Inference engine the Systems of the Dr.Watson Type. S.I. Goldberg MediSpectra, Inc. 100 Inman Street ,Cambridge, MA 02139 Saveli@aol.com The systems of decision making support (DSS) are a popular and quickly developing ranch of artificial intelligence. In a number of regions the application of decision support systems give a great economic effect [ 1 ]. It is natural that the value of the effect depends in many respects on the accuracy relation of DSS conclusion to a man's ability to decide the problem, task complexity by himself. DSS are applied most successfully when its ability considerably surpasses the man's capacities. This happens either when it is possible to create adequate formal mathematical models, or when DSS are constructed on the basis of experts' knowledge considerably more qualified than potential users. Lately, DSS have been widely applied for the opposite flank as well, and namely for the users inventing new knowledge (invention machines) [ 2 ]. Besides, DSS in principle cannot make correct final conclusions, but create conditions for the activation of an inventor's work. It is natural that there is an intermediate situation as well when the conclusion accuracy of DSS and a user are comparable between themselves. Usually these are creative, but relatively standard and badly formalized tasks, which have to be decided every day by specially trained experts intended for this. To such kind of tasks are related many of the problems of high level specialists in medicine, business, criminology and so on. There are known numerous attempts of DSS' creation for such specialists. However , more than not the similar systems are used for teaching or , at the best , for a posterior checking by a specialist’s conclusion . For example, in medicine, as shown in [3] with a very great number of completed DSS , there are utterly few used in real practical work. There are offered various ways of DSS' creation really used for specialists: application of self-learning DSS [4 ]: integration of DSS with other informational technologies [5]; maximum use of an interactive regime in DSS [6]; intensification of organizing and administrative measures when implementing DSS [7]. Considering all the recommendations absolutely correct, nevertheless it is thought that the problem of unsuccessful use of DSS is in the very nature of the given tasks. Most likely it will be impossible in the foreseeable future to make conclusions of DSS surpassing specialists' conclusions in accuracy. Absence of formal theories, a great percentage of subjective information, the trend of results in time do not allow to construct precise mathematical models and the experts' conclusions (at least what presents them in a computer) differ a little from professionals' judgments. In conditions when the accuracy of DSS' conclusions is close or even inferior to the accuracy of a specialist's conclusions, the specialist's knowledge of DSS’ conclusions ,even formally, does not improve the accuracy of his conclusions. And from the point of view of psychologists such a prompt can even lead to the so-called effect of locking [8], when the specialist fully loses control of the situation. The reasonings are question able us well that the situation of constant competition with DSS - conclusions highness activity and responsibility of specialists. In [ 13 ] Yates explained that the competition situation of a person constantly solving complex tasks can lead to worsening the decision quality at the expense of developing mentality stereotypization. Moreover, a prompt at the consideration moment can lead to undesirable consequences even if their accuracy surpasses the accuracy of specialists. Mistakes in conclusions of specialists are often not the result of the absence of logical thinking or ignorance of the actual material but of a tendentious selection of the material for analysis. Gathering and filtration the material take place already under the influence of obviously or subconsciously selected hypothesis. And physicians, criminalists and businessperson know very well this phenomenon. If on the basic of such data, DSS make a decision favorable for the user, this will increase his misconception even more, though it may reduce his responsibility if only morally. Observing DSS' actions the specialist quickly enough learns with understanding in what situation DSS takes this or that decision and sometimes quite subconsciously he adjusts the data for DSS' response necessary for him. The author often observed this effect in the actual work of DSS created by him both on the basis of pattern recognition methods [9], and fully expert systems [10]. On the other hand, cases are known when less-skilled specialists being in the process of discussing some material help their more qualified colleagues to make correct decisions. The best known example of such sort is the joint activity of a famous detective Sherlock Holmes and his friend Dr.Watson. Moreover in his professional activity Holmes considerably more often turns to the comparatively poor detective Dr.Watson for aid than to his brother - a recognized authority Microft Holmes (an ideal expert system). Namely this story of interrelations of Holmes and Dr.Watson was an example for the creation of DSS' new class systems of the Dr.Watson Type. The main ideas of the systems were worked out by the author together with Prof. Meshalkin [11], [12]. A way out of the situation is seen in the refusal to present a decision to a user accepted