimated (Figure 15), that is, estimated surgery durations are generally shorter than realized surgery durations. If for a surgery the estimate would turn out to be correct, in the figure the surgery dot would be on the horizontal black line. On the figure we see that most of the dots are above that black line, which means that surgery durations generally take longer than previously estimated. This happens because surgeons try to plan as many patients into their own slots as they can without exceeding the capacity of their slots. Underestimating patients’ surgery durations gives them a tool to legally overfill their slots. The same phenomenon can be seen in Figure 13. Disciplines that only slightly underestimate their surgery durations are head and neck surgery (NKO), vascular surgery (VAT) and cardiology (CAH). We assume that surgery durations are independent of the DT. More generally, we assume that patients from different DT categories only differ in their urgency status and the rest of their attributes are statistically the same. This is a strong assumption but a necessary one. In Figure 2 we see that for most disciplines the mix of DT categories is unequal. Therefore, if we divide patients according to their DT categories, our sample size would become for some of the DT categories very small. In Figure 14 we see an indication that the assumption that surgery durations are independent of their DT categories is mostly true. The same can be seen for neurosurgery in Figure 16, where the marginal distributions of surgery durations are similar for all DT categories. One of the exceptions is abdominal surgery (Figure 17), where we see that DT 4 patients (in yellow) have a significantly lower mean surgery duration than patients of other DT categories. 28 Figure 14: The surgery durations are independent of the DT. Notable exceptions are ABD (DT 4 are shorter), THO (DT 4 are shorter) and CAH (DT 6 are longer). Figure 15: Surgery durations are systematically underestimated (many of the dots are above the black line). 29 Figure 16: The distribution of both estimated and realized surgery durations follows a bell curve. For some DT categories the shape takes the form of a multimodal distribution (distinct peaks), as interventions of the same type usually take similarly long. Results for other disciplines can be found on www.econ.kuleuven.be/healthcare/seminar_1. Figure 17: Estimated surgeries durations are suggested to the planner based on similar previous OR sessions. As the figure shows, estimated surgeries are usually a multiple of 30 minutes (stripes in the figure). The figure also shows that DT 4 (yellow) abdominal surgeries (ABD) are generally shorter. 30 3.2.4. Non-elective allocation schema Every week over 60 non-electives require surgery at Gasthuisberg. In this section, we will show how those non-electives are allocated to ORs. It is more likely that a non-elective patient arrives during slot time than during the weekend or the night (from 22:00 – 06:00). Slot time in a broad sense is the time period when OR slots can theoretically be open, that is, weekdays from 06:00 to 22:00. The night is defined as the time interval between 22:00-06:00. If non-elective patients were to arrive with the same probability around the clock, then theoretically 2/7 (~29%) of all non-elective patients should arrive on the weekend. For most disciplines, we see that this number is in reality significantly lower (Figure 18, red) with exceptions: for thoracic surgery (THO), traumatology (TRH) and Oral and maxillofacial surgery (MKA). The fact that fewer non-electives arrive on weekends can also be seen in Figure 11, last image. The same observation is true for night arrivals. Theoretically, 5/7 * 8/24 (~24%) of patients should arrive in a week night whereas in practice this value is significantly lower (Figure 18, blue). Non-electives are during slot time mostly assigned to an OR which at the moment of their arrival is allocated to their discipline. For example, a patient with a heart attack is probable to be assigned to an OR that at the moment the patients arrives is serving cardiology. From Figure 18 we see that the only exception to this rule is oncology which often serves its non-electives in the abdominal or the transplantation OR. One might think that this will only apply to non-electives that are not very urgent (DT 2 and 3) and that the most urgent non-electives (DT 1) are served immediately and therefore enter the first OR that becomes free. This is indeed true for MKA and to some extent for ONC, but for all other disciplines not (Appendix, Figure 55). 31 As a significant part of non-electives are allocated to ORs used by the elective program, it becomes necessary to recalcuate the MSS capacity calculations found in Figure 13. The result is shown in Figure 19. The calculations were done the following way: nonelective cardiology patients need 16.2 hours of OR time a week. From Figure 18 we know that only 67% will interfere with the elective cardiology program, i.e., arrive during slot hours and be served in an OR reserved for cardiology. As a result, in Figure 19 the red bars (DT 1-3) associated to cardiology will represent a length of 10.9 hours (0.67 * 16.2). Non-electives entering an OR other than their own discipline are assumed to be random and not included into the capacity calculations. This is an assumption valid for all ORs except B3 that despite being assigned to transplantation surgery (TX) is generally regarded to be the “emergency OR”. 32 Figure 18: Red columns are generally lower than the red line which means that less non-electives arrive on a day of the weekend than on a day during the week. The same logic applies to night arrivals (in blue). Figure 19: Where the lower column is larger than the capacity of the planned MSS are disciplines that are physically unable to serve all their patients in their planned slot time. Therefore, even if scheduled optimally, they will run into overtime. At Gasthuisberg the impact of overtime is reduced by the funneling mechanism (Section 3.2.5). 33 3.2.5. Funnel Not all ORs are planned to close at the same time but depending on the hour of the day some can remain open for longer: 8 (18:00 - 19:00), 4 (19:00 - 20:00) and 2 (20:00 – next morning). This means that gradually more and more ORs are closing towards the evening, therefrom the analogy of a “funnel” (Figure 20). This funneling mechanism helps the hospital to reduce overtime costs that would otherwise certainly occur (Figure 19 - some disciplines will unavoidably run into overtime). The amount of surgeries planned into an OR corresponds to 9 hours (7:45-16:45). An exception to this rule is cardiology, where for each day 1 or 2 surgeries are planned. As can be seen from Figure 15, cardiology surgeries tend to be long, therefore planning 2 cardiology surgeries into one OR almost certainly means that the OR goes into overtime (cardiology ORs have in average 1.2h overtime). One might wonder how it is decided what ORs to leave open, thus how the funnel is formed. First of all, ORs that finish in time do not join the funnel. It is the ORs that experience some sort of complication with one or more of their surgeries that might go into overtime and thus join the funnel. An OR can always be stopped from going into overtime by rescheduling its last surgery (cancel or reassign the surgery to another OR). To make those rescheduling decisions is difficult as it requires to balance between many competing criteria such as: patient satisfaction (not too many patients can be canceled), surgeon’s preference and OR staff requirements (nurses and anesthesiologist). At Gasthuisberg the rescheduling decisions are made by the head anesthesiologist and the head nurse. The anesthesiologist is responsible for decisions that affect the OR complex for the next 24 hours (OR reassignments) and is therefore in charge of the funnel, whereas the head nurse makes decisions that affect the OR complex on the longer run (cancelations). One might get the impression that non-electives are served as last ones (Figure 21), i.e., after the elective schedule. This can be 34 explained, on the one hand, by the fact that the number of ORs that can accommodate them is limited (Figure 18), i.e., it is unlikely that an OR just turns free. On the other hand, surgeons like to make sure that all of their electives scheduled for the day get served. In case they leave a non-elective patient as the last one of the day their schedule is “safe” as the head anesthesiologist will not force them to cancel that last non-elective case. If it would be the case that non-electives are scheduled last on a day (Figure 21), one might wonder whether there is a difference between DT 2 and 3, i.e., non-electives who have to be served within 6 hours and within 24 hours. As shown by Figure 22, there is a difference between the two: for DT 2 patients there is a 44.5% chance to be served within 3 hours of arrival, whereas for DT 3 patients this is only an 18.1% chance. For both DT 2 and 3 patients it is true that they rarely enter an OR immediately (within half an hour). By 48 hours (Figure 23, right) we see that most patients get served for all three DT categories. Non-electives that arrive during the weekend or on a holiday are excluded in Figure 22 and Figure 23, but are included in the Appendix in Figure 56 and Figure 57. DT categories 1-3 largely determine a patient’s expected waiting time in the first 6 hours (Figure 23, left). As the graph shows, the longer patients wait, the less it matters whether they are assigned to DT 1 or DT 2. By 6 hours, the chance of being served is 85% for DT 1, 76% for DT 2 and only 40.2% for DT 3 patients. 35 Figure 20 (day 205, 2012): The number of ORs that can remain open after 6:45 depends on the funnel (dashed line). The number of open ORs relates to nurse requirements: after 20:00, all surgeries could have be done in two ORs (E1, C3) despite the fact that physically five ORs were open (E1, C3, B2, D2 and D3). This means that only 2 OR teams were available. The staffing requirements for anesthesiologists follows a similar but slightly shifted funnel. The two ORs are open after 20:00 on weekdays and during the whole weekend mainly to serve non-electives. Figure 21 (day 332, 2012): All schedules for the year 2012 can be found on www.econ.kuleuven.be/healthcare/seminar_1. 36 Figure 22: Non-elective waiting time on weekdays. On the left side each bin is of size 30 min, e.g., within 30 minutes 19.8%, whereas after 30 minutes but before 60 minutes around 22% of DT 1 patients are served. On the right side the bin size is changed to 3 hours. Figure 23: Cumulative function of non-elective waiting time on weekdays. The bin size is 30 min on the left and 3 hours on the right. For example, for DT 1, within 30 min 19.8%, within 60 min 42% ... within 6 hours 85% of patients are served. Interesting to note is the fact that after 12 hours the probability of being served is higher for DT 2 than for DT 1. 37 3.2.6. Rescheduling Rescheduling is used to decrease the load on ORs that otherwise would go into overtime. Generally, in order to serve all the planned surgeries, the available capacity of an OR is enough. There are two major reasons why the capacity of an OR might turn out to be insufficient. Firstly, if during a surgery a complication occurs then the surgery duration can be larger than planned (Figure 15). Secondly, non-elective arrivals might interfere with the already fixed daily schedule (Figure 21). We distinguish between two types of rescheduling actions: OR reassignment and canceling. If a surgery is OR reassigned then the surgery will still be performed on the same day as planned and only the OR is changed (Figure 24). In case a surgery is canceled then the surgery is scheduled to the next slot available for the discipline even in case that slot is already fully booked. It is always preferred to perform a surgery on the planned surgery day and therefore cancelation is only used if there are no other options left. This is the case as it can become frustrating for a patient to physically and mentally prepare for a surgery that in the end is not carried out. Consequently, cancelations are not happening often: On average 1.4 surgeries are canceled a day while on the other hand 5.5 surgeries are OR reassigned a day. Examples of both OR reassignments (in blue) and cancelations (in black) are shown in Figure 20 and Figure 21. As Figure 25 shows, rescheduling activities are mostly concentrated in the morning and in the afternoon hours. Interestingly, the shapes of both the histograms of OR reassignments and canceling are similar with the difference being that the latter is shifted by 1-2 hours to the right, i.e., canceling decisions are generally made later in the day. Unexpectedly there is increased rescheduling activity observable in the early morning. This can be explained by the fact that sometimes complete ORs are switched, which in the system is registered as many individual OR reassignments. 38 Figure 24: OR reassignment schema. (1) It is preferred to OR reassign a patient to a slot with the same discipline. (2) Alternatively a surgery can be OR reassigned to another discipline within the kern. (3) Least favorable is to move a surgery within kern A, B, C and D. CAH and NCH can only be OR reassigned to slots of their own respective disciplines. Figure 25: The bin size represents the amount of rescheduling activity that is expected for the given hour of day. The figure shows, for example, that most OR reassignments happen between 13:00 and 14:00. 39 3.3. Model Within the simulation model, we implemented most of the patient and hospital specific mechanisms described in the previous section. In this section we will introduce three of the mechanisms in more detail. The first one relates to the relationship between real and estimated surgery durations. The second one relates to the way how rescheduling was modeled while the third one describes the way how the MSS is constructed and used in the simulation model. 3.3.1. Rescheduling model As described in Section 3.2.5, OR reassignment and canceling is used to decrease the load of ORs that otherwise would go into overtime. A surgery that is canceled is assigned to the slot closest in date. If on the closest date several slots of a discipline are available, the one with the lowest occupancy is chosen. A surgery that is canceled once is in the replanned OR always served first, ensuring that it is not canceled again. Neurosurgery cancelations are modeled slightly differently as in their case it is essential that the surgery is performed by the same surgeon. The surgeon’s next slot is one week later on the same day. Rescheduling actions are based on the estimated closing time of ORs. This estimate is the sum of two components: firstly, the sum of the estimated surgery durations still waiting and, secondly, the amount of time the current surgery is still expected to need. This last component is regarded to be zero in case the surgery already takes longer than its estimate. This is not a correct estimate, but is likely to be close to the value that is used in reality. The correct estimate would be the expected value of the conditional probability distribution of the surgery length given the amount of time the patient is already in the OR. 40 Figure 26: The expected closing time of Slot ‘A’ can fall into any of the three intervals: Green – OR closes in time, nothing has to be done; Yellow – OR goes overtime, if possible move last surgery, else keep it; Red – OR goes heavily into overtime, if possible move last surgery, else cancel it. The search procedure for an OR that can accept a moved surgery follows the hierarchy explained in Figure 24. An OR can only accept a moved surgery if, including the new surgery, the OR is still estimated to close before 16:45. Cancelation becomes an option for those ORs that are estimated to close after 18:00, but also in this case first it is always attempted to reassign its last surgery to another OR. As shown by Figure 28, rescheduling decisions are made continuously from the morning until the evening. Consequently, instead of rescheduling patients on one certain hour of the day, in our model we allow rescheduling interventions to happen on an hourly basis starting from 9:00 to 19:00. At each intervention point, we collect the ORs that are believed to run into overtime. For those ORs, it is attempted to move their last surgery to another less occupied OR. There is one problem with the method explained in Figure 26, namely, that if implemented, a disproportionally large number of surgeries will be rescheduled already early in the morning. This happens in case the first surgery is taking longer than expected. In reality, whether we believe that an OR goes overtime or not will depend on the degree of trust we put into our estimate. The later we are in the day, the more surgeries have been realized and consequently the better becomes our estimate. This relationship between the current time and the estimated OR closing time is captured by the formulas in Figure 27. In the formulas, the degree of trust we have in our estimates is modeled as a linear function of time (in blue). Put differently, we normalize the time of the day, that is, we map the time of the day onto the unit interval. Similarly we also normalize the estimated OR closing time. Estimates can be normalized in the two ways represented by the functions in yellow and red. The yellow function is used to check whether an OR fulfils the criteria to move its last surgery to another OR, that is, the OR is believed to go into overtime to a degree that justifies an OR reassignment. Similarly, the red function is used to test for cancelation. Important to note is that OR reassignments will never happen if the estimated OR closing time is before 16:45, as the yellow function will be on a constant zero. For the same reason we will never cancel a surgery from an OR that is estimated to close before 18:00. Applying this model in the simulation yields the results shown in Figure 28. 41 Figure 27: The last surgery of OR A is reassigned to another OR if, for example, it is 12:00 (blue function takes the value 1/3) and the estimated closing time of OR A is 18:00 (yellow function takes the value 1). The multiplication of 1/3 and 1 results in 1/3 which is larger than the threshold of 0.3. This would not be the case if the intervention point would be checked an hour earlier at 11:00. A similar logic applies to cancelations. The thresholds ‘0.3’ and ‘0.5’ were chosen on a ‘try and error’ basis, trying to fit the histograms in Figure 28. Noteworthy is the fact that the blue function reaches its maximum at 18.00 which means that after that point in time, we have full confidence in the estimates. Figure 28: The rescheduling statistics for Gasthuisberg are in blue and the statistics produced by our model are in green. 42 3.3.2. Copula The connection between realized and estimated surgery durations can be captured using a copula. Despite the fact that the theoretical background behind copulas is complex, they turn out to be relatively easy to use in practice [9]. The idea is the following: transform the durations (dots in Figure 16) into the unit space (e.g., using a kernel estimator of the cumulative distribution function). In the second step, measure the dependence between realized and estimated surgery durations (e.g., using maximum likelihood, estimate the linear correlation matrix and the degrees of freedom of the copula). Knowing the dependence between the variables it is easy to generate an arbitrary number of pairs of realized and estimated surgery durations in the unit space. Scaling them back into their original space yields the final random estimated and realized surgery duration pairs. The surgery durations generated with the copula method have the same pattern as the original, real data. 43 3.3.3. The applied Master Surgery Schedule In the simulation model we use a one week MSS that merges the two week cycle of the MSS (Table 2) used at Gasthuisberg. This can easily be done as the two weeks are almost identical. While creating the MSS applied in the simulation model, we were trying to find a trade-off between the MSS template, the planned MSS and the realized MSS. A practical problem we encountered in our simulation model concerns the disciplines cardiology and neurosurgery. The two disciplines can, in case they need to OR reassign, only do this to slots of their own disciplines. As a consequence, surgeries not fitting the slots will often be canceled. In case this happens, it is possible that chains of cancelations appear in the simulated schedule. This is to be avoided as, firstly, it is unrealistic because in reality this could be handled (e.g., by allowing more overtime or opening additional slots) and, secondly, it yields wrong results as many patients might be canceled to a slot that is not within their DT anymore. Figure 29: In the simulation model we use a MSS that constitutes a trade-off between the MSS template, the planned MSS and the realized MSS. 44 3.4. Implementation This section contains some of the technicalities related to the simulation model and the data processing step. 3.4.1. Simulation model Using the understanding of the Gasthuisberg patient scheduling setting, we needed a way to test and analyze newly developed scheduling policies. Because of the complexity of the setting we chose to use simulation. The simulation model was created based on the hospital’s data. Models of mechanisms that were not covered by the data were created on the basis of the explanations of our contacts at Gasthuisberg. Simulation models are often created using either custom healthcare related simulation software or from scratch using a general purpose language. Both methods have their benefits but also their drawbacks. The first option is quick to implement but there are mechanisms that are difficult to model within the software. Additionally, it is usually a standalone software and thus it is tedious and often impossible to properly integrate it with other environments. The second option, using a general purpose language, has the drawback of being time consuming to implement. Advantages are: flexibility, speed and high integrate-ability. We chose to create a simulation model that is based on a general purpose language but is seamlessly integrated with a custom discrete event simulation (DES) environment. We are using Matlab to code routines and Simulink’s SimEvents toolbox for the DES framework. 45 An advantage of using Matlab together with Simulink is that components from each environment are easily integrated. Practically, this means that we are able to call the Simulink DES model from Matlab while within Simulink we are able to use Matlab code. SimEvents is used to simulate the patient service process. This involves timing patient arrivals: for electives calling appropriate scheduling modules and for non-electives calling the appropriate non-elective to OR allocation schema (Figure 18 or appendix Figure 54). Also hospital mechanisms are implemented in SimEvents such as the surgery process in the OR and patient rescheduling actions. One of the drawbacks of SimEvents, in comparison to other DES environments, is its rudimentary nature, making it difficult to directly implement more complex mechanisms. This is to some extent compensated for by the fact that different models from other Simulink environments (e.g., state machines) or Matlab code can be mixed into SimEvents. One of the strong sides of SimEvents is that an entity (patient) can enter an attribute function block. The strength of this block comes from the fact that within the block it is, on the one hand, possible to arbitrarily change the entities’ attributes (e.g., scheduled OR) and, on the other hand, it allows to import object handles. This means that patient attributes can freely be processed within the simulation model using any class function created in Matlab. This practically means that an entity can be processed in an arbitrary way. This results in a highly flexible and capable environment. Matlab is used to create patient arrivals (Section 3.2.1) with their attributes (e.g., Section 3.2.3) and to model scheduling mechanisms. Matlab is a general purpose language and as such usable for many different tasks. For example, Matlab code is used in the simulation model to achieve parts of the funneling behavior, i.e., from 22:00 – 7:45 only 2 ORs are allowed to be open in parallel. Matlab code is also used to process and visualize simulation results. 