by DSS, but using a mechanism of decision making for creation of an informational environment which would promote a correct decision making by the user. It must be noted that giving up an already accepted decision is not our invention. As early as Platon in his treatise “Taetet” indicated that Socrates had helped his disciples to make correct decisions solely with the help of leading questions. The article covers the present notion on such systems on the basis of the first experience of their creation, and the inference engine for the systems of the Dr.Watson Type. The second paragraph gives definition of the systems of the Dr.Watson Type and a scheme of principle of the decision support algorithm in Dr.Watson Type Systems (WTS). The third paragraph describes a concrete appendix of the algorithm in WTS for an intensive therapy department physician, and briefly considers other methods of decision making support in WTS' ideology. The forth paragraph considers the experience of WTS' creation and application for an intensive therapy physician. The fifth paragraph is conclusion and summary. 2. Dr. Watson Type System. Ways to stimulate thinking is one of the directions of practical psychology. Common regulations are known well enough [14], [15]. This is necessity: 1) to present all the necessary information for decision making; 2) to create a conflict situation in the informational space for stimulating the user; 3) to direct the user's reflections towards a correct decision; 4) to provide a positive mind-set to the solution of a task; 5) to provide a motivation for continuing the work at a task. The methods of such a kind are actively applied for organization of a creative process by psychologists, specialists of system analysis, invention machines a distinctive feature of Dr. Watson systems is the presence of inference engine in them. Though the decision themselves are not offered, namely on the basis of making decision algorithms and results of their work, WTS create the above indicated conditions (W-technologies). Another specific feature is self-tuning of the system to a concrete specialist solving this problem, but not to a problem which is being solved (as it is done in invention machines). Let’s name as a Dr. Watson Type System the decision making support system a) possessing the conclusion algorithm, but using it for creating the conditions 1-5, but not for presenting a decision b) having an adaptation mechanism DSS to a concrete user. Below is given an example of the algorithm realizing W-technology of making decision for the system of Dr.Watson. Let specialist consider situation and according to information ( - whole possible information about ) make decision i( ii=1,..., n all the possible decisions). The rest of the information about can be taken in the form of answers to questions. Each question is put by the set of J answers to the question. There are the thresholds D and the making decision algorithm F, which for each decision i calculates the membership function f ( i , X ) . If f ( i , X ) D ( i ) , then F takes the meaning i. Action of WTS: In the beginning , D ( ) is formed taking into consideration a probability error of the specialist M in determining the decision i ( the first sort error ) D ( , M , 1) D( ). If f ( , X ) D(, M , 1) , then WTS displays itself in no way. If f ( , X ) D ( , M , 1) , then M receives additional information in the form of indicating elements from X , which do not increase the value f ( , X ) . Of the set \ X there is selected an admissible ( not contradictive X ) element j increasing the value f ( , X & j ) in the greatest manner and the question J containing j is put to a specialist. In case if in \ X admissible elements increasing f ( , X ) are not left then there must be a question with an admissible element j which mostly increases assurance in the alternative hypothesis D ( k ) --f ( k, X & j ) = min (D ( i ) --f ( i, X & l)) iil \ X If the answer j to the question J increases f ( , X & j ) , then D( , M , 2 ) decreases in comparison with D ( , M , 1) . If the answer j to the question J does not increase f ( , X & j ) , then D ( , M , 2) = D ( , M , 1). After the answer to the question, WTS continues its work with new set of thresholds and so on until the decisions M and F are concerted between themselves or M stops work with WTS. Let us consider WTS' influence on the accuracy of making decision by a specialist. Let us assume that in situation it is required to make decision . G ( M , X ( A) ) - decision of specialist M on information X (A) ; P (G ( M , X ( A) ) = ) - probability of making M decision ; p ( F ( X(A)) = ) - probability of making F decision ; If G ( M , X ( A) ) = F ( X(A)) , then even with P >> p probability of the error ( M - WTS ) decreases and makes: ( 1 - P (G ( M , X ( A) ) = ) ( 1 - p ( F ( X(A)) = ) ) < 1 - P (G ( M , X ( A)) = ) If G ( M , X ( A) ) F ( X(A)) , but as a result of the above indicated procedure there is an increase of M knowledge about the situation A, then in this case as well there is a decrease of probability of the error ( M -WTS ). It is clear that the assumption that the accuracy of making decision increases with the increase of information about a situation, is sufficiently strong and all the problems connected with WTS' creation consist in making the assumption a real one. Formally, the expert systems constructed on the principle of adverse conclusion begin their work from a conclusion analysis of the user as well, and, in the preocess of the work, ask the user questions. However, these questions serve for the control of the conclusion from the point of expert systems and because they are not fastened neither to a diagnostic process nor to a concrete user like in the Dr.WTS. Moreover, these queries are not similar to an indirect leading information which is proposed in the below described subsystems “ Effectiveness” and “ Adequacy” of the system for a physician of the intensive therapy department. 