46 Using a simulation model, and for that matter any model, one might wonder whether the results are correct. We will briefly elaborate on points regarding model verification, validity and credibility. The first point is model verification which is often defined as “ensuring that the computer program of the computerized model and its implementation are correct” [6]. The verification of our model happened on several levels. For example, we visualized and inspected the patient schedules (e.g., Figure 20), checked for all created entities the time they were in the system and whether they all left the system. We checked whether simplified inputs result in the correct output, e.g., used the same value for estimated and realized surgery durations and checked whether scheduling routines show the intended behavior. The second issue is model validation which is the “substantiation that a computerized model within its domain of applicability possesses a satisfactory range of accuracy consistent with the intended application of the model” [7]. The intended application of our model is only to show how different scheduling policies and methods compare against one another. That is, we do not try to predict the real performances of different scheduling policies in case they are implemented by the hospital. This would not be possible as the setting is too complex, e.g., some of the post-surgery units cannot be modeled as they serve patients from a variety of different facilities within the hospital but also from outside the hospital. The third issue is model credibility which is concerned with “developing in potential users the confidence they require in order to use a model and in the information derived from that model” [6]. We believe that our model is credible to the people interested in our results at Gasthuisberg. We created the model based on the data of Gasthuisberg and through several meetings we confirmed that we have the right understanding of the data. Also the model of surgery scheduling related processes are both based on the hospital’s data and on the managerial insights and expertise of the people at the hospital. 47 3.4.2. Data processing In order to test patient scheduling policies that are applicable to Gasthuisberg, we needed to understand their scheduling setting. Therefore, firstly, we analyzed the properties of the patients getting surgery at their hospital and, secondly, examined the service process of their ORs. In order to achieve the aforementioned tasks we used the hospitals historical data. The hospital provided us with patient scheduling records of the complete year 2012 and helped us to correctly interpret those records. The data was grouped into three datasets: Patient trajectory, OR information and Patient planning related data. The first data set, Patient trajectory, contains 468.599 entries and 23 attributes. The dataset contains, on the one hand, the date and time patients were transferred to different facilities in the hospital and, on the other hand, general information concerning the patient such as arrival date, ID of surgeon in charge and whether the patient arrived as an emergency. The second dataset, OR information, contains 17.310 entries and 14 attributes. The dataset includes surgery specific information, such as the IDs of the planned surgical interventions, estimated surgery duration and the DT of the patient. The last dataset, Patient planning, contains 28.514 entries and 19 attributes. The data contains surgery scheduling related records, such as the surgery’s planned date and OR. A new entry is created in the table anytime the planner changes the scheduling information of a patient. Before merging the three datasets, each of them was individually preprocessed. Preprocessing involved, for example, combining entries in a dataset that relate to the same surgery. The datasets were merged in a way that all the information attached to one surgery became easily accessible and processable. Some entries were lacking important attributes, such as the arrival data and were therefore removed from the dataset. After removing all uncompleted entries 15596 patients remained from 13 disciplines. 48 As surgeries can be replanned several times, the Patient planning table will generally contain more than one entry. Patient trajectory and OR information related datasets are merged using the attribute ‘aanwezigheidKey’. Patient planning related data is included using attributes ‘okBonNr’ and ‘OpnamevoorstelNr’. Further post-processing steps were made to make the dataset easier to interpret. One of the steps includes the interpretation of the patient planning points based on the type of rescheduling action they represent (Figure 30). Some planning points are regarded to be invalid. It can happen that planners after making a planning decision for a surgery change their mind and decide to plan the surgery differently: in this case we are only interested in the planners’ final decision. Therefore, if there are two entries which are made within 30 minutes then only the latter is considered. Furthermore, as Figure 30 shows, a surgery can be replanned before the surgery day in which case however it’s not regarded to be an OR reassignment or cancelation. The statistics related to the time of the day when surgeries are OR reassigned or cancelled can be found in Section 3.2.5. 49 Figure 30: Planning decisions that were made for one patient. This patient arrived the 1 th of December at 17:03 and was planned into OR C3. The patient was OR reassigned to E2 at 11:07 on the planned surgery date. As E2 was apparently overfilled, the patient was canceled at 20:37 and finally served 2 days later. 50 4. References 1. Adan I, Bekkers J, Dellaert N, Jeunet J, Vissers J (2011) Improving operational effectiveness of tactical master plans for emergency and elective patients under stochastic demand and capacitated resources. Eur J Oper Res In Press, Corrected Proof. doi:DOI: 10.1016/j.ejor.2011.02.025 2. Cardoen B, Demeulemeester E, Belien J (2009) Optimizing a multiple objective surgical case sequencing problem. Int J Prod Econ 119 (2):354-366. doi:10.1016/j.ijpe.2009.03.009 3. Cardoen B, Demeulemeester E, Belien J (2010) Operating room planning and scheduling: A literature review. Eur J Oper Res 201 (3):921-932. doi:10.1016/j.ejor.2009.04.011 4. Cardoen B, Demeulemeester E, Van der Hoeven J (2010) On the use of planning models in the operating theatre: Results of a survey in flanders. The International Journal of Health Planning and Management 25 (4):400-414. doi:10.1002/hpm.1027 5. Hans E, Wullink G, van Houdenhoven M, Kazemier G (2008) Robust surgery loading. Eur J Oper Res 185 (3):10381050. doi:10.1016/j.ejor.2006.08.022 6. Sargent RG (2005) Verification and validation of simulation models. Paper presented at the Proceedings of the 37th conference on Winter simulation, Orlando, Florida, 7. Schlesinger S, Crosby R, Cagne R, Innis G, Lalwani CS, Loch J, Sylvester R, Wright R, Kheir N, Bartos D (1979) {terminology for model credibility}. SIMULATION 32 (3):103-104. doi:citeulike-article-id:4813312 8. Schoenmaeker AD, Kimpen S Onthaalbrochure e905 & e906. 9. Trivedi P, Zimmer D (2006) Copula modeling: An introduction for practitioners. Foundations and Trends® in Econometrics 1 (1):1-111. doi:citeulike-article-id:6425795 51 5. Literature Review: OR planning Abstract Operating room (OR) planning and scheduling decisions involve the coordination of patients, medical staff and hospital facilities. The patients arriving to the hospital are assigned to a surgery date and a surgery time slot. At the time of surgery, a suitable OR, the attending surgeon, supporting anesthesiologists, nurses and, after the surgery, room in secondary facilities such as post anesthesia care unit (PACU), intensive care unit (ICU) and ward need to be available. In order to deal with the complexity and the variety of problems faced in OR scheduling, it is useful to involve methods from operations research. In this chapter, we review the recent literature on the application of operations research to OR planning and scheduling. We start by discussing the impact of planning and scheduling of the ORs on the overall performance of a hospital. Next, we discuss the criteria for included publications and summarize the structure of Cardoen et al. [29] that served as the guideline for organization of this chapter. In the remainder of the chapter, we describe the evolution of the literature over the last ten years with regards to the patient type, the different performance measures, the decision that has to be made, the incorporation of uncertainty, the operations research methodology and the applicability of the research. Moreover, each of these evolutions will be demonstrated with a short review of some relevant papers. The chapter ends with conclusions and a discussion of interesting topics for further research. 5.1. Introduction Healthcare has a heavy financial burden for governments within the European Union as well as oversees. While growing economies and newly emerging technologies could lead us to believe that supporting our respective national healthcare systems might get less expensive over time, data show the contrary is true. For example, within the U.S. the NHE (National Health Expenditure) as a share of the Gross Domestic Product (GDP) increased from 15.6% in 2004 up to 16.2% in 2009 [112]. The data suggest that increasing health expenditures are an ongoing trend with an estimated annual growth rate of 6.3%. The NHE share of the GDP in the U.S. is thereby being projected to hit 19.6% by the year 2019 [31]. 52 Similarly on the European continent, even though differences exist across member states, healthcare spending as a share of the GDP is increasing, where countries that are hit hard by the global recession are most affected. For example, in Ireland, the percentage of GDP devoted to healthcare increased from 7.6% in 2004 to 9.5% in 2009. In the UK during the same time interval, a rise from 8% to 9.8% was experienced [112]. A large amount, estimated at 31% of the spending on healthcare, pertains to hospitals [32], which are consequently being pressured to reduce costs. Hospitals are expensive from the patient perspective as well. For instance, Milliman Medical Index (MMI) estimates that, for a typical family of four, 48% of the family health care spending involves hospital costs [104]. Within the hospital, special attention is given to ORs as they represent the largest costs and provide the largest revenues [1]. It comes as no surprise that the body of literature dealing with topics related to OR efficiency and profitability is steadily increasing. Out of the many aspects, we focus our attention on planning and scheduling problems and do not include topics related to business process reengineering, the impact of introducing new technologies or facility design. Our work is the natural continuation of the literature review carried out by Cardoen et al. [29]. In this chapter we complement the work by adding the recent body of literature (2007-2010) and including a more in-depth analysis of the trends being followed by the research community. Trends are investigated on the interval starting from 2000 and ending at 2010, while detailed descriptions are only provided for the most recent contributions being published after 2006 and which were not included in [29]. In the past 60 years, a large body of literature on the management of ORs has evolved. Magerlein and Martin [95] distinguish between advance scheduling and allocation scheduling as they provide us with a review on surgical demand scheduling. Advance scheduling is the process of fixing a surgery date for a patient, whereas allocation scheduling determines the OR and the starting time of the procedure on the specific day of surgery. Blake and Carter [17] elaborate on this taxonomy in their literature review and add the domain of external resource scheduling, which they define as the process of identifying and reserving all resources external to the surgical suite necessary to ensure appropriate care for a patient before and after surgery. Przasnyski [125] structures the literature on OR scheduling based on general areas of concern, such as cost containment or scheduling of specific resources. The more recent review by Gupta and Denton [67] focuses on appointment scheduling from a general perspective and describes three commonly encountered health care scheduling environments, namely primary care, specialty care, and elective surgery (nonemergencies). In primary care usually a primary care physician, such as a general practitioner or a family physician, acts as the principle point of consultation. Specialty care is focused on a specific and often complex diagnosis and treatment, whereas elective surgery is focused on a specific procedure. They discuss various factors, which affect the performance of the appointment system, including arrival and service time variability, patient and provider preferences and performance measures such as patient waiting time, OR idle time and overtime. According to these factors, the existing literature is classified into three groups. The difference between Gupta and Denton’s review and our review is that our focus is limited to scheduling problems which arise in close relationship to the OR. Consequently, we include elective surgery scheduling and exclude considerations related to primary and specialty care. In the literature review of Guerriero and Guido [64], a selected number of articles are categorized according to the commonly used three hierarchical decision levels: strategic, tactical and operational. Strategic decisions involve defining both number and types of surgeries to be performed, and hence affect the OR function in the long term. The tactical level usually involves the construction of a 53 cyclic schedule, which assigns time blocks to surgeons or surgeon groups. The final, operational level does not influence the number and type of performed procedures, but deals mostly with daily staffing and surgery scheduling decisions. The three hierarchical levels give the OR planning problem structure. Nonetheless, in our literature review we chose not to use the three hierarchical levels but rather to define descriptive fields. The justification for our decision can be found in Section 2. Other reviews, in which OR management is covered as a part of global health care services, can be found in [21,124,135,165]. We searched the databases Pubmed and Web of Knowledge1 for relevant manuscripts, which are written in English and appeared in 2000 or afterwards. Search phrases included combinations of the following words: operating, surgery, case, room, theatre(er), scheduling, planning and sequencing. We searched in both titles and abstracts and in addition checked the complete reference list of any found article. As the search process happened in an unbiased way, we believe to have arrived at a set of articles, which objectively represents most of the articles in the field. At the end of the search procedure, a set of 181 articles was identified of which 136 [2,3,516,18-20,22-25,28,27,33-41,43,45,46,48-57,59-63,65,66,68-73,58,74-84,86-94,96-100,105-111,113-115,117-123,126-132,136141,143,142,144-148,150-157,161-164,166-168] were found to be technically oriented. We define an article as “technical” if it contains an algorithmic description of a method directly related to patient scheduling. Some articles missing this algorithmic component instead provide managerial insights. Those types of articles are classified as “managerial” and excluded from the classification process itself. However, these are mentioned in the text where they seem appropriate. The quantitative descriptions provided in later sections, which try to give insights into the dynamics of the trends followed by the research community, are exclusively based on the technical contributions. The distribution of included articles according to their year of publication is shown in Figure 31. 1 Includes: Web of Science, Current Contents Connect and Inspec 54 35 30 25 Both technical and managerial 20 Technical 15 Managerial 10 5 0 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 Figure 31: The distribution of included articles according to their year of publication. Figures appearing in the text are based on technical contributions only. The structure we use is meant to balance between simplicity and expressiveness. We provide a simplified, but in our belief for the majority of the readers, sufficiently accurate way to identify and select articles they are interested in. In the detailed tables, all researched manuscripts are listed and categorized. Pooling these tables over the several fields should enable the reader to reconstruct the content of specific papers. They furthermore act as a reference tool to obtain the subset of papers that correspond to a certain characteristic. To search for literature in a more direct way the database containing the detailed classification of each analyzed article is made available at www.econ.kuleuven.be/healthcare/review2011 in the form of an Excel [Microsoft, Redmond, WA] spreadsheet. 5.2. Organization of the review As in [64], many authors differentiate between strategic (long term), tactical (medium term) and operational (short term) approaches, and situate their planning or scheduling problem accordingly. With respect to the operational level, a further distinction can be made between offline (i.e. before schedule execution) and online (i.e. during schedule execution) approaches. The boundaries between these major categories can vary considerably for different settings and hence are often perceived as vague and interrelated [134]. Furthermore, this categorization seems to lack an adequate level of detail. Other taxonomies, for instance, are structured and categorized on a specific characteristic of the paper, such as the use of the solution or evaluation technique. However, when a 55 researcher is interested in finding papers on OR utilization, a taxonomy based on solution technique does not seem very helpful. Therefore, we propose a literature review that is structured using descriptive fields. As in Cardoen et al. [29] each field analyzes the manuscripts from a different perspective, which can be either problem or technically oriented. In particular, we distinguish between six fields: Patient characteristics (Section 3): reviewing the literature according to the elective (inpatient or outpatient) or non-elective (urgency or emergency) status of the patient. Performance measures (Section 4): discussion of the performance criteria such as utilization, waiting time, preferences, throughput, financial value, makespan, and patient deferral. Decision delineation (Section 5): indicating what type of decision has to be made (date, time, room or capacity) and whether this decision applies to a medical discipline, a surgeon or a patient (type). Uncertainty (Section 6): indicating to what extent researchers incorporate arrival or duration uncertainty (stochastic versus deterministic approaches). Research methodology (Section 7): providing information on the type of analysis that is performed and the solution or evaluation technique that is applied. Applicability of research (Section 8): information on the testing (data) of research and its implementation in practice. Each section clarifies the terminology if needed and includes a brief discussion based on a selection of appropriate manuscripts. Furthermore, a detailed table is included in which all relevant manuscripts are listed and categorized. Plots are provided to point out some of the trends followed by the research community. It should be noted that, if not stated otherwise, the percentages are calculated in relation to the total amount of technical papers. Also note that some categories are not interpretable for some methods and even though rare, some articles contain more than one single method. As a consequence, the sum of mutually exclusive categories does not necessarily add up to 100%. 5.3. Patient characteristics Two major patient classes are considered in the literature, namely elective patients and non-elective patients. The former class represents patients for whom the surgery can be planned in advance, whereas the latter class groups patients for whom a surgery is unexpected and hence needs to be planned on short notice. A non-elective surgery is considered an emergency if it has to be performed immediately and an urgency if it can be postponed for a short time (i.e. days). 56 As shown in Figure 32 and Table 3, the literature on elective patient scheduling is vast compared to the non-elective counterpart. Although many researchers do not indicate what type of elective patients they are considering, some distinguish between inpatients and outpatients. Inpatients refer to hospitalized patients who have to stay overnight, whereas outpatients typically enter and leave the hospital on the same day. As pointed out by [103], in reality an ongoing shift of services from the inpatient to the outpatient setting is present, which is reflected in a higher growth rate of the latter. Moreover, according to the MMI, outpatient costs increase with a yearly value of 10%, of which 90% are attributable to increasing prices of existing and more expensive emerging services [104]. Despite the increasing importance of outpatient care in general, the share of outpatient related academic articles as shown by Figure 32 remains constant in time. 100% 90% 80% 70% 60% Elective 50% Elective Inpatient 40% Elective Outpatient Non-elective 30% 20% 10% 0% 2005 2006 2007 2008 2009 2010 Figure 32: The majority of contributions relate to the elective patient. 2 Surprising is that, contrary to what might be expected, the share of outpatient related articles is not increasing. 2 As some articles deal with both elective and non-elective patients, the sum of the two will add up to a value higher than 100%. 57 Elective inpatient 2,8,13,14,16,23,25,36,48,62,67,72,74,84,94,108,109,118,123,136,137,139,140,143,144,148,154,156,167 outpatient 8,15,23,25,27,28,40,43,46,48,52,61,67,74,75,76,84,108,109,123,126,129,139,140,148,154,157,167 not specified 3,5,6,7,9,11,22,24,34,35,37,38,39,45,49,50,54,55,56,57,58,59,60,65,68,69,70,71,73,77,80,81,82,83,86,87,88,89,9 0,91,92,96,99,100,105,106,107,111,113,114,115,117,119,120,121,122,127,128,130,132,138,141,142,145,146,150 ,151,152,153, 155,161,162,163,164,166,168 Non-elective urgent 22,54,65,67,96,109,110,123,168 emergent 9,25,60,67,71,72,82,87,88,89,90,92,100,106,108,113,122,123,139,141,142,150,154,163,167 not specified 83,84,114,151,162 Table 3: It is not always specified what type of patients is considered and, especially for the elective patient case, not always clear whether an inpatient or outpatient setting is used. Scientific contributions on outpatients often (22 out of 37 articles) investigate ORs in an integrated way, i.e., a preoperative or postoperative unit is taken into account. For example, Huschka et al. [76] use both an intake and a recovery area as part of a simulation model of an outpatient procedure center. Several daily scheduling and sequencing heuristics are applied and tested on their impact on patient waiting time and the amount of OR overtime. The authors found that defining the order of surgeries has less influence on patient waiting time and OR overtime as the arrival time schedules. Additionally, the importance of a proper daily surgery mix is stressed. Other methods focus on assigning patients to days and do not define the exact procedure starting times. Lamiri et al. [88] consider several stochastic optimization methods to plan elective surgeries in case OR capacity is shared by both elective and emergency patients and present a solution method combining Monte Carlo sampling and mixed integer programming. Several heuristics were also tested, from which the most efficient one proved to be tabu search. In their problem setting, it is surprising that the quality of heuristic solutions degrades as the variability on the amount of emergency arrivals decreases, i.e., the stochastic problem is easier than the deterministic one. Planning for stochastic emergency arrivals poses a problem because it leaves an uncertain amount of capacity left for elective patients. Ferrand et al. [61] investigate whether eliminating this source of uncertainty by channeling emergency arrivals to dedicated emergency ORs is beneficial. This requires however that a constant number of ORs be reserved for emergencies, and therefore leaves less free capacity for elective patients. Nevertheless, based on their results using a simulation model, they find elective patients benefit from this, whereas emergency arrivals have an increased waiting time. A scenario where a hospital dedicates all of its ORs to emergency services is the case of a disaster. As a consequence, all elective surgeries are cancelled while resources are redirected to provide quick care to non-electives. This type of non-elective patient is an urgency, as quick but not necessarily immediate care is required. Nouaouri et al. [110] sequence a large number of patients resulting 58 from a disaster, with the objective of maximizing patient throughput. Their approach identifies patients which cannot be served by the given hospital and therefore have to be transported to another one. Zonderland et al. [168] refers to patients who have to be treated within 1-2 weeks as semi-urgencies and uses queuing theory to analyze the trade-off between reserving too much OR time for their arrival (which results in unused OR time) and excessive cancellations of elective surgeries. Additional complexity results from the fact that cancelled elective patients turn into semi-urgent patients, which consequently need to be served within the following two weeks. An insight gained by the authors is that focusing only on the average behavior of the system can result in undesired system outcomes, i.e., shortage of capacity, which ultimately leads to the cancellation of many elective patients. 5.4. Performance measures Different performance measures emphasize different priorities and will favor the interests of some stakeholders over others. A hospital administrator could, for example, be interested in achieving high utilization levels and low costs. Medical staff, on the other hand, might care less about cost factors and rather aim to achieve low overtime. The patient as the client of the hospital might care little about the above factors and only desires high quality service and short waiting times. Many authors in the scientific community try to find a compromise between the interests of different stakeholders and simultaneously enforce several kinds of performance criteria. As a matter of fact, as shown by Figure 33, a gradually decreasing number of authors restrict themselves to only one single performance measure. 59 100% 90% 80% 70% 60% 50% Multiple performance criteria 40% Single performance criteria 30% 20% 10% 0% 2005 2006 2007 2008 2009 2010 Figure 33: It is increasingly popular to use multiple instead of single performance criteria. The criteria are not interpretable for all articles, their sum is thus lower than 100%. The major performance measures we distinguish can be found in Table 4: utilization, waiting time, leveling, preference, throughput, financial measures, makespan, and patient deferral. As shown by Figure 34, patient waiting time is a frequently used performance measure, which is understandable as one of the major problems in general health care consists of long waiting lists but also extensive waits on the day of surgery. Wachtel and Dexter [159,160] investigated increases in waiting time on the day of surgery, for both surgeon and patient, caused by tardiness from scheduled start times. They concluded that the total duration of preceding cases is an important predictor of tardiness, i.e., the tardiness per case grew larger as the day progressed. If case durations are systematically underestimated, tardiness results. The estimated amount of under-utilized (over-utilized) time, which was to be expected at the end of the workday, caused average tardiness to increase (decrease). A reduction of tardiness can be achieved by modifying the OR schedule to incorporate corrections for both lateness of first cases of the day and case duration bias. 60 Utilization underutilization /undertime OR 2,19,20,36,50,55,56,57,58,59,68,73,78,79,80,90,91,100,111,113,115,140,142,148,152,156,161,164,167,16 8 Ward 156 ICU 2,36,156 PACU 2,36 overutilization /overtime OR 2,11,22,25,33,36,37,38,39,40,41,50,51,54,55,56,57,58,59,60,64,65,69,73,76,77,78,79,80,87,88,89,90,91,9 6,98, 100,105,106,111,113,115,119,121,122,123,127,128,132,137,140,141,142,145,148,150,151,156,161,163,1 64 Ward 54,156 ICU 2,36,115,156 PACU 2,27,28,36 general 3,8,9,12,22,23,25,33,34,48,50,54,60,61,69,71,72,91,111,122,132,139,142,145,148,150,151,163 Waiting time patient 3,9,25,33,36,37,38,40,54,60,61,64,65,67,76,78,79,81,91,93,106,107,109,111,118,120,121,122,132,136,13 9,141, 142,143,144,148,150,154,157,163,167 surgeon 11,37,38,91,157 Leveling OR 15,98,99,111 Ward 13,14,16,68,94,130,140,153 PACU 15,27,28,75,96,97,131,139,152 61 Patient volume 111,140,142 Preference 16,18,27,28,34,35,52,62,83,88,94,105,106,107,114,115,118,138,140,143,144,145,146,155,162,166 Throughput 3,8,9,12,71,100,110,126,130,132,142,145,154 Financial 11,18,24,39,43,45,46,48,49,65,84,93,100,108,138 Makespan 5,6,7,35,57,58,59,73,74,75,86,96,123,129,137,147 Patient deferral 22,25,36,54,71,72,81,82,92,120,121,122,132,139,145,168 Other 2,9,10,12,33,35,36,61,68,77,83,87,89,90,93,96,99,100,117,121,123,127,128,136,141,142,153,155 Table 4: The main performance criteria are: Utilization, Waiting time, Leveling, Preference (e.g., priority scooring), Financial (e.g., minimization of surgery costs), Makespan (completion time), Patient deferral, and Other (e.g., number of required porter teams). Figure 34 shows that patient throughput is rarely used as a performance measure, and patient or surgeon preferences are increasingly gaining popularity. Preference most often covers some qualitative aspect, whereas throughput is associated with volume. Noteworthy is the fact that both in general healthcare [149] and in the operations research literature value or quality based approaches seem to be getting increasingly important. For example, the preferences of cataract surgery patients of one surgeon are investigated by Dexter et al. [42]. The surgeon’s patients place a high value on receiving care on the day chosen by them, at a single site, during a single visit, and in the morning. Preferences can also be embodied in patient priorities, i.e., preference to perform surgery first on patients who have a more acute condition. Priorities are most often defined at the level of a patient group. Testi et al. [144,146] define a model where the position of a patient on a waiting list is defined by a priority scoring algorithm, which considers both patient urgency (progression of disease, pain or dysfunction and disability) and time spent on the surgical waiting list. It is easy to see that priority scoring minimizes the total weighted waiting time of all patients, i.e., using an algorithm where patient priorities are uniform, we will minimize the average patient waiting time. 62 100% 90% 80% 70% 60% Overtime 50% Patient waiting time 40% Preference Throughput 30% 20% 10% 0% 2005 2006 2007 2008 2009 2010 Figure 34: Some selected performance criteria. Overtime, despite used slightly less in 2010, it is still the most frequently used performance measure. From 2008 on, preference related measures seem to become increasingly popular. Including patient priorities drives OR scheduling in a more patient oriented direction. Min and Yih [105] go a step further and explicitly incorporate an additional payment factor, defined as the cost of overtime. In their model, if many high priority patients are on a waiting list, ORs will be kept open longer in order to avoid high postponement costs, i.e., the surgery postponement costs are balanced versus OR overtime costs. The authors establish that patient prioritization is only useful if the difference between the cost coefficients associated to different priority classes is high, as otherwise a similar schedule can be obtained by using average postponement costs. Additionally, the relative cost ratio between the cost of patient postponement and OR overtime should not be low, as a low ratio would imply high overtime costs and therefore prioritization would only marginally affect the surgery schedule. Purely financial objectives are rarely used in the literature. In Stanciu and Vargas [138], protection levels, the amount of OR time reserved in a partitioned fashion for each individual patient class, are used to determine which patients to accept and which to postpone during the planning period under study. A patient class is a combination of the patient reimbursement level and the type of surgery. A patient class enjoys higher priority if its expected revenue per unit surgery time is higher. The goal is to maximize expected revenues incurred by the surgical unit. Patients, given their class, are accepted or postponed to a later date when the protection level for their class can accommodate them. The central question becomes how many requests to accept from low revenue patients, and how much capacity to reserve for future high revenue patients. Financial considerations are also expressed by Wachtel and Dexter [158], who argue that if additional OR capacity is expanded, it should be assigned to those subspecialties that have the greatest contribution margin per OR hour (revenue minus variable cost), that have the potential for growth, and that have minimal need for a limited resource such as ICU beds. 63 Figure 34 also reveals the fact that minimizing overtime is a popular performance measure. This is understandable as overtime results both in the dissatisfaction of the surgical staff and high costs for the hospital (as higher wages typically apply for the time beyond the normal working hours). Reducing overtime is consequently highly beneficial and often desirable to practitioners. Dexter and Macario [47] establish that a correction of systematically underestimated lengths of case durations would not markedly reduce overutilization of ORs. Tancrez et al. [141] proposed an analytical approach that takes into account both stochastic operating times and random arrivals of emergency patients. They showed how the probability of overtime in the OR changes as a function of the total number of scheduled operations per day. As shown in Table 4, we relate underutilization to undertime and overutilization to overtime, although they do not necessarily represent the same concept. Utilization actually refers to the workload of a resource, whereas undertime or overtime includes some timing aspect. Hence, it is possible to have an underutilized set of ORs, although with overtime in some of the ORs. In some manuscripts it is unclear which view is applied. Therefore, we group underutilization with undertime and similarly overutilization with overtime. Adan et al. [2] formulated an optimization problem which minimizes the deviation from a targeted utilization level of the OR, the ICU, the medium care unit and nursing hours. The deviation is measured as the sum of overutilization and underutilization. They recommend using stochastic time durations as well as a broad perspective on supporting resources such as the ICU and the wards. Pandit and Dexter [116] defined rules to determine whether an OR should be staffed for 8 or 10 hours. They concluded that in case the average OR time is less than 8 h 25 min, 8 h staffing should be planned, while in case the average OR time is over 8 h 50 min, 10 h of staffing is needed. For averages in between, the full analysis of McIntosh et al. [102] should be performed. Augusto et al. [7] minimized the sum of the makespan (completion time) of patients undergoing surgery. Makespan in general defines the time span between the entrance of the first patient and the finishing time of the last patient. Since minimizing the makespan often results in a dense schedule, deviations from the plan can result in complications requiring adjustments to the schedule. An example is the arrival of an emergency patient to the hospital which results in a dense schedule with no free ORs available. In a case like this, a likely scenario includes the deferral of an elective patient, who will have to be reassigned to another surgery slot at a later point in time. Zonderland et al. [168], used queuing theory to investigate the trade-off between cancellations of elective surgeries due to semiurgent surgeries and unused OR time. They provide a decision support tool, which assists the scheduling process of elective and semiurgent cases. General reasons for patient deferrals in one specific hospital are discussed by Argo et al. [4]. In addition to emergency patient arrivals, other reasons can cause cancelation of elective surgeries. For example, a fully occupied PACU would prohibit patients who have already completed surgery from leaving the OR. A blocked OR will delay the service of preceding elective surgeries, which as a final measure may have to be cancelled. This situation can be avoided if the OR schedule is constructed in a way that resource utilization is leveled. Similar approaches may be used for other resources. For instance, Ma and Demeulemeester [94] maximize the availability of the number of expected spare beds and investigate bed occupancy levels at wards, whereas Van Houdenhoven et al. [152] target the ICU. Some of the articles in the literature used methods which have not been covered in the previous paragraphs. Does et al. [53] used Six Sigma to decrease the tardiness of surgeries, which are performed first on a day. Applied to two hospitals in the Netherlands, substantial savings were achieved and the number of surgeries was increased by 10% without requiring additional resources. Epstein and Dexter 64 [44] introduced a method through which analysts can screen for the economic impact of improving first-case starts. In [2,77] nursing hours are considered, while [128] considers the number of open ORs, and [141] the number of disruptions. 5.5. Decision delineation A variety of planning and scheduling decisions are studied in the literature. In Table 5, we provide a matrix that indicates what type of decisions are examined, such as the assignment of a date (e.g., on Friday, February 25), a time indication (e.g., at 2 p.m.), an OR (e.g., OR 1, OR of type B) or the allocation of capacity (e.g., one hour of OR time). The manuscripts are further categorized according to the decision level they address, i.e., to whom the particular decisions apply. We distinguish between the discipline (e.g., pediatrics), the surgeon and the patient level. Figure 35 shows that a large part of the literature aims at the patient level, whereas the discipline as well as the surgeon level receives less attention. 100% 90% 80% 70% 60% Patient 50% Discipline 40% Surgeon Other 30% 20% 10% 0% 2005 2006 2007 2008 2009 2010 Figure 35: In the literature, most decision problems relate to the patient level. A typical problem setting, for example, consists of finding the optimal patient to day / OR assignment. The discipline level unites contributions in which decisions are taken for a medical specialty or department as a whole. This frequently involves the construction of a cyclic timetable, which aims at minimizing the underallocation of a specialties’ OR time with respect to its predetermined target time. On a lower level, the target is a surgeon group or a single surgeon. In Denton et al. [39], surgeries 65 consecutively carried out by one surgeon define a surgery block, which are subsequently assigned to ORs. The problem is formulated as a stochastic optimization model, which balances the cost of opening an OR with the cost of overtime. Discipline Level Surgeon level Patient level Other Date 54,143,156 13,19,20,24,33,49,65,13 14,15,16,25,74,80,120,1 2,25,33,34,36,48,49,54,55,56,57,58,59,64,65,68,69,72,73, 42 0,132, 78,79,80,81,82,83,87,88,89,90,94,105,106,107,111,115, 144,145,167 118,119,120,121,123,127,128,132,136,140,143,144,145, 146,152,153,155,161,164,166 Time 13,33,49,65,71,132,145 11,14,15,16,25,375,6,7,11,25,27,28,33,37,38,40,41,49,57,58,59,65,70,71,73, 11,12 75,76,77,78,79,83,86,91,96,97,99,110,123,127,128,132, 137,145,147,150 Room 19,20,33,62, 130,132,144, 145,167 11,143 11,15,16,39,74,80,120,1 11,23,27,28,33,34,35,41,50,51,55,56,57,58,59,60,64,68,69, 42 70,73,76,77,78,79,80,83,86,87,90,96,98,99,106,107,110, 111,115,118,120,121,123,127,128,129,132,137,143,144, 145,152,153,161,163,164 Capacity 22,24,33,49,65,67,71,13 11,18,25,37,43,45,46,80, 2,11,22,25,33,36,37,49,54,64,65,68,71,72,80,94,105,108, 11,12,52,54,61,92,93, 84,120 114,120,121,132,138,141,145,168 109,117,122,126,131,154 0,132, 139,145,167 Other 151 120 3,7,35,54,111,120,136,143,157,162,168 52,54,143,148 Table 5: Type and level of decisions. For example, articles dealing with the sequencing problem are found in column 3 and row 2. Articles dealing with the problem generally referred to as the assignment step are found in column 3 and row 1. Defining patient capacity requirement for a given day of the week are articles found in column 3 and both row 1 and row 4. On the patient level, the decision variables are formulated on the individual patient or patient type. Fei et al. [59] describe a two stage approach, where in the first stage patients are assigned to days and rooms, and the exact daily sequence of surgeries (timing aspect) is set in the second stage. This is a common way of patient scheduling, as the assignment of the day and the room of a given surgery is more easily planned ahead than the exact time, which is often fixed close to the actual surgery date. Date, room, time and capacity questions as shown in Table 5 are answered on all three decision levels. Figure 36 shows that the share of manuscripts dealing with decisions on exact times is decreasing, whereas date, room or capacity problems are continuously addressed in the literature. A capacity problem is discussed by Masursky et al. [101] who forecasted longterm anesthesia and OR workload. They concluded that forecasting future workload should be based on historical and current workload related data and advise against using local population statistics. The problem of forecasting workload is addressed by Gupta et al. [68] as well. In his case study, simulation is used to answer capacity related questions. They concluded that a one-time infusion of capacity 66 in the hope to clear backlogs will fail to reduce waiting times permanently, while targeting extra capacity to highest urgency categories reduces all-over waiting times including those of low urgency patients. In situations where arrival rates increased, even if only within a specific urgency class, waiting times increased dramatically and failed to return to the baseline for a long time. 100% 90% 80% 70% Assignment of date 60% Assignment of time 50% Assignment of room 40% Assignment of capacity 30% Other 20% 10% 0% 2005 2006 2007 2008 2009 2010 Figure 36: Solution approaches related to the assignment of dates and rooms are increasingly popular in the literature, whereas the time assignment step (e.g., sequencing) is slightly less popular than it used to be. We added both a row and a column (other) to Table 5 to provide entries for manuscripts that study OR planning and scheduling problems in a way that is not well captured by the main matrix. Manuscripts that are categorized in this column or row examine for instance, capacity considerations with regards to beds [92,131], OR to ward [143] and patient to week assignments [168] or different timing aspects, such as the amount of recovery time spent within the OR [7]. As OR planning and scheduling decisions affect facilities throughout the entire hospital, it seems to be useful to incorporate facilities, such as the ICU or PACU, in the decision process in an attempt to improve their combined performance. If not, we believe that improving the OR schedule may worsen the efficiency of those related facilities. Figure 37 shows that the ratio of manuscripts between 2005 and 2010, which deal with the OR in an integrated way, and those which deal with the OR in an isolated way, are oscillating around the 50% mark. This is surprising as we would expect to observe an increasing interest in integrated approaches. Whether a manuscript uses an integrated or an isolated approach can be looked up in Table 6. 67 100% 90% 80% 70% 60% 50% OR is isolated 40% OR is integrated 30% 20% 10% 0% 2005 2006 2007 2008 2009 2010 Figure 37: Every second article deals with the OR planning and scheduling process in an integrated way. Isolated OR 5,10,11,19,20,22,33,34,37,38,39,41,45,48,49,50,51,52,55,56,57,58,59,60,64,65,67,68,69,73,74,83,84,87,88,89,9 0,91,98,99,105,107,110,111,113,114,117,118,119,120,122,127,128,132,138,141,146,147,148,150,151,157,161,1 62,163,164, 166,168 Integrated OR 2,3,6,7,8,9,12,13,14,15,16,18,23,24,25,27,28,36,40,43,46,54,58,61,62,65,68,70,71,72,73,75,76,77,78,79,82,86,9 4,96,97,100,106,108,109,115,121,123,126,130,131,136,137,139,140,142,143,145,152,153,154,155,156,167 Table 6: In an integrated OR, supporting facilities such as the ICU, PACU, and wards are considered. The problem of the congested PACU, which previously had been scarcely addressed in the literature, received more attention between 2008 and 2010. In this problem, patients are not allowed to enter the fully occupied PACU and are therefore forced to start their recovery phase in the OR itself, keeping it blocked. Iser et al. [77] used a simulation model to tackle the problem and compare overtime occurring in the OR with PACU specific performance measures. Augusto et al. [7] showed the benefit of using a mathematical model to plan ahead the exact amount of recovery time a patient will spend within the OR. As it is typical for highly utilized systems, there is a sensitive relationship between overall case volume, capacity (of the PACU) and the effect on waiting time (to enter the PACU). This relationship is described in more detail by Schonmeyr et al. [131] using queuing theory. 68 Besides the PACU, a downstream facility, which could affect the function of the OR is the ICU. Kolker [82] reduced diversion of an ICU to an acceptable level by defining the maximum number of elective surgeries per day that are allowed to be scheduled along with the competing demand from emergency arrivals. Litvak et al. [92] went a step further and tackled the ICU capacity problem in a cooperative framework. In their model, several hospitals of a region jointly reserve a small number of beds in order to accommodate emergency patients and achieve an improved service level for all patients. 5.6. Uncertainty One of the major problems associated with the development of accurate OR planning and scheduling strategies is the uncertainty inherent to surgical services. Deterministic planning and scheduling approaches ignore uncertainty, whereas stochastic approaches explicitly try to incorporate it. In Table 7, we list the relevant manuscripts classified according to the type of uncertainty they incorporate. 