3. W- technologies in DSS for a physician of the intensive therapy department. The System of supporting decision making of a therapeutic department physician of the in-hospital ("System Analysis Guide" or "SAGe") was supposed that the physician cures several patients simultaneously, but immediate medical procedures are carried out by medical nurses or doctor's assistants. They perform information entry into the system as well. As a result of the program activity there appears an analog of a case history. Before creating the system there was held a questioning of 8 physicians-intensivists with no less than two years experience of work with DINAR.2 [16] . They were offered a list of 25 possible functions of the hypothetical ideal system. Each function was to be evaluated by a 5 point-grading scale as to the usefulness for a physician. The most useful functions were recognized : treatment adequacy assessment (point sum = 40); treatment efficiency assessment ( 39 points ) ; and state dynamics assessment ( 39 points ) in distinction for example from treatment selection (30 points) and diagnostics (27 points). Namely these 3 problems were considered as central ones in the system being under construction, though the system had to help in solving other intellectual problems, such as diagnostics, state development prognoses and so on. Necessity to solve such complex, numerous and versatile tasks has made the choice of Dr.WTS' approach not only desirable, but forced as well. The SAGe is written on the language C++ for a 486SX- compatible personal computer and consists of 5 subsystems : "Diagnostics", "Treatment Effectiveness", "Treatment Adequacy", "Integral assessment of patients in the department" and "Information import". 3.1 Diagnostics. A task of the given subsystem is help in selection of a diagnostic hypothesis about pathologic syndromes and assessment of a patient state severity. For assessment of state severity a special scale [17] was used. For helping selection of leading pathologic syndromes there has the algorithm of W-technology realization (proposed above). been realized As f ( i , X ) was considered: r (i , x ) x X Where -- hypothesis i . < r (i , x ) < + the sign weight x in relation to In case if the first kind error for M when determining then If e ( , M) < 0.1 then D ( , 2 ) . If e ( , M) 0.1 then D ( , is the e ( , M) M , 1) = D ( ) + D ( ) ( 0.1 - e ( , M) M,1)=D( ) . In the work process of the algorithm, if there is T-step, when receiving the information j. f ( , X & j ) >> f ( , X ) , then D ( , M , T +1 ) = D ( , M , T ) + (f ( , X & j ) - f ( , X ) ). After putting the two questions, which are not dependent upon the physician's actions, SAGe's diagnostic algorithm stops its work and doesn't inform about its conclusions. An important role in the diagnostic subsystem is played by information visualization with the help of special circles painted in a special manner. The graphic image, original for every case and reflecting all the collected information, will better help to assess the specific situation, to remember a like case, to determine contradictions and empty spaces in information. 3.2 Treatment efficiency. The task of the given system is help in selecting treatment and determining its efficiency. For the assessment of the treatment effectiveness there was used a multidimensional evaluation of an organism reaction on the current therapy. Changes of integral assessment of the patient’s state severity cannot be the only index of this notion . As a treatment effectiveness can be considered avoiding of complications or stabilization of the patient’s state when decreasing intensity of the treatment or improving separate concrete indexes. On the other hand, when there is a common improvement of the patient’s state, an unfavorable dynamics of separate indexes can turn out with irreparable problems in future. In this connection for effectiveness assessment the SAGe gives the physician a state severity dynamics and changes of character and intensity of leading pathological syndromes in comparison with dynamics of type and intensity of a treatment. Description of the treatment is considered in conformity with [18]. At will the physician can compare these curves with changes of any quantitative and qualitative signs. On its part, the SAGe, by itself, offers dynamics of signs, which from "its point of view", needs attention when selecting a treatment. The five groups of such signs are considered: 1 group - indexes reflecting a lesion of an organism system not affected by a disease by this moment; 2 group - indexes with a paradoxical dynamics ( for example, indexes majority increases, but a meaning of the given index decreases); 3 group - indexes with the most severe dynamics; 4 group - the most threatening indexes; 5 group - indexes earlier studied by the user in this situation. For each of the group indexes , the SAGe gives a rank from 1 to 7 depending upon its potential importance. Irrespective of the groups the user is offered 4 