100% 90% 80% Deterministic 70% 60% Stochastic (arrival + durration + other) 50% Stochastic arrival 40% Stochastic durration 30% 20% Stochastic other 10% 0% 2005 2006 2007 2008 2009 2010 Figure 38: Uncertainty incorporation. Implicit in the figure is the fact that if duration uncertainty is accounted for, it is often the case that arrival uncertainty is considered as well. 69 Deterministic 6,7,10,14,15,18,19,20,24,27,28,33,34,35,43,46,52,55,56,57,58,59,64,68,70,73,75,77,78,79,80,83,84,86,97,99,107,1 08,110, 111,114,115,118,120,121,123,127,128,130,136,137,140,143,144,145,146,150,152,155,156,164,166,167 Stochastic Arrival 3,9,13,16,22,23,25,36,48,49,54,62,65,67,71,72,74,81,82,87,89,90,93,94,100,105,109,117,121,122,126,132,138,139, 141, 142,145,150,154,162,163,167,168 Duration 2,3,5,8,9,11,12,13,16,22,25,36,37,38,39,40,41,49,50,54,62,65,67,68,69,71,72,76,81,87,88,89,91,93,94,96,97,98,100 ,105, 106,109,113,117,119,122,126,131,132,139,141,142,145,148,150,151,153,154,157,161,162,163,167 Other 3,5,9,25,45,46,72,92,94,106,126,162 Table 7: Methods frequently take stochasticity into account. Most common forms are duration and arrival uncertainty. As shown in Figure 38, stochasticity in the form of uncertain patient arrival and surgery duration is frequently used in the OR literature. If we narrow the literature to recent contributions, which explicitly incorporate non-elective patients, we see that over 80% of methods incorporate some sort of uncertainty. Non-elective patient arrivals are in most cases impossible to predict in advance and additionally occupy a random amount of OR time, which often leaves OR managers with no option but to keep a safety margin to accommodate them [141]. In contrast, the arrival of elective patients contains a lower amount of uncertainty, and as shown in Figure 39, is frequently considered as deterministic in the literature. 70 100% 90% 80% Non-elective patient and stochastic duration 70% 60% Non-elective patient and stochastic arrival 50% Elective patient and stochastic duration 40% 30% Elective patient and stochastic arrival 20% 10% 0% 2005 2006 2007 2008 2009 2010 Figure 39: Stochasticity in the elective and non-elective patient setting.3 If uncertainty is considered in an elective or non-elective patient setting, stochastic aspects are in either case slightly more often applied to surgery durations than to patient arrival times. Duration uncertainty is a central element in Denton et al. [39] as well as in Batun et al. [11]. In Denton et al. [39] decisions include the number of ORs to open and surgery block to OR assignments, whereas in Batun et al. [11] this is supplemented by patient sequencing and setting surgeon start times. Both methods aim at minimizing OR opening and OR overtime costs, where Batun et al. [11] additionally consider surgeon idle times. The functional difference between their methods lies in the way surgery to OR assignments are carried out. In Denton et al. [39], the common practice of assigning a surgery block to a single surgeon (block scheduling) is followed, whereas Batun et al. [11] consider the scenario of pooled ORs, and therefore surgeons are allowed to switch between ORs. OR pooling allows carrying out surgeries in parallel as the main surgeon only needs to be present during the critical part of the surgery and can move to the next patient before the close-up of the patient. A timing aspect, which is different from the actual surgery duration but is characterized by large variations, is the patient length of stay (LOS) in e.g. PACU, ICU or ward. The variability of patient time spent in the ward is considered by Ma and Demeulemeester [94] in which the rate of patient misplacements caused by bed shortages is minimized. It should be clear that operations research techniques are able to deal with stochasticity, especially simulation techniques (included in 66% of the stochastic literature) and analytical procedures (included in 22% of the stochastic literature), and that an adequate planning and scheduling approach may lower the negative impact of uncertainty. Mostly, studies assumed a certain level of variability, based on analyzing historical data, and used this information as input for models. However, only limited attention is paid to the reduction of variability within the individual processes. As an example, consider the estimation of surgery durations. Instead of the immediate determination of the distribution of a surgery duration, one should examine whether the population of patients for which the durations 3 Percentages are calculated in ratio to the total number of articles dealing with the respective patient type. 71 are taken into account is truly homogenous. If not, separating the patient population may result in a decreased variability even before the planning and scheduling phase is executed. As the estimation of surgery durations exceeds the scope of this literature review, we do not elaborate further on this issue. 5.7. Operations Research Methodology The literature on OR planning and scheduling exhibits a wide range of methodologies that fit within the domain of operations research and that combine a certain type of analysis with some solution or evaluation technique. Table 8 provides an overview of the ways in which OR planning and scheduling problems are analyzed. The table shows that mathematical programming and heuristics are frequently applied, generally to patient sequencing or assignment type of problems (e.g., patient to surgeon/OR assignment). These types of problems are combinatorial optimization problems. In some approaches the impact of specific changes to the problem setting is examined. We refer to this type of analysis as scenario analysis since multiple scenarios, settings or options are compared to each other with respect to the performance criteria. 72 Analytical procedure 22,37,59,62,65,88,89,92,93,105,113,114,131,141,151,157,162,168 Mathematical Programming Linear Programming 7,37,43,46,84,108,119 Goal Programming 2,18,36,115,140 Integer Programming 19,20,27,33,52,110,130,142,143,144,145,153 Mixed integer programming 11,13,16,39,68,78,79,80,87,88,89,106,107,118,120,121,123,128,167 Column generation 55,57,58,59,68,73,86,87,90,152,153 Branch-and-price 14,28,56 Dynamic programming 6,7,14,28,56,73,90,105,164 Other 6,7,13,16,45,70,99,119 Dedicated branch-and-bound 27,39,111,156 Scenario analysis 3,5,7,8,9,12,15,18,22,23,25,34,35,36,37,38,39,40,41,43,46,48,49,50,51,54,55,57,58,59,61,62,67, 69,71,72,74,75,76,78,81,82,84,86,91,92,93,94,96,97,98,100,106,107,108,109,111,113,115,117,1 18,121,122,126,128,130,131,132,136,138,139,141,142,145,146,148,150,151,152,154,156,157,1 62,163,167 Simulation Discrete-event 3,5,8,9,12,22,23,25,36,48,49,50,54,58,60,61,67,71,72,74,76,77,81,82,92,93,94,96,97,98,100,106 ,109,121,122,126,132,139,142,145,148,150,154,163,167 Monte-Carlo 22,40,46,69,87,88,89,91,111,117 Heuristics improvement heuristics Simulated annealing 13,16,40,69,88,150 Tabu search 35,73,75,88 Genetic algorithm 34,58,72,73,127,128,136,137,147,161,166 other 19,20,38,69,87,88,90,98,106,138,150,164 constructive heuristics 5,6,13,16,33,37,38,39,49,50,51,55,58,59,64,69,76,77,83,86,87,88,90,126,142,143,150 73 Table 8: Different solution techniques are used in the literature: Analytical procedures (e.g., queuing theory or new vendor model), mathematical programming, dedicated branch-and-bound, scenario analysis (or sensitivity analysis), simulation, and heuristics. As is shown by Figure 40, performing scenario analysis is popular, especially in the discrete-event simulation (DES) modeling literature. Scenario analysis can be done by plotting the efficiency frontier formed by respective scenarios’ (possibly multidimensional) performance scores. This helps to identify and distinguish between advantageous and disadvantageous scenarios. The performance criteria most frequently used in the DES modeling literature are patient waiting time and different kinds of utilization measures such as the overtime related to the OR, ward, ICU or PACU. 100% 90% 80% 70% Analytical 60% Mathematical programming 50% Scenario or sensitivity analysis 40% Discrete-event simulation 30% Heuristics 20% 10% 0% 2005 2006 2007 2008 2009 2010 Figure 40: Only selected solution techniques are shown. Most articles include a scenario or sensitivity analysis. In Section 4, we expressed our surprise about the lack of an increasing use of integrated approaches. In the DES modeling literature, however, the proportion of integrated approaches does increase, as OR supporting facilities such as the PACU or ward are increasingly taken into account. We think that modeling the OR in an integrated way is an important step towards the construction of more realistic and applicable models. An integrated DES model is introduced by Steins et al. [139], in which pre-operative care as well as a PACU are considered. The arrival of case types, the surgery time and the LOS in the PACU are represented as probabilistic distributions. The patients in their model are differentiated according to their urgency status, i.e., whether a patient is elective or non-elective. It is true in general that DES modeling approaches take explicit account of non-electives. In the literature, the analytical approach is less often encountered than DES models. Aside from their differences on the methodological side, both DES and analytical methods are often related to a similar problem setting. In case of a stochastic environment where capacity questions have to be answered and non-electives possibly play a role, both of the approaches are useful. Tancrez et al. [141] define 74 the amount of OR capacity, which is needed to accommodate for non-elective patients in a Markovian model setting. Simulation is used to show that the assumptions required to build the Markov chain have a minor influence on their final analytical results. In their work, the stochasticity in OR capacity is the consequence of randomly arriving non-elective patients occupying an uncertain amount of OR time. Also without non-elective patient arrivals it is difficult to predict the required OR capacity on a day, as surgery durations are unknown in advance and can vary considerably in length. In Olivares et al. [113], the decision making process of reserving OR capacity is investigated using the newsvendor model. In the analytical approach, an estimate is given of the cost placed by the hospital on having idle capacity and the cost of a schedule overrun. Their results reveal that the hospital under study places more emphasis on the tangible costs of having idle capacity than on the costs of a schedule overrun and long working hours for the staff. As shown by Figure 40, MPs are popular. As opposed to DES and analytical models, MPs, such as mixed integer programs, deal with combinatorial optimization problems. In the majority of cases (>60%), the objective function of the optimization problem includes under/overtime or under/overutilization. Those performance criteria are rarely used by themselves but are usually part of a multiple objective formulation. The use of multiple objectives in MPs, as is the case in general, is increasingly popular. In 2010, less than 20% of mathematical formulations found in the literature still restrict themselves to a single objective. Their popularity can be explained in two ways. First, the development of better solvers makes it increasingly practical to use them. Second, defining multiple objectives allows capturing stakeholder preferences more realistically. Despite the increasing complexity of the objective functions of MPs, there are no indications that the same would be true in respect to their constraints. In other words, the variety of constraints used in MPs seems to be constant. The most frequently used constraints are resource related, which in many cases relate to the OR (under-, regular or overtime) or medical personnel. Regularly applied constraints, which do not focus on a given resource, are priority constraints (a high priority patient always needs to be served before a low priority patient), demand related constraints (a given specialty needs to be given a certain amount of OR time) and release related (a patient belonging to a given category needs to be served before a given deadline). As in Min and Yih [106], the decisions in most of the mathematical programs apply to the elective patient. In their work, a stochastic mixed integer programming model is proposed and solved by a sampling based approach. The surgery durations, the LOS, the availability of a downstream facility (ICU) and new demand are assumed to be random with known distributions. In some cases mathematical programs are too difficult to solve within a reasonable time limit and therefore heuristics are proposed. In Fei et al. [60] a column generation based heuristic is used to solve the patient scheduling problem. In their setting, a column corresponds to a feasible plan, in other words, the assignment of surgical cases to an OR. Roland et al. [128] propose a method, which includes the assignment of cases to ORs, planning days and operating time periods. The NP-hard problem is tackled by means of a genetic algorithm. Similarly to mathematical programming, also heuristics are in most of the cases used for scheduling tasks involving the elective patient. Noteworthy is that in 2006 all heuristic methods found in the literature were time assignment problems whereas in 2010, as a result of a gradual decrease, this was true for only 20% of the articles as date and room assignment problems become more popular. 75 5.8. Applicability of Research in Practice Many researchers provide a thorough testing phase in which they illustrate the applicability of their research. Whether applicability points at computational efficiency or at showing to what extent objectives may be realized, a substantial amount of data is desired. From Figure 41 and Table 9, we notice that most of this data is on real health care practices. This evolution is noteworthy and results from the improved hospital information systems from which data can be easily extracted. Unfortunately, a single testing of procedures or tools based on real data does not imply that they finally get implemented in practice. Lagergren [85] indicates that the lack of implementation in the health services seems to have improved considerably. Figure 41 shows, however, that only a very small share of the articles report on actual implementation. An exception to this is Wachtel and Dexter [157] who introduce a website, which is used by the hospital under study to decide on the exact times patients have to arrive to their surgery appointment. The problem tackled by the authors arises from the fact that a case is often started earlier than scheduled, but it cannot be known in advance if it will happen or not. Patient availability must therefore be balanced against patient waiting times and fasting times. Daily applicability is entailed by their method. However, there are problems, which have to be solved on a less frequent basis. An example is the application of a case mix model that is applied every year, clearly resulting in a different degree of implementation. A clear comparison of manuscripts on this aspect is hence not straightforward. Even if the implementation of research can be assumed, authors often provide little detail about the process of implementation. Therefore, we encourage the provision of additional information on the behavioral factors that coincide with the actual implementation. Identifying the causes of failure, or the reasons that lead to success, may be of great value to the research community [26]. 100% 90% 80% 70% 60% No testing 50% Theoretic data 40% Based on real data Implemented and applied 30% 20% 10% 0% 2005 2006 2007 2008 2009 2010 Figure 41: Even though most data used in the literature is based on real data, this does not mean that the methods are applied in reality. 76 In many contributions a problem is solved and applied to the problem setting specific to one single hospital and it is unclear whether or to what extent a method is applicable to another setting. In order to justify the generality of their modeling assumptions Schoenmeyr et al. [131] surveyed several hospitals. Introducing generalizable methods makes it easier to spread and implement good working operations research practices to more than one hospital. Only limited research has been done to study which planning and scheduling expertise is currently in use in hospitals. Using a survey, Sieber and Leibundgut [133] reported that the state of OR management in Switzerland is far from excellent. A similar more recent exercise for Flemish (Belgium) hospitals is described in Cardoen et al. [30]. It seems contradictory that so little research is effectively applied in a domain as practical as OR planning and scheduling. No testing 65,161,162 Data for testing theoretic 6,7,13,14,37,49,51,54,55,56,64,73,77,78,79,83,86,87,88,89,90,91,94,96,98,110,111,120,123,136,137,138,143, 164 based on real data 2,3,5,8,9,10,11,12,15,16,22,23,24,25,27,28,33,34,35,36,38,39,40,41,43,45,46,48,50,51,52,57,58,59,60,61,62,6 7,68,69,70,71,72,74,75,76,80,81,82,84,92,93,97,99,100,105,106,107,108,109,111,113,114,115,117,118,119,1 20,121,122,123,126,127,128,130,131,132,139,140,141,144,145,146,147,148,150,151,152,153,154,155,156,15 7,163,166,167,168 Implemented and applied18,19,20,60,70,128,141,144 Table 9: Both theoretic and real data are frequently used for testing purposes. 5.9. Opportunities for Future Research Our review suggests that methods introduced in the literature are rarely implemented at hospitals and, if implemented, details usually remain unpublished. Both the problem of low success rates of implementations and the lack of reports could be mitigated by actively involving surgeons, head nurses and IT personnel as coauthors. In order to avoid developing scheduling software that is only specified to the needs of one hospital, it might be wise that projects cover several hospitals. Including more than one hospital in a study provides, besides generalizability, also other opportunities: resource pooling on the level of emergency ORs, anesthesia rooms, equipment or even nursing staff can lead to an integrated approach, which profits all participating hospitals. Integration is also an important concept within the hospital itself. Considering its importance, it appears there is an opportunity to study the role of supporting facilities such as 77 the anesthesia department, PACU, ICU and/or wards in an integrated way with the OR. Similarly, important as integrality is the incorporation of uncertainty. First and most prominently, uncertainty is accounted for in respect to patient surgery times, which are unknown until surgeries are actually realized. Second, it is unknown whether the operating room will be available at the planned surgery start as an emergency could occupy it. Third, while booking a surgery into a certain slot, it is unknown whether the slot might be needed to allocate a future more urgent patient. Even though we see that uncertainty is frequently incorporated, the question arises whether it should be a prerequisite for an algorithm to take aspects of uncertainty into account, i.e., is a strictly deterministic scheduling approach able to provide the robustness required in reality? These and other open questions present many opportunities for future research related to OR planning and scheduling. 5.10. Conclusion In this chapter we have studied and described recent trends in the field of OR planning and scheduling. Based on the data, we found that most attention is given to elective patients, and even though often not stated explicitly, it is in many cases implied to be an inpatient setting. Less frequently occurring in the literature are methods which consider outpatients. This is surprising as outpatient care is gaining in importance and therefore we would expect to observe an increasing amount of literature dealing in this area. We also observed a gradual shift from determining the exact time of a surgery to problems related to date and/or room assignments. With respect to the performance measures considered, we found that overtime is the most popular measure and that preference related criteria are gaining popularity. Noteworthy is the fact that preferences are increasingly used in multi-criteria settings. Mathematical programming, DES and Heuristics are the techniques most frequently used. It is also true that the majority of articles present results based on real data. However, it is important to note that this does not imply that the 5.11. References 1. Achieving operating room efficiency through process integration (2003). Healthc Financ Manage 57 (3):supl. 1supl. 8 78 2. Adan I, Bekkers J, Dellaert N, Vissers J, Yu XT (2009) Patient mix optimisation and stochastic resource requirements: A case study in cardiothoracic surgery planning. Health Care Management Science, vol 12. doi:DOI 10.1007/s10729-008-9080-9 3. Antonelli D, Taurino T (2010) Application of a patient flow model to a surgery department. 2010 IEEE Workshop on Health Care Management (WHCM):6 pp. doi:10.1109/whcm.2010.5441247 4. Argo JL, Vick CC, Graham LA, Itani KMF, Bishop MJ, Hawn MT (2009) Elective surgical case cancellation in the veterans health administration system: Identifying areas for improvement. Am J Surg 198 (5):600-606. doi:10.1016/j.amjsurg.2009.07.005 5. Arnaout JPM, Kulbashian S (2008) Maximizing the utilization of operating rooms with stochastic times using simulation. 2008 Winter Simulation Conference, Vols 1-5:1617-1623 6. Augusto V, Xie X, Perdomo V (2008) Operating theatre scheduling using lagrangian relaxation. Eur J Ind Eng 2 (2):172-189 7. Augusto V, Xie XL, Perdomo V (2010) Operating theatre scheduling with patient recovery in both operating rooms and recovery beds. Comput Ind Eng 58 (2):231-238. doi:10.1016/j.cie.2009.04.019 8. Ballard SM, Kuhl ME (2006) The use of simulation to determine maximum capacity in the surgical suite operating room. Proceedings of the 2006 Winter Simulation Conference, Vols 1-5:433-438 9. Barkaoui K, Dechambre P, Hachicha R (2002) Verification and optimisation of an operating room workflow. Proceedings of the 35th Annual Hawaii International Conference on System Sciences:2581-25902590. doi:10.1109/hicss.2002.994236 10. Basson MD, Butler T (2006) Evaluation of operating room suite efficiency in the veterans health administration system by using data-envelopment analysis. Am J Surg 192 (5):649-656. doi:10.1016/j.amjsurg.2006.07.005 11. Batun S, Denton BT, Huschka TR, Schaefer AJ (2010) Operating room pooling and parallel surgery processing under uncertainty. Informs J Comput:ijoc.1100.0396. doi:10.1287/ijoc.1100.0396 12. Baumgart A, Zoeller A, Denz C, Bender HJ, Heinzl A, Badreddin E (2007) Using computer simulation in operating room management: Impacts on process engineering and performance. Proceedings of the 40th Annual Hawaii International Conference on System Sciences:10 pp.-10 pp.10 pp. 13. Belien J, Demeulemeester E (2007) Building cyclic master surgery schedules with leveled resulting bed occupancy. Eur J Oper Res 176 (2):1185-1204. doi:DOI 10.1016/j.ejor.2005.06.063 14. Belien J, Demeulemeester E (2008) A branch-and-price approach for integrating nurse and surgery scheduling. Eur J Oper Res 189 (3):652-668. doi:10.1016/j.ejor.2006.10.060 15. Belien J, Demeulemeester E, Cardoen B (2006) Visualizing the demand for various resources as a function of the master surgery schedule: A case study. J Med Syst 30 (5):343-350 16. Belien J, Demeulemeester E, Cardoen B (2009) A decision support system for cyclic master surgery scheduling with multiple objectives. J Sched 12 (2):147-161. doi:10.1007/s10951-008-0086-4 17. Blake JT, Carter MW (1997) Surgical process scheduling: A structured review. J Soc Health Syst 5 (3):17-30 18. Blake JT, Carter MW (2002) A goal programming approach to strategic resource allocation in acute care hospitals. Eur J Oper Res 140 (3):541-561 19. Blake JT, Dexter F, Donald J (2002) Operating room managers' use of integer programming for assigning block time to surgical groups: A case study. Anesth Analg 94 (1):143-148 20. Blake JT, Donald J (2002) Mount sinai hospital uses integer programming to allocate operating room time. Interfaces 32 (2):63-73 21. Boldy D (1976) Review of application of mathematical-programming to tactical and strategic health and socialservices problems. Operational Research Quarterly 27 (2):439-448 22. Bowers J, Mould G (2004) Managing uncertainty in orthopaedic trauma theatres. Eur J Oper Res 154 (3):599608. doi:10.1016/S0377-2217(02)00816-0 23. Bowers J, Mould G (2005) Ambulatory care and orthopaedic capacity planning. Health Care Manag Sci 8 (1):4147 24. Calichman MV (2005) Creating an optimal operating room schedule. AORN J 81 (3):580-588 79 25. Cardoen B, Demeulemeester E (2008) Capacity of clinical pathways - a strategic multi-level evaluation tool. J Med Syst 32 (6):443-452. doi:10.1007/s10916-008-9150-z 26. Cardoen B, Demeulemeester E (2011) A decision support system for surgery sequencing at uz leuven's day-care department. Int J Inf Tech Decis 10 (3):435-450. doi:10.1142/S0219622011004397 27. Cardoen B, Demeulemeester E, Belien J (2009) Optimizing a multiple objective surgical case sequencing problem. Int J Prod Econ 119 (2):354-366. doi:10.1016/j.ijpe.2009.03.009 28. Cardoen B, Demeulemeester E, Belien J (2009) Sequencing surgical cases in a day-care environment: An exact branch-and-price approach. Comput Oper Res 36 (9):2660-2669. doi:10.1016/j.cor.2008.11.012 29. Cardoen B, Demeulemeester E, Belien J (2010) Operating room planning and scheduling: A literature review. Eur J Oper Res 201 (3):921-932. doi:10.1016/j.ejor.2009.04.011 30. Cardoen B, Demeulemeester E, Van der Hoeven J (2010) On the use of planning models in the operating theatre: Results of a survey in flanders. The International Journal of Health Planning and Management 25 (4):400-414. doi:10.1002/hpm.1027 31. Centers for Medicare & Medicaid Services OotA (2010) National health expenditures and selected economic indicators, levels and annual percent change: Calendar years 2004-2019. 32. Centers for Medicare & Medicaid Services OotA, National Health Statistics Group (2009) The nation’s health dollar, calendar year 2009: Where it went. 33. Chaabane S, Meskens N, Guinet A, Laurent M (2006) Comparison of two methods of operating theatre planning: Application in belgian hospital. 2006 International Conference on Service Systems and Service Management, Vols 1 and 2, Proceedings:386-392 34. Conforti D, Guerriero F, Guido R (2010) A multi-objective block scheduling model for the management of surgical operating rooms: New solution approaches via genetic algorithms. 2010 IEEE Workshop on Health Care Management (WHCM):5 pp. doi:10.1109/whcm.2010.5441264 35. Dekhici L, Belkadi K (2010) Operating theatre scheduling under constraints. Journal of Applied Sciences:13801388 36. Dellaert N, Jeunet J (2008) Hospital admission planning to optimize major resources utilization under uncertainty. Paper presented at the 3rd World Conference on Production and Operations Management, 37. Denton B, Gupta D (2003) A sequential bounding approach for optimal appointment scheduling. Iie Trans 35 (11):1003-1016. doi:10.1080/07408170390230169 38. Denton B, Viapiano J, Vogl A (2007) Optimization of surgery sequencing and scheduling decisions under uncertainty. Health Care Manag Sci 10 (1):13-24 39. Denton BT, Miller AJ, Balasubramanian HJ, Huschka TR (2010) Optimal allocation of surgery blocks to operating rooms under uncertainty. Operations Research 58 (4):802-816. doi:DOI 10.1287/opre.1090.0791 40. Denton BT, Rahman AS, Nelson H, Bailey AC (2006) Simulation of a multiple operating room surgical suite. Proceedings of the 2006 Winter Simulation Conference, Vols 1-5:414-424 41. Dexter F (2000) A strategy to decide whether to move the last case of the day in an operating room to another empty operating room to decrease overtime labor costs. Anesth Analg 91 (4):925-928 42. Dexter F, Birchansky L, Bernstein JM, Wachtel RE (2009) Case scheduling preferences of one surgeon's cataract surgery patients. Anesth Analg 108 (2):579-582. doi:DOI 10.1213/ane.0b013e31818f1651 43. Dexter F, Blake JT, Penning DH, Sloan B, Chung P, Lubarsky DA (2002) Use of linear programming to estimate impact of changes in a hospital's operating room time allocation on perioperative variable costs. Anesthesiology 96 (3):718-724 44. Dexter F, Epstein RH (2009) Typical savings from each minute reduction in tardy first case of the day starts. Anesth Analg 108 (4):1262-1267. doi:DOI 10.1213/ane.0b013e31819775cd 45. Dexter F, Ledolter J, Wachtel RE (2005) Tactical decision making for selective expansion of operating room resources incorporating financial criteria and uncertainty in subspecialties' future workloads. Anesth Analg 100 (5):1425-1432. doi:10.1213/01.Ane.0000149898.45044.3d 80 46. Dexter F, Lubarsky DA, Blake JT (2002) Sampling error can significantly affect measured hospital financial performance of surgeons and resulting operating room time allocations. Anesth Analg 95 (1):184-188. doi:10.1213/01.Ane.0000018821.20416.5a 47. Dexter F, Macario A, Ledolter J (2007) Identification of systematic underestimation (bias) of case durations during case scheduling would not markedly reduce overutilized operating room time. J Clin Anesth 19 (3):198-203. doi:DOI 10.1016/j.jclinane.2006.10.009 48. Dexter F, Macario A, Lubarsky DA (2001) The impact on revenue of increasing patient volume at surgical suites with relatively high operating room utilization. Anesth Analg 92 (5):1215-1221 49. Dexter F, Macario A, O'Neill L (2000) Scheduling surgical cases into overflow block time - computer simulation of the effects of scheduling strategies on operating room labor costs. Anesth Analg 90 (4):980-988 50. Dexter F, Traub RD (2002) How to schedule elective surgical cases into specific operating rooms to maximize the efficiency of use of operating room time. Anesth Analg 94 (4):933-942 51. Dexter F, Traub RD, Macario A (2003) How to release allocated operating room time to increase efficiency: Predicting which surgical service will have the most underutilized operating room time. Anesth Analg 96 (2):507512. doi:10.1213/01.ane.0000042450.45778.ab 52. Dexter F, Wachtel RE, Epstein RH, Ledolter J, Todd MM (2010) Analysis of operating room allocations to optimize scheduling of specialty rotations for anesthesia trainees. Anesth Analg 111 (2):520-524. doi:Doi 10.1213/Ane.0b013e3181e2fe5b 53. Does R, Vermaat TMB, Verver JPS, Bisgaard S, Van den Heuvel J (2009) Reducing start time delays in operating rooms. Journal of Quality Technology 41 (1):95-109 54. Everett JE (2002) A decision support simulation model for the management of an elective surgery waiting system. Health Care Manag Sc 5 (2):89-95. doi:10.1023/a:1014468613635 55. Fei H, Chu C, Meskens N (2009) Solving a tactical operating room planning problem by a column-generationbased heuristic procedure with four criteria. Ann Oper Res 166 (1):91-108. doi:DOI 10.1007/s10479-008-0413-3 56. Fei H, Chu C, Meskens N, Artiba A (2008) Solving surgical cases assignment problem by a branch-and-price approach. Int J Prod Econ 112 (1):96-108. doi:10.1016/j.ijpe.2006.08.030 57. Fei H, Combes C, Chu C, Meskens N (2006) Endoscopies scheduling problem: A case study. Paper presented at the INCOM 06, 2006-05-17 58. Fei H, Meskens N, Chu C (2007) An operating theatre planning and scheduling problem in the case of a "block scheduling" strategy. 2006 International Conference on Service Systems and Service Management (IEEE Cat No06EX1406):7 pp.|CD-ROM 59. Fei H, Meskens N, Chu C (2010) A planning and scheduling problem for an operating theatre using an open scheduling strategy. Comput Ind Eng 58 (2):221-230. doi:10.1016/j.cie.2009.02.012 60. Fei HY, Meskens N, Combes C, Chu CB (2009) The endoscopy scheduling problem: A case study with two specialised operating rooms. Int J Prod Econ 120 (2):452-462. doi:10.1016/j.ijpe.2007.09.016 61. Ferrand Y, Magazine M, Rao U (2010) Comparing two operating-room-allocation policies for elective and emergency surgeries. 2010 Winter Simulation Conference (WSC 2010):2364-2374. doi:10.1109/wsc.2010.5678933 62. Ferrin DM, Miller MJ, Wininger S, Neuendorf MS (2004) Analyzing incentives and scheduling in a major metropolitan hospital operating room through simulation. Proceedings of the 2004 Winter Simulation Conference, Vols 1 and 2:1975-1980 63. Gonzalez P, Herrero C (2004) Optimal sharing of surgical costs in the presence of queues. Mathematical Methods of Operations Research 59 (3):435-446. doi:10.1007/s001860400350 64. Guerriero F, Guido R (2011) Operational research in the management of the operating theatre: A survey. Health Care Manag Sc 14 (1):89-114. doi:10.1007/s10729-010-9143-6 65. Guinet A, Chaabane S (2003) Operating theatre planning. Int J Prod Econ 85 (1):69-81. doi:Doi 10.1016/S09255273(03)00087-2 66. Gupta D (2007) Surgical suites' operations management. Production and Operations Management 16 (6):689700. doi:10.1111/j.1937-5956.2007.tb00289.x 81 67. Gupta D, Denton B (2008) Appointment scheduling in health care: Challenges and opportunities. Iie Trans 40 (9):800-819. doi:10.1080/07408170802165880 68. Gupta D, Natarajan MK, Gafni A, Wang L, Shilton D, Holder D, Yusuf S (2007) Capacity planning for cardiac catheterization: A case study. Health Policy 82 (1):1-11. doi:DOI 10.1016/j.healthpol.2006.07.010 69. Hans E, Nieberg T, van Oostrum JM (2007) Optimization in surgery planning. Medium Econometrische Toepassingen 15 70. Hans E, Wullink G, van Houdenhoven M, Kazemier G (2008) Robust surgery loading. Eur J Oper Res 185 (3):10381050. doi:10.1016/j.ejor.2006.08.022 71. Hanset A, Meskens N, Duvivier D (2010) Using constraint programming to schedule an operating theatre. 2010 IEEE Workshop on Health Care Management (WHCM):6 pp. doi:10.1109/whcm.2010.5441245 72. Harper PR (2002) A framework for operational modelling of hospital resources. Health Care Manag Sc 5 (3):165173. doi:10.1023/a:1019767900627 73. Helm JE, Lapp M, See BD (2010) Characterizing an effective hospital admissions scheduling and control management system: A genetic algorithm approach. 2010 Winter Simulation Conference (WSC 2010):2387-2398. doi:10.1109/wsc.2010.5678935 74. Hongying F, Meskens N, El-Darzi E (2010) Evaluating alternative surgery plans with discrete-event simulation model. 2010 IEEE Workshop on Health Care Management (WHCM):6 pp. doi:10.1109/whcm.2010.5441241 75. Hsu VN, de Matta R, Lee CY (2003) Scheduling patients in an ambulatory surgical center. Naval Research Logistics 50 (3):218-238. doi:10.1002/nav.10060 76. Huschka TR, Denton BT, Gul S, Fowler JW (2007) Bi-criteria evaluation of an outpatient procedure center via simulation. Proceedings of the 2007 Winter Simulation Conference, Vols 1-5:1489-1497 77. Iser JH, Denton BT, King RE (2008) Heuristics for balancing operating room and post-anesthesia resources under uncertainty. 2008 Winter Simulation Conference, Vols 1-5:1601-1608 78. Jebali A, Alouane ABH, Ladet P (2006) Operating rooms scheduling. Int J Prod Econ 99 (1-2):52-62. doi:10.1016/j.ijpe.2004.12.006 79. Jebali A, Hadj-Alouane AB, Ladet P Performance comparison of two strategies for operating room scheduling. In: International Symposium on Computational Intelligence and Intelligent Informatics, 2003. 80. Kharraja S, Albert P, Chaabane S (2006) Block scheduling: Toward a master surgical schedule. 2006 International Conference on Service Systems and Service Management, Vols 1 and 2, Proceedings:429-435 81. Kim SC, Horowitz I (2002) Scheduling hospital services: The efficacy of elective-surgery quotas. OmegaInternational Journal of Management Science 30 (5):335-346 82. Kolker A (2009) Process modeling of icu patient flow: Effect of daily load leveling of elective surgeries on icu diversion. J Med Syst 33 (1):27-40. doi:10.1007/s10916-008-9161-9 83. Krempels K-H, Panchenko A An approach for automated surgery scheduling. In: Sixth International Conference on the Practice and Theory of Automated Timetabling, Brno, Czech Republic, 30th August – 1st September 2006 2006. 84. Kuo PC, Schroeder RA, Mahaffey S, Bollinger RR (2003) Optimization of operating room allocation using linear programming techniques. J Am Coll Surgeons 197 (6):889-895. doi:10.1016/j.jamcolsurg.2003.07.006 85. Lagergren M (1998) What is the role and contribution of models to management and research in the health services? A view from europe. Eur J Oper Res 105 (2):257-266. doi:10.1016/s0377-2217(97)00233-6 86. Lamiri M, Augusto V, Xie XL (2008) Patients scheduling in a hospital operating theatre. 2008 IEEE International Conference on Automation Science and Engineering, Vols 1 and 2:627-632 87. Lamiri M, Dreo J, Xiaolan X (2007) Operating room planning with random surgery times. Proceedings of the 3rd Annual IEEE Conference on Automation Science and Engineering:521-526 88. Lamiri M, Grimaud F, Xie XL (2009) Optimization methods for a stochastic surgery planning problem. Int J Prod Econ 120 (2):400-410. doi:10.1016/j.ijpe.2008.11.021 89. Lamiri M, Xie XL, Dolgui A, Grimaud F (2008) A stochastic model for operating room planning with elective and emergency demand for surgery. Eur J Oper Res 185 (3):1026-1037. doi:10.1016/j.ejor.2006.02.057 82 90. Lamiri M, Xie XL, Zhang SG (2008) Column generation approach to operating theater planning with elective and emergency patients. Iie Trans 40 (9):838-852. doi:Doi 10.1080/07408170802165831 91. Lebowitz P (2003) Schedule the short procedure first to improve or efficiency. AORN J 78 (4):651-654 92. Litvak N, van Rijsbergen M, Boucherie RJ, van Houdenhoven M (2008) Managing the overflow of intensive care patients. Eur J Oper Res 185 (3):998-1010. doi:DOI 10.1016/j.ejor.2006.08.021 93. Lovejoy WS, Li Y (2002) Hospital operating room capacity expansion. Management Science 48 (11):1369-1387 94. Ma G, Demeulemeester E (2010) Assessing the performance of hospital capacity planning through simulation analysis. KBI_1031 edn. Katholieke Universiteit Leuven, Leuven 95. Magerlein JM, Martin JB (1978) Surgical demand scheduling - review. Health Services Research 13 (4):418-433 96. Marcon E, Dexter F (2006) Impact of surgical sequencing on post anesthesia care unit staffing. Health Care Manag Sci 9 (1):87-98 97. Marcon E, Dexter F (2007) An observational study of surgeons' sequencing of cases and its impact on postanesthesia care unit and holding area staffing requirements at hospitals. Anesth Analg 105 (1):119-126. doi:10.1213/01.ane.0000266495.79050.b0 98. Marcon E, Kharraja S, Simonnet G (2003) The operating theatre planning by the follow-up of the risk of no realization. Int J Prod Econ 85 (1):83-90. doi:Doi 10.1016/S0925-5273(03)00088-4 99. Marcon E, Kharraja S, Smolski N, Luquet B, Viale JP (2003) Determining the number of beds in the postanesthesia care unit: A computer simulation flow approach. Anesth Analg 96 (5):1415-1423. doi:10.1213/01.ane.0000056701.08350.b9 100. Marjamaa RA, Torkki PM, Hirvensalo EJ, Kirvela OA (2009) What is the best workflow for an operating room? A simulation study of five scenarios. Health Care Manag Sc 12 (2):142-146. doi:10.1007/s10729-008-9073-8 101. Masursky D, Dexter F, O'Leary CE, Applegeet C, Nussmeier NA (2008) Long-term forecasting of anesthesia workload in operating rooms from changes in a hospital's local population can be inaccurate. Anesth Analg 106 (4):1223-1231. doi:DOI 10.1213/ane.0b013e318167906c 102. McIntosh C, Dexter F, Epstein RH (2006) The impact of service-specific staffing, case scheduling, turnovers, and first-case starts on anesthesia group and operating room productivity: A tutorial using data from an australian hospital. Anesth Analg 103 (6):1499-1516. doi:DOI 10.1213/01.ane.0000244535.54710.28 103. Medpac (2010) Report to congress: Medicare payment policy. Medicare Payment Advisory Commission, Washington, DC 104. Milliman (2011) 2011 milliman medical index. Milliman Inc., 105. Min DK, Yih Y (2010) An elective surgery scheduling problem considering patient priority. Comput Oper Res 37 (6):1091-1099. doi:10.1016/j.cor.2009.09.016 106. Min DK, Yih Y (2010) Scheduling elective surgery under uncertainty and downstream capacity constraints. Eur J Oper Res 206 (3):642-652. doi:10.1016/j.ejor.2010.03.014 107. Molina JM, Framinan JM Testing planning policies for solving the elective case scheduling phase: A real application. In: 35th international conference on operational research applied to health services, Leuven, Belgium, 2009. 108. Mulholland MW, Abrahamse P, Bahl V (2005) Linear programming to optimize performance in a department of surgery. J Am Coll Surgeons 200 (6):861-868. doi:10.1016/j.jamcollsurg.2005.01.001 109. Niu Q, Peng Q, ElMekkawy T, Tan YY Performance analysis of the operating room using simulation. In: CDEN and CCEE Conference, Manitoba, Canada, July 22 - 24, 2007 2007. 110. Nouaouri I, Nicolas JC, Jolly D (2009) Scheduling of stabilization surgical cares in case of a disaster. Proceedings of the 2009 IEEE International Conference on Industrial Engineering and Engineering Management (IEEM 2009):1974-1978. doi:10.1109/ieem.2009.5373510 111. Noyan Ogulata S, Erol R (2003) A hierarchical multiple criteria mathematical programming approach for scheduling general surgery operations in large hospitals. J Med Syst 27 (3):259-270. doi:10.1023/a:1022575412017 112. OECD (2011) Total expenditure on health 2011. The Organisation for Economic Co-operation and Development, 83 113. Olivares M, Terwiesch C, Cassorla L (2008) Structural estimation of the newsvendor model: An application to reserving operating room time. Management Science 54 (1):41-55. doi:10.1287/mnsc.1070.0756 114. Oostrum JMv, Parlevliet T, Wagelmans APM, Kazemier G (2008) A method for clustering surgical cases to allow master surgical scheduling. Erasmus University Rotterdam, Econometric Institute, 115. Ozkarahan I (2000) Allocation of surgeries to operating rooms by goal programing. J Med Syst 24 (6):339-378. doi:10.1023/a:1005548727003 116. Pandit JJ, Dexter F (2009) Lack of sensitivity of staffing for 8-hour sessions to standard deviation in daily actual hours of operating room time used for surgeons with long queues. Anesth Analg 108 (6):1910-1915. doi:DOI 10.1213/ane.0b013e31819fe7a4 117. Paoletti X, Marty J (2007) Consequences of running more operating theatres than anaesthetists to staff them: A stochastic simulation study. Brit J Anaesth 98 (4):462-469. doi:Doi 10.1093/Bja/Aem003 118. Pariente JMM, Torres JMF, Cia TG (2009) Policies and decision models for solving elective case operating room scheduling. Cie: 2009 International Conference on Computers and Industrial Engineering, Vols 1-3:112-117 119. Pérez Gladish B, Arenas Parra M, Bilbao Terol A, RodrIguez UrIa MV (2005) Management of surgical waiting lists through a possibilistic linear multiobjective programming problem. Applied Mathematics and Computation 167 (1):477-495. doi:10.1016/j.amc.2004.07.015 120. Persson M, Persson JA Optimization modelling of hospital operating room planning: Analyszing strategies and problem settings. In: Annual Conference of OR Applied to Health Services, Worclaw, Poland, July 23 - July 28 2006. 121. Persson M, Persson JA (2009) Health economic modeling to support surgery management at a swedish hospital. Omega-International Journal of Management Science 37 (4):853-863. doi:10.1016/j.omega.2008.05.007 122. Persson MJ, Persson JA (2010) Analysing management policies for operating room planning using simulation. Health Care Manag Sc 13 (2):182-191. doi:10.1007/s10729-009-9122-y 123. Pham DN, Klinkert A (2008) Surgical case scheduling as a generalized job shop scheduling problem. Eur J Oper Res 185 (3):1011-1025. doi:10.1016/j.ejor.2006.03.059 124. Pierskalla WP, Brailer DJ (1994) Applications of operations research in health care delivery. In: Operations research and the public sector. pp 469--505 125. Przasnyski ZH (1986) Operating room scheduling. A literature review. AORN J 44 (1):67-79 126. Ramis FJ, Palma JL, Baesler FF (2001) The use of simulation for process improvement at an ambulatory surgery center. Proceeding of the 2001 Winter Simulation Conference (Cat No01CH37304):1401-1404 vol.1402|1402 vol.(xxxiii+xx+1678). doi:10.1109/wsc.2001.977462 127. Roland B, Di Martinelly C, Riane F (2006) Operating theatre optimization: A resource-constrained based solving approach. 2006 International Conference on Service Systems and Service Management, Vols 1 and 2, Proceedings:443-448 128. Roland B, Di Martinelly C, Riane F, Pochet Y (2010) Scheduling an operating theatre under human resource constraints. Comput Ind Eng 58 (2):212-220. doi:10.1016/j.cie.2009.01.005 129. Ruey-Kei C, Yu-Chen Y (2010) Fuzzy-based dynamic scheduling system for health examination. 2010 International Conference on Machine Learning and Cybernetics (ICMLC 2010):636-641. doi:10.1109/icmlc.2010.5580551 130. Santibanez P, Begen M, Atkins D (2007) Surgical block scheduling in a system of hospitals: An application to resource and wait list management in a british columbia health authority. Health Care Manag Sci 10 (3):269-282 131. Schoenmeyr T, Dunn PF, Garnarnik D, Levi R, Berger DL, Daily BJ, Levine WC, Sandberg WS (2009) A model for understanding the impacts of demand and capacity on waiting time to enter a congested recovery room. Anesthesiology 110 (6):1293-1304 132. Sciomachen A, Tanfani E, Testi A (2005) Simulation models for optimal schedules of operating theatres. International Journal of Simulation 6:26-34 133. Sieber TJ, Leibundgut DL (2002) Operating room management and strategies in switzerland: Results of a survey. Eur J Anaesth 19 (06):415-423. doi:doi:10.1017/S0265021502000662 134. Slack N (1999) The blackwell encyclopedic dictionary of operations management. Wiley, 84 135. Smith-Daniels VL, Schweikhart SB, Smith-Daniels DE (1988) Capacity management in health-care services review and future-research directions. Decision Sci 19 (4):889-919 136. Souki M, Ben Youssef S, Rebai A (2009) Memetic algorithm for operating room admissions. Cie: 2009 International Conference on Computers and Industrial Engineering, Vols 1-3:519-524 137. Souki M, Rebai A (2009) Memetic differential evolution algorithm for operating room scheduling. Cie: 2009 International Conference on Computers and Industrial Engineering, Vols 1-3:845-850 138. Stanciu A, Vargas L, May J (2010) A revenue management approach for managing operating room capacity. 2010 Winter Simulation Conference (WSC 2010):2444-2454. doi:10.1109/wsc.2010.5678940 139. Steins K, Persson F, Holmer M (2010) Increasing utilization in a hospital operating department using simulation modeling. Simul-T Soc Mod Sim 86 (8-9):463-480. doi:10.1177/0037549709359355 140. Tan YY, ElMekkawy TY, Peng Q, Oppenheimer L Mathematical programming for the scheduling of elective patients in the operating room department. In: CDEN/C²E² Winnipeg, 2007. 141. Tancrez J-S, Roland B, Cordier J-P, Riane F (2009) How stochasticity and emergencies disrupt the surgical schedule. In: McClean S, Millard P, El-Darzi E, Nugent C (eds) Intelligent patient management, vol 189. Studies in computational intelligence. Springer Berlin / Heidelberg, pp 221-239. doi:10.1007/978-3-642-00179-6_14 142. Tanfani E, Testi A (2010) Improving surgery department performance via simulation and optimization. 2010 IEEE Workshop on Health Care Management (WHCM):6 pp. doi:10.1109/whcm.2010.5441255 143. Tanfani E, Testi A (2010) A pre-assignment heuristic algorithm for the master surgical schedule problem (mssp). Ann Oper Res 178 (1):105-119. doi:DOI 10.1007/s10479-009-0568-6 144. Testi A, Tanfani E (2009) Tactical and operational decisions for operating room planning: Efficiency and welfare implications. Health Care Manag Sc 12 (4):363-373. doi:10.1007/s10729-008-9093-4 145. Testi A, Tanfani E, Torre G (2007) A three-phase approach for operating theatre schedules. Health Care Manag Sci 10 (2):163-172 146. Testi A, Tanfani E, Valente R, Ansaldo G, Torre G (2008) Prioritizing surgical waiting lists. J Eval Clin Pract 14 (1):59 - 64 147. Tsoy G, Arnaout JP, Smith T, Rabadi G (2004) A genetic algorithm approach for surgery operating rooms scheduling problem. Managing in a Dangerous World:299-304 148. Tyler DC, Pasquariello CA, Chen CH (2003) Determining optimum operating room utilization. Anesth Analg 96 (4):1114-1121. doi:10.1213/01.Ane.0000050561.41552.A6 149. Value in health care: Current state and future directions (2011). The Healthcare Financial Management Association, westchester, illinois 150. van der Lans M, Hans E, Hurink JL, Wullink G, van Houdenhoven M, Kazemier G (2006) Anticipating urgent surgery in operating room departments. BETA working paper WP-158. University of Twente 151. Van Houdenhoven M, Hans E, Klein J, Wullink G, Kazemier G (2007) A norm utilisation for scarce hospital resources: Evidence from operating rooms in a dutch university hospital. J Med Syst 31 (4):231-236. doi:10.1007/s10916-007-9060-5 152. Van Houdenhoven M, Oostrum JMV, Wullink G, Hans E, Hurink JL, Bakker J, Kazemier G (2008) Fewer intensive care unit refusals and a higher capacity utilization by using a cyclic surgical case schedule. J Crit Care 23 (2):222-226. doi:10.1016/j.jcrc.2007.07.002 153. van Oostrum JM, Van Houdenhoven M, Hurink JL, Hans EW, Wullink G, Kazemier G (2008) A master surgical scheduling approach for cyclic scheduling in operating room departments. Or Spectrum 30 (2):355-374. doi:DOI 10.1007/s00291-006-0068-x 154. VanBerkel PT, Blake JT (2007) A comprehensive simulation for wait time reduction and capacity planning applied in general surgery. Health Care Manag Sci 10 (4):373-385 155. Velasquez R, Melo T, Kufer KH (2008) Tactical operating theatre scheduling: Efficient appointment assignment. Operations Research Proceedings 2007:303-308 156. Vissers JMH, Adan IJBF, Bekkers JA (2005) Patient mix optimization in tactical cardiothoracic surgery planning: A case study. IMA J Manag Math 16 (3):281-304304. doi:10.1093/imaman/dpi023 85 157. Wachtel RE, Dexter F (2007) A simple method for deciding when patients should be ready on the day of surgery without procedure-specific data. Anesth Analg 105 (1):127-140. doi:DOI 10.1213/01.ane.0000266468.09733.4d 158. Wachtel RE, Dexter F (2008) Tactical increases in operating room block time for capacity planning should not be based on utilization. Anesth Analg 106 (1):215-226. doi:106/1/215 [pii] 159. Wachtel RE, Dexter F (2009) Influence of the operating room schedule on tardiness from scheduled start times. Anesth Analg 108 (6):1889-1901. doi:DOI 10.1213/ane.0b013e31819f9f0c 160. Wachtel RE, Dexter F (2009) Reducing tardiness from scheduled start times by making adjustments to the operating room schedule. Anesth Analg 108 (6):1902-1909. doi:DOI 10.1213/ane.0b013e31819f9fd2 161. Wang D, Xu JP (2008) A fuzzy multi-objective optimizing scheduling for operation room in hospital. In C Ind Eng Eng Man:614-618 162. Wang QN (2004) Modeling and analysis of high risk patient queues. Eur J Oper Res 155 (2):502-515. doi:10.1016/s0377-2217(02)00916-5 163. Wullink G, Van Houdenhoven M, Hans EW, van Oostrum JM, van der Lans M, Kazemier G (2007) Closing emergency operating rooms improves efficiency. J Med Syst 31 (6):543-546. doi:DOI 10.1007/s10916-007-9096-6 164. Ya L, Chengbin C, Kanliang W (2010) Aggregated state dynamic programming for operating theater planning. 2010 IEEE International Conference on Automation Science and Engineering (CASE 2010):1013-1018. doi:10.1109/coase.2010.5584004 165. Yang Y, Sullivan KM, Wang PP, Naidu KD (2000) Applications of computer simulation in medical scheduling. Proceedings of the Fifth Joint Conference on Information Sciences, Vols 1 and 2:A836-A841 166. Yu W, Jiafu T, Gang Q (2010) A genetic algorithm for solving patient-priority-based elective surgery scheduling problem. Life System Modeling and Intelligent Computing International Conference on Life System Modeling and Simulation, LSMS 2010, and International Conference on Intelligent Computing for Sustainable Energy and Environment, ICSEE 2010:297-304|xvi+518. doi:10.1007/978-3-642-15597-0_33 167. Zhang B, Murali P, Dessouky MM, Belson D (2009) A mixed integer programming approach for allocating operating room capacity. J Oper Res Soc 60 (5):663-673. doi:10.1057/palgrave.jors.2602596 168. Zonderland ME, Boucherie RJ, Litvak N, Vleggeert-Lankamp C (2010) Planning and scheduling of semi-urgent surgeries. Health Care Manag Sc 13 (3):256-267. doi:10.1007/s10729-010-9127-6 86 6. Literature Review: Patient scheduling ABSTRACT Hospitals are under growing pressure to cut costs, meanwhile sustaining the level of service provided to the patients. One part of the solution to this difficult task is to improve the efficiency of the operating rooms by using a sound patient scheduling framework. This can be done by identifying and applying good operations research methods from the healthcare literature. This literature review tries to help the reader to achieve this goal, firstly by pointing out relevant articles and, secondly by identifying and describing some of the major research groups in the field. Each research group is approaching the patient planning problem in a different way as, for example, some groups look at timing decisions such as finding an appropriate surgery date (e.g., Dec. 16) or time (e.g., 11:30) for a patient, whereas some include decisions regarding the surgery location, that is, determining an appropriate operating room. We found that seven major research groups are focusing on patient scheduling related problems and that their findings can be highly relevant to achieve a better and more efficiently used operating room. Reading about the research groups hopefully helps the reader, on the one hand, to get an overview over the patient planning literature and, on the other hand, to identify useful methods. 6.1. Introduction Hospitals are under increased pressure to cut costs meanwhile requested to keep or even improve their level of service. Given this difficult task, many managers are trying to realize cost savings at one particular part of the hospital, namely the operating rooms (OR). Tools from the operations research literature can help hospital managers to achieve this difficult task and to increase the efficiency of the OR’s by improving the way patient planning and scheduling is done. This review focuses on those topics from the surgery planning and scheduling literature that are primarily about the patient. Problems of this kind involve timing decisions such as finding an appropriate surgery date (e.g., Dec. 16 2012) and time (e.g., 11:30 AM) for 87 a patient but also decisions regarding the surgery location such as determining an appropriate OR. A more general literature review on OR scheduling can be found in Demeulemeester et al. [10] and Cardoen et al. [8]. Excluded from this review are scheduling problems related to the discipline level (e.g., pediatrics), the surgeon itself or to any other medical entity. Not many research groups devote a substantial amount of their scientific effort on researching problems specific to patient scheduling. Moreover, each research group uses, from the methodological perspective, a restricted set of tools and reapplies or extends similar methods across the scientific contributions they make. It is therefore straightforward to introduce the research groups one by one, together with their applied methodology. Discussing a research group at an earlier (or at a later) stage does not mean that their method is less (or more) sophisticated than the one of another group, but merely reflects a sequencing choice, which supports a logical and continuous way of describing the research field. In hospitals the demands of both patients and medical personnel are important and have to be considered alongside cost considerations. This means, on the one hand, that patient demands have to be satisfied both at the individual level (e.g., by decreasing patient waiting time) and on the patient group level (e.g., considering patient priorities). Additionally, surgeon and hospital requirements need to be met (e.g., allowing surgeons to schedule patients into preferred slots). Cost considerations, on the other hand, translate into strategies, where the amount of salary paid to medical personnel is decreased by ensuring that an OR closes in time and therefore no overtime costs occur. The strategy can also mean that we try to increase profitability by serving, for example, more patients. To cope with the many aspects related to patient scheduling, operations researchers are continuously developing new and innovative methods or searching and defining best practices. Recent problems found to be important by the operations research community are OR integration (section 2.1-2.2) and uncertainty (section 2.6). OR integration ensures that the created surgery schedule is to allow for a smooth patient care in pre/post-surgery facilities such as the post anesthesia care unit (PACU). Uncertainty is incorporated in order to achieve robust scheduling. It is considered on different levels. First and most prominently, it is accounted for with respect to patient surgery times, which are unknown until surgeries are actually realized (section 2.6). Second, it is unknown whether the OR will be available at the planned surgery start time as the arrival of an emergency patient could occupy it (section 2.1). Third, while booking a patient for surgery into a certain slot, it is unknown whether that slot could be needed to allocate a future more urgent patient (section 2.5). The second and third problems are similar as in both problems future arriving patients have to be anticipated and prepared for. 88 6.2. Literature review The following section contains a summary on the research activity observed in the field of patient scheduling. For each group only those key articles are given, which are directly linked to patient scheduling. Some groups are more focused on patient scheduling problems than others. The research output of the former groups will be described in more detail. Articles of a group that are highly related are grouped together and are referred to by the same capital letter (e.g., A1, A2 and A3). Single contributions not associated to a particular research group are discussed at the end of the text. Throughout the text, boldface and uppercase letters indicate vectors and random variables respectively. 6.2.1. The managerial way The first group publishes material that is highly managerial in nature and deals with applicability and implementability issues. The group is centered around Jan Vissers who works at the Erasmus University Medical Center in the Netherlands. The people associated to the group are: Vissers (Jan), Bekkers (Jos) – Erasmus University Medical Centre (Rotterdam) | Adan (Ivo), Dellaert (Nico) – Eindhoven University of Technology | Jeunet (Jully) - CNRS (Université Paris Dauphine) Article: 2011 | Improving operational effectiveness of tactical master plans for emergency and elective patients under stochastic demand and capacitated resources [1] The article describes and provides routines which can be followed step by step by practitioners. First, an approach is given to determine the amount of optimal OR capacity dedicated to patient types (tactical). Second, rules are given to guide the process of populating the dedicated capacities with patient instances. Third and finally, rules are given to guide the decision making process on the day of surgery. What distinguishes their article from most others in the field is the fact that it presents a holistic approach including many levels of the patient scheduling process. Moreover, it is also rare that authors give instructions usable on the surgery day itself. The authors of the article use a highly descriptive way of introducing their methods and rules. In order to give a rudimentary understanding of the basics of their work, we will use an explanation method that is based on state vectors. The values of the state vectors are calculable by the routines given in the article. The vectors are: Capacity vector c, where ci,t is the number of category ‘i’ patients to be operated on day ‘t’. The vector c is defined in advance (weeks or months) and is based on the average number of arriving patients of that category (similar to a master surgery schedule). It does not contain actual arriving patient instances. 89 Booking state s (operational plan), where si,t is the number of ‘i’ patients assigned to day ‘t’. The vector s is set ‘7’ or more days prior to the surgery date and contains real patient instances. Patients in s are getting notified about their surgery date. Planning vector p (executed plan), where pi,t is the number of ‘i’ patients to be certainly operated on day ‘t’. The vector is defined in the morning of the surgery day and is therefore regarded to be fixed. Goal programming is used to determine the value of vector c. In order to prepare for arriving non-electives (urgent or emergency patients), slack capacity is included. Vector c provides a template guiding the process of allocating actual patients to vector s. The allocation can happen in a strict or flexible way, depending on whether unused space is allowed to be occupied by patients from a different patient category. There might be unused space available if, of a given patient category, a less than expected number arrives. In order to calculate p, in the morning of the surgery day, a rule based daily scheduling algorithm is applied on s. This allows p to incorporate information with respect to still ongoing surgeries leftover from the night shift. At this point, as daily emergency arrivals are still not being realized, capacity for emergencies is still being reserved in the form of slack capacity. While determining p in the morning of the surgery day, if an elective surgery is estimated not to fit, it will be late canceled (severe action) and scheduled for day t+7. This may, in return, interfere with a patient assigned to s i,t+7 causing an early cancelation (less severe action). Arriving emergencies are handled in the same way as electives, as also they have to fit the schedule, i.e., they are not allowed to take the place of an elective. The difference between an emergency and an elective patient is that the emergency patient can be allocated into slack time. If an emergency does not fit the schedule, it is deferred to another hospital. Figure 42 gives an illustration of the whole procedure. Figure 42: Goal programming is used to arrive to the tactical plan ‘c’. Scheduling rules are applied to derive an operational plan ‘s’. Then, each morning, the daily scheduling algorithm is applied in order to derive the executed plan ‘p’. The same algorithm is used, at the moment an emergency arrives, to decide whether the emergency should be accepted and scheduled or deferred to another hospital. [2] Whether a surgery is expected to fit a schedule depends not only on the OR occupancy but also on the bed loads in the medium care unit (MCU) and intensive care unit (ICU). Additionally, it is required that nurses are available to support the ICU. An evaluation of the scheduling system is given through a Discrete-event simulation (DES) study where hospital efficiency measures are compared against patient satisfaction factors. Hospital efficiency measures the consistency in which a plan was carried out whereas patient satisfaction relates to waiting time, elective late cancelation and non-elective deferral. In their model hospital resources are exclusively devoted to a targeted discipline, namely cardiothoracic pathology. Hence, it would be interesting to see the results of their method applied to the general OR setting. 90 6.2.2. The exact way Developing efficient patient scheduling algorithms is a challenging task as patient scheduling problems are generally combinatorially complex and therefore difficult to solve. The combinatorial complexity of a problem depends on factors such as the problem objective and the restricting constraints. Research groups would generally, if possible, provide exact solution approaches. Exact solutions are especially from an academic perspective more interesting and favorable than approximate ones. The research group centered around Erik Demeulemeester is using exact methods in the major part of their research. The people associated to the group are: Demeulemeester (Erik), Ma (Guoxuan), Samudra (Michael) – KU Leuven | Belien (Jeroen) – Hogeschool-Universiteit Brussel | Cardoen (Brecht) – Vlerick Leuven Gent Management School Articles: A: 2009 | Optimizing a multiple objective surgical case sequencing problem [6] B: 2008 | Sequencing surgical cases in a day-care environment: An exact branch-and-price approach [7] In articles A and B, patient sequences and surgery starting times are determined. Their scheduling algorithm is applied on the fixed patient sets of each surgery block. A separate OR assignment step is therefore obsolete as a block is by definition allocated to an OR. Even though in A and B identical problems are solved, each of the articles describes different solution methods. In A, solutions based on Mixed Integer Linear Programming (MILP) are proposed, whereas in B, column generation as part of an enumerative branch-and-price framework is discussed. The problem objectives of both A and B are: Minimize the sum of surgery starting times carried out on children; Minimize the sum of surgery starting times of prioritized patients; Minimize the number of patients coming from further than 150km and scheduled for early surgery; Minimize the number of overtime periods in the recovery areas; Minimize the peak bed occupancy in PACU 1 and PACU 2 (patients first enter PACU 1 and are later transferred to PACU 2). The cost coefficients associated to each objective are a product of two components. The first component is a normalization factor mapping the coefficient to a value between 0 and 1. The second component is chosen by the hospital management and reflects the importance of the objective relative to other objectives. Constraints in their model are: 91 Surgeons need to allocate their surgeries into the time window provided to them (by the master surgery schedule); Each surgery has to be given a surgery time; Patients required to undergo pre-surgical tests need to be scheduled after a reference period; Peak occupancy in PACU 1 and PACU 2 cannot exceed the respective capacities of the units; The number of medical equipment used in parallel cannot exceed their availability; The OR requires a long cleaning session after the treatment of a patient infected by the bacterium called Methicillinresistant Staphylococcus aureus (MRSA). In A, three MILP based solution procedures are introduced. The first and the second are a basic and a preprocessed MILP. In the preprocessed version, the implications of fixing each binary variable to 0 or 1 (probing) are investigated. Additionally, the possible starting times of both MRSA and non-infected patients are limited. In the 3rd, the iterated MILP, the starting times of a particular set of surgeries is iteratively fixed. In B, a column (representing a surgery sequence) oriented reformulation of the MILP is given. This allows to enforce most constraints within the column, i.e., independently of other columns. The constraints which cannot be enforced within a column are the ones related to shared resources (beds, medical instruments). Figure 43 gives an outline of the column generation approach. In the model, the restricted master problem is solved with an initial set of existing columns. Through a process called pricing, new columns with reduced costs are iteratively identified and added to the set of already selected columns. The process terminates when no more columns price out negatively (in which case the relaxed problem is solved to optimality). Figure 43: In the column generation approach, iteratively newly generated columns (surgery sequences) are added to the problem. In the restricted master problem, the decision variable (selecting a column) is relaxed. In order to arrive to an integer solution, the column generation approach is embedded into an enumerative branch-and-bound framework. The authors propose several branching strategies and argue that it is more effective to branch on the individual surgeries than on the decision variables themselves. If the MILP or column generation based procedures are given a time limit (e.g., 5 minutes), we are likely to arrive at a sub-optimal solution. In many cases though, the algorithm terminates within the time limit and hence provides the optimal solution (iterated MILP is an exception as a probability based selection procedure is used). 6.2.3. Heuristics 92 Approximate solutions are used if it is difficult to find an efficient and exact algorithm to solve a particular problem. The group centered around Nadine Meskens seems to prefer solution methods of exact nature but turns to heuristics if they seem necessary. An example is their column generation approach. While the column generation of Demeulemeester’s group is ensured to be optimal by a branch-and-bound framework, Meskens’ group uses a column generation based heuristic. The people associated to the group are: Meskens (Nadine), Fei (Hongying) – Louvain School of Management, Catholic University of Mons | Chu (Chengbin) – Ecole Central Paris | Combes (Catherine) – University of Lyon Articles: A: 2009 | Solving a tactical operating room planning problem by a column-generation-based heuristic procedure with four criteria [15] B: 2009 | The endoscopy scheduling problem: A case study with two specialized operating rooms [17] C: 2010 | A planning and scheduling problem for an operating theatre using an open scheduling strategy [16] D: 2010 | Using constraint programming to schedule an operating theatre [20] Articles A, B and C are related and are hence discussed together. In A, the date and OR assignment problem is solved in an open scheduling problem setting. In open scheduling, a surgery can be assigned to any OR-day (e.g., OR 5 next Monday) within the planning horizon. Their objective is to minimize costs of unexploited opening hours and overtime. A column based formulation is used, where a column represents the surgeries that are assigned to an OR-day. This way, a set of columns will represent a solution. The decision variables select from the pool of possible columns the subset that represents the optimal solution. A problem related to column generation based approaches is that the decision variables may not be integral, i.e., the columns are not selected decisively. To overcome this problem a heuristics procedure is applied. B and C extend the problem setting of A by including a sequencing step. The assignment and sequencing step are independent of each other, i.e., the sequencing step provides no feedback to the OR-day assignment step. In B, a special case of the scheduling problem is introduced where only 2 ORs are considered. The problem is modeled as a two-machine open-shop scheduling problem and is solved by the Gonzalez-Sahni [5] algorithm. A more general problem setting is used in C, where the sequencing problem is solved by a genetic algorithm. The genetic algorithm ensures compliance with two major constraints. Firstly, it is ensured that a surgeon will perform one surgery at a time. Secondly, recovery bed limitations are obeyed as patients are simultaneously scheduled for the OR and the recovery room. The objective of their problem is to minimize the makespan in ORs and recovery rooms. In D, the problem of determining good surgery sequences and starting times is solved by a constraint programming approach. Constraints ensure that two surgeries are not carried out simultaneously in the same OR. Their model is based on a threedimensional Boolean matrix otr(o,t,r), where ‘o’ denotes the surgery, ‘t’ the time slot and ‘r’ the room. The Boolean corresponding to a surgery ‘o’ scheduled at time ‘t’ in operating room ‘r’ takes the value 1, whereas all other variables are set to 0. Constraints ensure that a surgery is performed continuously without stops in between. Constraints are also set for earliest availability and duedates. 93 Meskens’ group deals separately with the date and the time assignment step. In their work, an inpatient setting is presumed which means highly variable surgery durations and frequent emergency disturbances. An interesting extension of their work could therefore use stochastic surgery durations and include aspects of emergency patient arrivals. 6.2.4. Uncertainty Noticeable is the way how research groups relate to approximate methods. Demeulemeester’s group generally avoids them, Meskens’ applies them intermittently, and the group centered around Xiaolan Xie heavily, almost exclusively uses heuristics. Whether a group uses exact or approximate methods might be, on the one hand, dependent on the group’s attitude. On the other hand, the group might have no other option, as after a problem reaches a certain level of complexity, it can become impossible to solve it exactly. The level of complexity when researches tend to turn to heuristics seems to be reached if uncertainty is introduced. The people associated to the group are: Xie (Xiaolan), Lamiri (Mehdi), Augusto (Vincent), Grimaud (Frédéric) – Ecole Nationale Supérieure des Mines de Saint Etienne (Engineering and Health Division) Articles: A1: 2008 | A stochastic model for operating room planning with elective and emergency demand for surgery [24] A2: 2009 | Optimization methods for a stochastic surgery planning problem [23] A3: 2008 | Column generation approach to operating theater planning with elective and emergency patients [25] A4: 2007 | Operating room planning with random surgery times [22] B: 2010 | Operating theatre scheduling with patient recovery in both operating rooms and recovery beds [3] Xie’s group focuses on the date assignment step in all articles except B. In B, a deterministic time assignment problem is solved. Additionally, the patient to OR assignment step complements the general patient to date assignment in A3 and A4. In A1 and A2, it is unnecessary to determine the patient to OR assignment as only the cumulated capacity of ORs is considered. Since non-electives might occupy some of the OR’s capacity, the accessible capacity left to electives is uncertain. Implicit in the problem is therefore the determination of allocated elective OR capacity. Formally, an assignment between elective patient ‘i’ and day ‘t’ has to be found, given that the total OR capacity is decreased by random capacity Wt (arriving non-electives). The objective includes the minimization of patient related costs (cit ) and overtime. Patient related costs will depend on the patient type and the assigned date. Patient costs are independent of non-elective arrivals and are not influenced by the realizations of random capacity Wt , i.e., the stochastic problem only affects OR overtime costs. The objective is: + J∗ = Minimize J(x) = ∑ ∑ cit xit + ∑ ct Ewt [(Wt + ∑ di xit − Tt ) ] i 94 t t i ∈ It (1) where ct is the unit cost of overtime and Tt is the total capacity available on day ‘t’. In A1 and A2, a Monte Carlo type of solution method is presented. In A1, the solution algorithm involves a MIP, while in A2 a heuristic is used. Since a Monte Carlo solution method is inherently approximate, the solution to the MIP can only be an approximation. In the Monte Carlo solution, Ewt is approximated by sample averages. The method is consequently called sample average approximation (SAA). Other names of the method are: the sample path or stochastic counterpart method. The resulting objective function is formulated as: + K 1 JK∗ = Minimize JK (x) = ∑ ∑ cit xit + ∑ ct ∑ [(wtk + ∑ di xit − Tt ) ](2) K k=1 i t t i ∈ It where wtk is the k-th sample from Wt and ‘K’ is the total sample size. The SAA method optimizes over a finite number of scenarios where a scenario represents the outcome of a random process. In A1 and A2, the random process relates to Wt only. Other examples of a random process in general include: elective surgery times, length of stay (LOS) in the PACU or other units. Stochastic considerations can also be applied to hard constraints. In that case, it is ensured that the constraint holds for every scenario ‘k’. In A2, the authors prove that their problem’s SAA convergence rate to the true optimum is exponential. The quality of the SAA method with respect to the deterministic counterpart is investigated in A1. The authors show that for a moderate sample size K (~20) a better result is obtained by SAA than by the deterministic method. In the deterministic model, Wt = E[Wt(fixed non-elective OR time)] and the corresponding solution is calculated by a MIP. In A2, several constructive and improvement heuristics as well as meta-heuristics are introduced. Depending on the heuristic, different moves are defined. Moves are, for example, to add a surgery ‘i’ to period ‘t’ or to exchange two surgeries. The quality of a move is determined by assessing the gained benefit according to objective (1). The authors’ results suggest that Tabu search is the most suitable heuristic and performs best for large instance sizes. It outperforms the Mont-Carlo type of MIP solution with fixed time budget. The best constructive heuristic is to sort surgeries in increasing order of their surgery durations. The surgeries are added to the schedule one by one, maximizing at each step objective (1). A counterintuitive observation of the authors is that as the variability of Wt decreases, so does the solution quality of the heuristic. In A3, the problem setting is widened by the patient to OR assignment step. The applied solution method is column generation. The decision variables are yip (patient ‘i’ is assigned to plan ‘p’) and ztsp (plan ‘p’ is assigned to OR ‘s’ on day ‘t’). A set of feasible plans will give the solution. The variable λp indicates whether plan ‘p’ is selected or not. Similarly to the work of Meskens, heuristics are used to ensure integrality of λp’s and to improve existing feasible solutions. Overtime costs and the penalty for exceeding the OR-day availability are set by the authors to 500€/hour and 3000€/hour respectively. The cost of underutilization is 1.75 times less than the cost of overtime. The authors set the costs as fixed and choose values which are typically used in French hospitals. We think that an interesting future extension of their method could investigate the case when emergencies are regarded to arrive to the entire hospital instead of single ORs, i.e., instead of excluding a random amount of emergency capacity from each OR, a total amount of emergency capacity could be excluded from all the ORs in a shared manner. This is similar to a queuing system, 95 where modeling one common queue for all the servers or a separate queue for each server will cause the system to behave differently. In A4, the problem setting is in two major points different from A3. First, not only non-elective capacity requirements are stochastic but also elective surgery durations. Second, the objective function is a simplified version of the one in A3 as it does not include a penalty component for capacity violations nor excessive idle time. The time assignment problem is solved in B. In their model, an integrated approach is considered as the availability of recovery beds and porter teams is included. Additionally, the model allows for patient recovery in the OR. In the article, the authors chose to minimize makespan and conclude that recovery in the OR is beneficial as soon as the ratio of recovery beds to ORs is lower than 3/2. 6.2.5. Dynamic allocation problem The group centered around Yuehwern Yih also incorporates aspects of uncertainty with a Monte Carlo type of solution approach. The dynamic scheduling problem, arising when future patient arrivals are taken into account, is solved by value iteration. The people associated to the group are: Yih (Yuehwern), Min (Daiki) – Purdue University (School of Industrial Engineering, West Lafayette) Articles: A: 2010 | An elective surgery scheduling problem considering patient priority [26] B: 2010 | Scheduling elective surgery under uncertainty and downstream capacity constraints [27] In A, the dynamic date assignment problem with ‘i’ different types of priority patients is solved. The described model finds the best action vector a = (a1, a2, …, aI) given a state vector s = (s1, s2, …, sI), where ai is the number of patients of priority class ‘i’ selected to undergo surgery in the next period and si is the number of priority class ‘i’ patients waiting for surgery. Vector a has to be chosen carefully as if too many patients are booked and many high priority patients will arrive in the future, high overtime costs become inevitable. Future demand is denoted by vector d = (d1, d2, …, dI). The problem objective is to minimize the amount of overtime and patient postponement costs. The corresponding Bellman optimality equation is: v(s) = min c(s, a) + λ ∑ p(d)v(s − a + d) a (3) d where v(s) is the value function and λϵ[0,1] the discount factor. The discount factor denotes the present value of future costs c(s, a) and is usually chosen to be close to 1. The cost vector c(s, a) is the sum of postponement and overtime costs. I ∞ c(s, a) = ∑ ci (si − ai ) + co ∫ (x − cap) dFa (x) i 96 cap (4) where ci is the cost coefficient of priority class ‘i’, co is the unit overtime cost, cap is the capacity of ORs (without overtime) and Fa (x) is the cumulative distribution function of total surgery durations. Figure 44 shows how the different cost factors sum up to yield value function v(s). Figure 44: As more and more patients are scheduled, overtime is increasingly dominating the cost function. [26] As shown by Figure 45, the scheduling rules are monotonic. For example, the more patients in the waiting list, the more patients will be scheduled. The same is true for priorities, i.e., a waiting list containing a larger number of high priority patients will schedule at least as many patients as if the list would contain many low priority patients. Figure 45: The solid line denotes the case without priorities. Different points represent scenarios consisting of different mixes of priority classes. [26] Their model shows how practices from the reinforcement learning literature are applicable to healthcare scheduling problems. Their procedure, the value iteration algorithm, is one of the elementary reinforcement learning methods. We would be interested to see whether some other methods from the field would be applicable in a similar fashion. For example, reinforcement learning algorithms are used in nondeterministic environments where state s and action a result in an uncertain state s’. This could be used to model patient cancelation. 97 In B, a technique resembling Xie’s A1 is introduced. They resemble as they apply the same methodology (SAA) and define coinciding objectives (minimize patient related and overtime costs). In B, stochastic aspects related to surgery durations, LOS in the ICU and non-elective patient arrivals are considered. The complexity of their problem setting is increased by restricting surgery assignments to a subset of surgery blocks. This is an important consideration as many hospitals use block scheduling in reality. To deal with this increased complexity of the problem, a stochastic programming model with recourse is defined where stochasticity is only included in the recourse function. As stochasticity only affects overtime, the objective of the recourse function likewise includes the minimization of overtime only. A general problem is that the solution quality of approximation algorithms is hard to assess. In B, a method based on solution averages is given to find the lower bound of a solution. The upper bound is found by evaluating the “true” objective value of a suboptimal solution. Yih’s group deals with two highly relevant parts of the patient to day assignment problem. In A, patient prioritization is considered and capacity is reserved for future bookings. In B, uncertainty related to surgery duration, LOS and non-elective patient arrivals is considered. Both A and B deal with different parts of the same problem, namely the robust assignment of electives to days or blocks. It would be very interesting to see whether and how A and B could be effectively unified in one model. 6.2.6. Stochastic programming The group centered around Brian Denton, similarly to Xie’s and Yih’s group, generally includes stochastic elements in its models. The group often combines Monte Carlo type of solution methods and mathematical programming. Their preferred problem setting includes the patient to day assignment step, sequencing and determining surgery starting times. The people associated to the group are: Denton (Brian), Viapiano (James), Erdogan (S. Ayca), Berg (Bjorn) – North Carolina State University | Gupta (Diwakar) – University of Minnesota (Dep. Of Mechanical Engineering) | Huschka (Todd R.), Rohleder (Thomas) – Mayo Clinic | Batun (Sakine), Schaefer (Andrew) – University of Pittsburgh (Dep. of Industrial Engineering) Articles: A1: 2007 | Surgical suits’ operations management [18] B1: 2003 | A sequential bounding approach for optimal appointment scheduling [11] B2: 2006 | Simulation of a multiple operating room surgical suite [13] B3: 2011 | Dynamic appointment scheduling with uncertain demand [14] C1: 2007 | Optimization of surgery sequencing and scheduling decisions under uncertainty [12] C2: 2010 | Operating room pooling and parallel surgery processing under uncertainty [4] Throughout their articles, the group consistently uses similarly defined performance measures and solves similarly formulated decision problems. Therefore first both their preferred objective function and their regularly used decision variables will be introduced and the exact details of each single article will only be shown afterwards. 98 Their universally used decision variable ‘xi’ denotes the time allowance for patient ‘i’, i.e., the amount of OR time reserved for the patient. The surgery starting time of a patient is the sum of the time allowances assigned to all preceding patients. The variable ‘Zi ′ represents the actual surgery time of patient ‘i’. The two performance measures generally included in their models are waiting ‘W’ and facility overtime ‘L’. Waiting time targets either the surgeon or the patient. The third performance measure, idle time ‘S’, is used in some of their articles and is usually defined with respect to the facility. Given patient ‘i’, the three measures are defined as: Wi = max(Wi−1 + Zi−1 − xi−1 , 0) Si = max(−Wi−1 − Zi−1 + xi−1 , 0) n−1 L = (−Wn − Zn − ∑ xi + d) i=1 where ‘n’ is the total number of scheduled patients and ‘d’ is the facility’s total capacity. Given cost coefficients c w, cs and c l , the objective function can then be formulated as: n min {∑(ciw E[Wi ] + cis E[Si ]) + c l E[L]} x (5) i=1 It is presumed that the surgeries are already sequenced in all articles of type B. In articles of type C also sequencing decisions are included. The group models surgery start times, as in railway scheduling [33], in a fashion that surgeries never start earlier than planned (in case a previous surgery requires less time). This is the case as the performance measures are recursive in the variable ′Wi ′, a variable which is always bigger than or equal to 0. This models a setting where, for example, each surgery is performed by a different surgeon who is only available at the exact starting time of the surgery. Denton’s group formulates their problems as two-stage stochastic programs. In stochastic programming, “the decision maker takes some action in the first stage, after which a random event occurs affecting the outcome of the first-stage decision. A recourse decision can then be made in the second stage that compensates for any bad effects that might have been experienced as a result of the first-stage decision. The optimal policy from such a model is a single first-stage policy and a collection of recourse decisions (a decision rule) defining which second-stage action should be taken in response to each random outcome” (Wikipedia - stochastic programming). The first stage decisions are ′xi ’, whereas the second stage decisions correspond to the calculated amount of waiting, idling and overtime. Until now, the common properties of the group articles were described. Next, a short description on article A1 is given. The article describes three models, of which two will be mentioned together with their given preliminary solution approaches. The first of the two models describes the elective surgery booking problem of prioritized patients. It is solved, as in Yih, with the value iteration algorithm. While in Yih the question over the amount of OR capacity to be reserved for the next time period is being answered, in A1, the (preliminary) answer to a different question is given. The question at hand is: how much downstream capacity should we reserve on each future day ‘t’ for priority class ‘i’, given stochastic future demand? The second of the two problems deals with the surgery sequencing problem. The objective of the introduced model is defined by (5). The sequencing problem is computationally expensive since the set of possible sequences grows factorially in the number of 99 surgeries. Additionally, the objective value does not depend on the sequence only, but also on the surgery allowances. In order to gain insights into the problem structure, the author radically simplifies the modeling assumptions. In the article, solution guidelines based on stochastic ordering are being formulated, suggesting that surgeries should be sequenced in the order of increasing variance of their surgery duration. Articles B1, B2 and B3 exclude the sequencing step and contain methods to determine OR time allowances. In B3, the sequencing step is obsolete as the problem setting excludes patient priorities. The patients are therefore scheduled on a FCFS (first-come, first-served) basis. In the articles of type B, the objective function (5) is used, with the exception of B3, where idle time is excluded. In B1, time allowances are determined based on objective (5). In the article the stochastic linear programming formulation is introduced which will reappear in many of their later works. As it is common for the group, the authors try to exploit the models’ structural properties in an effort to increase its solvability. The optimal solution is derived using a variation of the L-shaped algorithm [9]. A finding discussed in the article is that the optimal solution will exhibit a dome shape. In other words, assuming i.i.d. surgery times and equal waiting and idling costs, allowances will initially increase and towards the end of the surgery day decrease again. Also in B2, time allowances are determined with objective (5). In the article, a Simulated Annealing (SA) approach using Monte Carlo sampling is introduced. The SA procedure will in each step evaluate a new schedule on the basis of the Monte Carlo procedure given sample size ‘K’ (similarly to (2)). A new schedule is generated by perturbating the current schedule. In B3, as in previous B articles, time allowances are determined. The objective is defined by (5), but without considering idle time. The article, as A1 and Yih’s A1, deals with the dynamic appointment scheduling problem. The major difference in respect to Yih’s A1 considers the task description and the methodology. The task description is different from Yih’s A1: a set of differently prioritized patients is considered whereas in B3 patients are equally prioritized, i.e., patients are considered in sequence of their arrival. Additionally, in Yih’s A1, patients are assigned to a virtual one period overall capacity (sum across ORs), while in B3 surgery starting times are determined for one OR-day. The two models are also from a methodological perspective different. Yih applies the value iteration algorithm, whereas in B3 a multi-stage stochastic linear program is used. In Figure 46, the modeling idea followed in B3 is exemplified by a simple case. In the example, the first patient is scheduled and two additional patients may arrive. Patients are scheduled with imperfect knowledge about both the number of future arriving patients and their respective surgery durations. Each new request is therefore treated as an additional stage in the stochastic program. Figure 46: Multi-stage stochastic linear program. Outcomes and probabilities in case 1 patient is scheduled and 2 additional ones may arrive. [14] 100 The C type of articles are different from the B type of articles in one major point. Whereas in B allowances were determined and a surgery sequence was assumed to be given, in C also the sequencing step is carried out. In C1, patient sequences and allowances are determined using objective (5). The model is represented as a two-stage stochastic program. As the model is highly complex and an exact solution approach would be computationally expensive, an interchange heuristic is proposed. The heuristic, at each iteration step, performs a randomly generated pairwise interchange and checks the resulting solution quality. The quality of a solution is evaluated by a traditional two-stage stochastic model with patient sequences fixed. In addition to the interchange heuristics some easy to implement constructive heuristics are presented. The most effective constructive heuristic sequences surgeries in increasing order of the variance of their surgery duration. This is a strategy that is also praised in many of their other articles. In C2, ideas from C1 are further developed. The problems tackled in C2 involve: The number of ORs to open; Surgery to OR assignment; The sequence of surgeries; The start time for each surgeon in the morning. The objective function is similar to (5), but instead of including waiting time a factor representing OR opening costs is added. In the article, the authors exploit the fact that a surgery is dividable into 3 parts: preincision, incision and postincision. Surgeons are only required to be present at the incision part of the surgery and are able to freely switch between ORs. The authors refer to this situation as parallel surgery processing, since surgeries assigned to the same surgeon might be carried out in parallel (as long as the incision parts are not overlapping). In order to make effective use of parallel surgery processing, it is necessary to handle ORs as a pooled resource. This means that an open scheduling environment is assumed, i.e., not block scheduling. Important is the fact that the length of each of the three parts of the surgeries is stochastic. The problem is once again tackled by a two-stage stochastic model, where the first stage variables are the basic decision variables. Second stage decisions are completion times, surgeon idle times and the overtime in each OR. As the problem is computationally difficult to solve, the authors restrict the search space and define antisymmetry constraints with respect to OR orders. Induced feasibility constraints stemming from the problem structure are used in a standard L-shape and an L-shaped-based branch-and-cut algorithm. The theoretical cost reductions that can be achieved by using their method range from ~21% to ~59%. Denton’s group is specialized in optimization methods where stochasticity plays an important role. Since stochastic optimization problems are highly complex, the group generally searches for ways to exploit the structural properties of their problem instances. Besides mathematical programming, they often turn to either constructive or improvement heuristics. 6.2.7. Robust schedule 101 The group centered around Erwin Hans approaches stochasticity in a less traditional way than previous groups and uses solution methods, which are based on common sense. The group deals with two main problems. Firstly, the maximization of OR usage given a fixed probability of overtime is solved. Secondly, they reduce the maximum waiting time of urgent arrivals. The people associated to the group are: Hans (Erwin), van der Lans (Marieke) – University of Twente | Wullink (Gerhard), van Houdenhoven (Mark), Kazemier (Geert) – Erasmus Medical Centre Articles: A: 2006 | Anticipating urgent surgery in operating room departments [34] B: 2008 | Robust surgery loading [19] In B, the authors consider the robust surgery loading problem. Given a predetermined schedule (surgery to OR-days), the authors discuss different reassignment strategies aiming at the maximization of OR utilization and the minimization of the risk of overtime. The risk of overtime in a given OR depends on the variation of the surgery durations in the OR and on the amount of slack time (buffer). In other words, the lower the total variance of the surgery durations, the less slack time is required. Figure 47 shows how the portfolio effect can reduce the variance of the summed surgery durations in a 2 OR setting. In the figure (left), the standard deviation of the summed surgery durations equals 2 ∗ √502 + 102 = 102 time units. In the second scenario (right), the amount is reduced to √502 + 502 + √102 + 102 = 84.9 time units. The shown reduction in variance allows to use less slack time and therefore a larger amount of OR time can be effectively used by surgeons. Figure 47: Decreasing the variance of summed surgery time using the portfolio effect. [34] In A, the focus is on reducing the waiting time of urgent patients. In order to achieve this, possible urgent surgery entry points, referred to as break-in-moments (BIMs), are distributed evenly during the day. In Figure 48, BIMs are shown for a 2 OR case. The break-in-interval (BII) represents the time intervals between BIMs. The optimization problem boils down to minimizing the maximum BII. 102 Figure 48: Break-in-moments and break-in-intervals in a 2 OR setting. [34] Different heuristics are introduced, which try to approximate the case where BIIs are of even length (ideal case). The authors introduce different strategies with respect to the way how slack time is divided among ORs. They conclude that distributing slack over all the ORs while performing BIM optimization performs best. We think that using BIM optimization in a setting with a high number of ORs might have a lessened impact since if there are many ORs also BIMs will occur more frequently. This is however not really a problem as we know from practice that in large hospitals emergency patients are often only allowed to enter ORs allocated to their own discipline. BIM optimization can then simply be restricted to ORs of the same discipline. 6.2.8. Summarizing table In Table 10 and Table 11 the groups are compared according to a few selected attributes. The tables are not meant to give detailed information over the groups, but to merely provide a way to easily classify and compare the different research groups. Therefore attributes are selected only if they are valid for at least half of the articles of the given research group (e.g., if in 2 out of 4 articles considerations over non-elective arrivals are incorporated then the keyword ‘Non-elective’ is going to be included into the column ‘Patient type’ of that group). 103 Group Patient type Patient assigned to Uncertainty Supporting facilities Date LOS ICU, MCU 2.1 The way managerial Inpatient, Non-elective 2.2 The exact way Outpatient Time, Room Pacu Elective Date, Time, Room Pacu Elective, Non-elective Date, Room Emergency capacity Elective, Non-elective Date, Room Surgery LOS in ICU Elective Time Surgery duration Elective, Non-elective Date, Time, Room Surgery duration 2.3 Heuristics 2.4 Uncertainty 2.5 Dynamic allocation problem ICU duration, 2.6 Stochastic programming 2.7 Robust schedule Table 10: Some groups frequently incorporate considerations over non-electives whereas other groups focus on electives only. The table also shows that the patient to date assignment problem is solved most often. 104 Group Constraints Performance measure Solution technique 2.1 The way managerial Patient waiting time, Deferral, Cancelation programming, DES Utilization, OR time, ICU/MCU Goalbeds, ICU nurse hours 2.2 The exact way Prioritized patient starting Pacutime, Leveling bed usage in Medical Pacu,equipment Pacu overtime Branch-and-Price, capacity, Column generation MIP, OR utilization, Makespan Medical personnel Constructive Column generation heuristic, Patient OR overtime waiting Slot time, time MIP, SAA Patient OR overtime waiting Slot time, time Dynamic MIP, SAA programming, Patient waiting OR overtime, OR idle time time, Linear SAA Programming, 2.3 Heuristics 2.4 Uncertainty 2.5 Dynamic allocation problem 2.6 Stochastic programming 2.7 Robust schedule OR overtime Patient waiting time, OR overtime, OR utilization Constructive annealing heuristic, Simulated Table 11: Most groups include patient waiting time as one of their performance criteria. The table also shows that mathematical programming is a popularly used method among the research groups. 6.2.9. Single Contributions The following section contains articles that are not published by any of the previously introduced research groups. In Jebali et al. [21] a MIP is introduced to solve the patient planning problem in two steps. In other words, patients are assigned to ORs first and 105 are then sequenced. Again a MIP is used by Pham and Klinkert [29]. In their model, the availability of resources is stressed and modes are defined. A mode is a subset of resources that are occupied by a given patient. The usability of their method is limited by the capabilities of general purpose MILP solvers. To overcome complexity problems Roland et al. [31] turn to genetic algorithms. They define a 4-dimensional decision variable denoting for each surgery the operating room, the date and the time of the procedure. The same decision problem is solved heuristically by Riise and Burke [30]. The authors focus on the properties of the heuristics and describe the search space defined by the neighborhood structure of a relocate and two-exchange operator. The objective value is the weighted sum of patient waiting time (days), overtime and children waiting time in the morning of their surgery. Assigning different cost coefficients results in different fitness surfaces. Using search space analysis, properties of the fitness space are described with respect to ruggedness and the distance properties of local optima. An approach based on the combination of simulation and optimization is presented by Persson and Persson [28]. In their approach, if requested by the patient, their surgery has to be carried out within a time window of 90 days. This affects the schedule and increases the mean waiting time of medium priority patients. Patient priorities are also considered in the revenue maximization problem of Stanciu and Vargas [32]. The focus of their work is on capacity decisions pertaining to patient classes. A patient class is a combination of the patient reimbursement level and type of surgery. A patient class enjoys higher priority if its expected revenue per unit surgery time is higher. The problem shares some similarities with the airlines resource scheduling problem. The difference is that passenger requests are discrete (one seat), while surgery patient requests are continuous and random (surgery time). 6.3. Conclusion In this review we have focused on the healthcare literature over the surgery planning and scheduling problem of patients. The publications of seven research groups were used to describe the field. We saw that the seven research groups are approaching the planning problem in different ways. For example, some groups give very general, almost holistic models, meant to guide managerial decision making, whereas others are focusing on smaller problems and their solutions. Moreover, some groups give priority to exact methods whereas for other groups different objectives are important such as incorporating the stochastic aspects of the problem. Those different goals require the use of different operations research methods. Having different research groups dealing with the same problem leads thus to a large diversity of solution methods and helps to get a better and more diverse understanding of the surgery planning and scheduling problem. 106 6.4. References 1. Adan I, Bekkers J, Dellaert N, Jeunet J, Vissers J (2011) Improving operational effectiveness of tactical master plans for emergency and elective patients under stochastic demand and capacitated resources. Eur J Oper Res 213 (1):290-308. doi:http://dx.doi.org/10.1016/j.ejor.2011.02.025 2. Adan I, Bekkers J, Dellaert N, Jeunet J, Vissers J (2011) Improving operational effectiveness of tactical master plans for emergency and elective patients under stochastic demand and capacitated resources. Eur J Oper Res In Press, Corrected Proof. doi:DOI: 10.1016/j.ejor.2011.02.025 3. Augusto V, Xie XL, Perdomo V (2010) Operating theatre scheduling with patient recovery in both operating rooms and recovery beds. Comput Ind Eng 58 (2):231-238. doi:10.1016/j.cie.2009.04.019 4. Batun S, Denton BT, Huschka TR, Schaefer AJ (2010) Operating room pooling and parallel surgery processing under uncertainty. Informs J Comput:ijoc.1100.0396. doi:10.1287/ijoc.1100.0396 5. Breit J, Schmidt G, Strusevich VA (2001) Two-machine open shop scheduling with an availability constraint. Oper Res Lett 29 (2):65-77 6. Cardoen B, Demeulemeester E, Belien J (2009) Optimizing a multiple objective surgical case sequencing problem. Int J Prod Econ 119 (2):354-366. doi:10.1016/j.ijpe.2009.03.009 7. Cardoen B, Demeulemeester E, Belien J (2009) Sequencing surgical cases in a day-care environment: An exact branch-and-price approach. Comput Oper Res 36 (9):2660-2669. doi:10.1016/j.cor.2008.11.012 8. Cardoen B, Demeulemeester E, Belien J (2010) Operating room planning and scheduling: A literature review. Eur J Oper Res 201 (3):921-932. doi:10.1016/j.ejor.2009.04.011 9. Dantzig G, Glynn P (1990) Parallel processors for planning under uncertainty. Ann Oper Res 22 (1):1-21. doi:10.1007/bf02023045 10. Demeulemeester E, Belien J, Cardoen B, Samudra M (2013) Operating room planning and scheduling. In: Denton B (ed) Handbook of healthcare operations management: Methods and applications. Springer, 11. Denton B, Gupta D (2003) A sequential bounding approach for optimal appointment scheduling. Iie Trans 35 (11):1003-1016. doi:10.1080/07408170390230169 12. Denton B, Viapiano J, Vogl A (2007) Optimization of surgery sequencing and scheduling decisions under uncertainty. Health Care Manag Sci 10 (1):13-24 13. Denton BT, Rahman AS, Nelson H, Bailey AC (2006) Simulation of a multiple operating room surgical suite. Proceedings of the 2006 Winter Simulation Conference, Vols 1-5:414-424 14. Erdogan SA, Denton B (2011) Dynamic appointment scheduling of a stochastic server with uncertain demand. 15. Fei H, Chu C, Meskens N (2009) Solving a tactical operating room planning problem by a column-generationbased heuristic procedure with four criteria. Ann Oper Res 166 (1):91-108. doi:DOI 10.1007/s10479-008-04133 16. Fei H, Meskens N, Chu C (2010) A planning and scheduling problem for an operating theatre using an open scheduling strategy. Comput Ind Eng 58 (2):221-230. doi:10.1016/j.cie.2009.02.012 17. Fei HY, Meskens N, Combes C, Chu CB (2009) The endoscopy scheduling problem: A case study with two specialised operating rooms. Int J Prod Econ 120 (2):452-462. doi:10.1016/j.ijpe.2007.09.016 18. Gupta D (2007) Surgical suites' operations management. Production and Operations Management 16 (6):689-700. doi:10.1111/j.1937-5956.2007.tb00289.x 19. Hans E, Wullink G, van Houdenhoven M, Kazemier G (2008) Robust surgery loading. Eur J Oper Res 185 (3):1038-1050. doi:10.1016/j.ejor.2006.08.022 20. Hanset A, Meskens N, Duvivier D (2010) Using constraint programming to schedule an operating theatre. 2010 IEEE Workshop on Health Care Management (WHCM):6 pp. doi:10.1109/whcm.2010.5441245 21. Jebali A, Alouane ABH, Ladet P (2006) Operating rooms scheduling. Int J Prod Econ 99 (1-2):52-62. doi:10.1016/j.ijpe.2004.12.006 22. Lamiri M, Dreo J, Xiaolan X (2007) Operating room planning with random surgery times. Proceedings of the 3rd Annual IEEE Conference on Automation Science and Engineering:521-526 107 23. Lamiri M, Grimaud F, Xie XL (2009) Optimization methods for a stochastic surgery planning problem. Int J Prod Econ 120 (2):400-410. doi:10.1016/j.ijpe.2008.11.021 24. Lamiri M, Xie XL, Dolgui A, Grimaud F (2008) A stochastic model for operating room planning with elective and emergency demand for surgery. Eur J Oper Res 185 (3):1026-1037. doi:10.1016/j.ejor.2006.02.057 25. Lamiri M, Xie XL, Zhang SG (2008) Column generation approach to operating theater planning with elective and emergency patients. Iie Trans 40 (9):838-852. doi:Doi 10.1080/07408170802165831 26. Min DK, Yih Y (2010) An elective surgery scheduling problem considering patient priority. Comput Oper Res 37 (6):1091-1099. doi:10.1016/j.cor.2009.09.016 27. Min DK, Yih Y (2010) Scheduling elective surgery under uncertainty and downstream capacity constraints. Eur J Oper Res 206 (3):642-652. doi:10.1016/j.ejor.2010.03.014 28. Persson M, Persson JA (2009) Health economic modeling to support surgery management at a swedish hospital. Omega-International Journal of Management Science 37 (4):853-863. doi:10.1016/j.omega.2008.05.007 29. Pham DN, Klinkert A (2008) Surgical case scheduling as a generalized job shop scheduling problem. Eur J Oper Res 185 (3):1011-1025. doi:10.1016/j.ejor.2006.03.059 30. Riise A, Burke E (2011) Local search for the surgery admission planning problem. Journal of Heuristics 17 (4):389-414. doi:10.1007/s10732-010-9139-x 31. Roland B, Di Martinelly C, Riane F, Pochet Y (2010) Scheduling an operating theatre under human resource constraints. Comput Ind Eng 58 (2):212-220. doi:10.1016/j.cie.2009.01.005 32. Stanciu A, Vargas L, May J (2010) A revenue management approach for managing operating room capacity. 2010 Winter Simulation Conference (WSC 2010):2444-2454. doi:10.1109/wsc.2010.5678940 33. Tian W, Demeulemeester E (2013) Railway scheduling reduces the expected project makespan over roadrunner scheduling in a multi-mode project scheduling environment. Ann Oper Res:1-21. doi:10.1007/s10479-012-1277-0 34. van der Lans M, Hans E, Hurink JL, Wullink G, van Houdenhoven M, Kazemier G (2006) Anticipating urgent surgery in operating room departments. BETA working paper WP-158. University of Twente 108 7. Perspectives The remainder of this project will be focused on improving our understanding of the two step scheduling procedure. This includes finding appropriate protection levels for each urgency class and patient discipline. In other words, for each urgency class we want to find the appropriate amount of reserved weekly capacity so that, in total, most of the patients are served within their DT. It is a general problem in hospitals that the exact time future high urgency patients will be arriving is unknown in advance. This makes it necessary to reserve some amount of buffer capacity for them. As seen from the revenue management literature in the airline industry, the optimal strategy could be one where for the coming weeks less capacity is being reserved (protected) than for weeks further away. This can be intuitively understood in the following way. The further away a week is in time, the more uncertainty there is about the amount of capacity we will require to use from that week. As a consequence, we will protect a larger amount of capacity for high urgency patients. For weeks closer in time, we have more certainty about the amount of capacity we will need for the high urgency class and therefore we let a larger amount of lower urgency class patients occupy that space, i.e., we protect less. This means that the protection levels have to be dynamic. Using static protection levels, many patients of lower urgency categories are served late. The solution approach we are currently pursuing is based on Markov Decision Processes (MDP). One of the largest problems we encountered is that the search space even for a simplified problem setting is very large and thus an approximate solution needs to be found. At this point in time, we are not yet sure whether an MDP solution is possible. Alternatively, a heuristic could be used that is called expected marginal seat revenue (EMSR). The problem with the heuristic approach is 109 that it requires us to attach a value to each patient urgency class. This is difficult since a more urgent patient is not more valuable per se, but simply needs to be served faster. Another part of the problem, for which we currently only have a simple heuristic, is the within week scheduling. This is the problem of assigning the patients scheduled for a particular week to an exact OR, to a weekday (Monday-Friday) and a time, i.e., for each patient an appropriate slot and for each OR an appropriate sequence of surgeries. The exact components to include into the scheduling algorithm are still to be selected. The components will be selected in a way that the resulting problem, firstly, is from an operations research perspective challenging to solve while, secondly, is of practical use. In order to ensure that the problem is practically usable, we will rely on the insights of our collaborators at Gasthuisberg. Depending on the selected components it could be possible that a mathematical programming based optimal solution can be found. If this is not the case, then a branch-and-bound or a branch-and-price based solution approach will be proposed. Besides finding appropriate protection levels and solving the within week scheduling problem, one of the important targets for us is to help to explore with Gasthuisberg the practical consequences of switching from the direct slot scheduling to the two-step scheduling procedure. 110 8. Conclusion The first stage of this project involved understanding the literature and thus the research done so far. We did this by analyzing 181 relevant articles and categorized them based on several descriptive fields. The exact categorization can be found at: http://www.econ.kuleuven.be/healthcare/review2011. We found that most attention is given to elective patients, and even though often not stated explicitly, it is in many cases implied to be an inpatient setting. Less frequently occurring in the literature are methods that consider outpatients. This is surprising as outpatient care is gaining in importance and therefore we would expect to observe an increasing amount of literature concerning this area. We also observed a gradual shift from determining the exact time of a surgery to problems related to date and/or room assignments. With respect to the performance measures considered, we found that overtime is the most popular measure and that preference-related criteria are gaining popularity. Noteworthy is the fact that preferences are increasingly used in multi-criteria settings. Mathematical programming, discrete event simulation and heuristics are the techniques most frequently used. It is also true that the majority of articles present results based on real data. However, it is important to note that this does not imply that the methods are applied in practice. The applicability aspect of methods is one of the major points that we emphasize in our own research. After learning about the current research, we were exploring some of the practical aspects of the patient scheduling problem. This required us to analyze many of the different patient related attributes of the 13 patient disciplines served in the 22 inpatient ORs at Gasthuisberg. Those attributes relate, for example, to the weekday dependent arrival patterns, to urgency categories as well as to both estimated and realized surgery durations. 111 Analyzing the distribution of estimated and realized surgery durations, we noticed that for many disciplines the length of the surgery durations is consistently underestimated. This can be explained by the fact that surgeons try to plan as many patients into their own slot time as they can while not being allowed to exceed the capacity of those slots. Underestimating the patient’s surgery durations gives them a tool to legally overfill their slots. We captured this relationship between estimated and realized surgery durations by modeling the marginal distribution of the two components and a copula connecting them. Naturally, overfilled ORs will result in a larger probability of the OR to run into actual overtime. To avoid excessive overtime, rescheduling actions are taken in practice, i.e., moving a patient from one OR to another or canceling a patient. In order to arrive to a realistic simulation model, we also included rescheduling actions. The rescheduling model is created using the managerial insights of the head nurse of the hospital. The implementation of the rescheduling method into the simulation model was followed by an investigation on whether the distribution of the number of performed simulated rescheduling actions for each hour of the day matches the one observed in reality. We found that the most realistic rescheduling model takes into consideration several factors such as the hour of day, the patient discipline and the OR kern. Similarly, also the allocation of emergency arrivals to ORs is modeled in detail. An example is the emergency discipline to OR discipline time dependent allocation schema, e.g., oncology emergencies are during the day allocated to oncology slots if available, otherwise to abdominal or transplantation slots, whereas in the nights or weekends to OR ‘B2’ or ‘B3’. Finally, the simulation model is used to investigate the impact of different scheduling policies. The main base for these policies is a concept called the DT. The DT of a patient is a time interval within which surgery should be performed or we risk the worsening of the patient’s health condition. The exact waiting time within the interval is of lesser importance as the objective is to simply serve as many patients as possible within their respective DT. We investigated the effects of pushing patients with a longer DT interval into the future in order to serve a larger 112 part of patients with a shorter DT in time. Our simulation results suggest that this strategy might be too rigid as it does not allow to effectively distribute high peaks of demand. We are currently working on a more selective strategy. In most hospitals, as in Gasthuisberg, patients are assigned to slots directly. This means that a patient who is planned for surgery will be directly planned to an OR and a day. This system of direct slot assignment is going to be changed in the near future to a two-step procedure. Instead of assigning patients directly to slots, they will be assigned to a week first. This means that a second step is required, where for all the patients assigned to a given week a suitable OR and weekday is selected. The advantage of the two-step procedure comes from the fact that the second part of the procedure, the within week scheduling part, can be done just before the start of that given week. We use the simulation model to test for the implications of switching to this two-step procedure, focusing primarily on the effect it has on the amount of patients scheduled after their respective DT, i.e., too late. Our current results suggest that in case of the two-step procedure it is very important to allow patients with higher urgencies to break into the already fixed weekly schedules. Additionally, it is important that the second step, the within week scheduling, is guided by the patients urgency category. Interestingly, we found that reserving a constant amount of capacity for high urgency patients is from a whole system perspective less beneficial. 113 9. Appendix Figure 49: One-step strategies. FCFS benefits all DT categories. 114 Figure 50 Figure 51: DT 6 and 7 patients receive surgery mostly in time. 115 Figure 52 Figure 53. The arrival rates for ABD for each weekday and DT category 116 Figure 54: Non-electives will also during the weekend or the night enter a predetermined set of ORs. Figure 55: Also the most urgent non-elective category (DT 1) is during slot hours assigned to ORs of their own discipline. Exceptions are oncology (ONC) and Oral and maxillofacial surgery (MKA) patients. 117 Figure 56 Figure 57 118