quantitative and 2 qualitative signs with the most great ranks for demonstration. If in this case the number of the signs with the same rank is too great, then preferences are formed according to the group numbers. An important role in achieving the subsystem's purposes is played by the requirement to the physician to make independently a disease development prognosis in response to the treatment administered by him. In case of discrepancy of prognosis with reality, the physician must find explanations of the discrepancy, causes for state aggravation and causes for treatment inefficiency, and indicate evidently the causes in the SAGe. Experience of using similar questions in DINAR.2 revealed their great stimulating effect for comprehension by the physician of his own actions [17], and SAGe itself, tries to assess the treatment efficiency. However, its conclusion is very shifted towards the negative response and therefore a priori it can't be directly used by the physician. 3.3 Treatment adequacy. The purpose of the given subsystem is help in determining adequacy of an administered treatment. We suppose that the adequacy is a harmony between a treatment and an organism. It is not only to treatment effectiveness. The effectiveness can be attained by an excessive therapy. In a certain sense, an adequate therapy for an organism. However, necessary for us therapeutic resources can be absent or be limited. With the limitations being accounted, the most adequate therapy can be ineffective. But at least the adequate therapy should not be blind. Selecting the most adequate therapy out of possible ones, the physician should understand to what consequences this therapy can lead and what cause of the present state of the organism is. On the presentation and analysis of the physician’s reflections about the situation with a patient there, the system of supporting making decision in ‘Adequacy’ has been constructed. A special cognitive graphics is used. The whole course of treatment and pathologic process development and mainly the physician’s reflections dynamics is represented in the form of the growth of the flower. Each element of this picture responds to a certain side of this process: the root is anamnesis, the stem is the curve of changing the state severity, the branches are unrealized prognoses, the floscule is the final state of physiological parameters (an analog of the circle from "Diagnostics") , the leaves are leading pathological syndromes, props and strings are the rendered treatment , the atmospheric phenomena are reasons for the aggravation of the state , the roughness of the soil and problems with props and strings are reasons for inefficiency and inadequacy and so on. Difference in time of the reflections (for example, an assessment of treatment adequacy occurs in time of treatment administration, and an assessment of its efficiency - in the subsequent moment of time), control of the reflections by the pathologic process course, necessity of forecast the future generate difficulties of coordination and direct contradictions in the reflections. The view and number of these reflections give the physician assessment of his understanding the treatment course. In its turn, SAGe tries to assess the degree of controllability of the therapeutic process. According to its index the background intensity of the cognitive picture of “Adequacy” is selected.. By disorders of naturalness in drawing pictures, SAGe reflects contradictions in the physician's reflections about reasons for prognosis discrepancy, reasons for state aggravations, reasons for inefficiency and inadequacy . For example, a treatment was recognized by a physician as adequate and the prognosis is supposed as a state improvement, however the state has been gone bad. Direct logical contradictions are obviously indicated, such as for example : the cause of the treatment inefficiency is its wrong use and the cause of an aggravation is the wrong administration of the treatment. 3.4 . Integral assessment of patients in the department. The given subsystem starts and completes the work with patients. Display screen with the system action presents the department plan with its wards and medical beds. Each bed is a window able to reflect interesting information about a patient on it. Using information from other subsystems we get data on patient with leading severity of state, unfavorable state dynamics, financial expenses on treatment, medical complexity of a case etc. The physician has a possibility to receive integral characteristics and reports on an individual patient or on all patient treated or on patients from the archive. In the subsystem the physician indicates what data about the state and treatment of a patient and with what regularity must be entered by a doctor's assistant on each concrete patient in subsystem of assessment of diagnostics and treatment. 4. SAGe as an example of Dr.Watson system. When SAGe was system have been subsystems. under construction the basic principles of Dr.Watson formulated. The principles determined creating each of the Creation of conflict situations in the " Diagnostics" is provided with revealing information which doesn't go into a diagnostically hypothesis of a physician; in the “Efficiency"- with a disparity between prognosis and reality, in the "Adequacy" - with showing contradictions in the physician's reflections. Attempts to change the direction of the physician's reflections occur in the "Diagnostics" with the help of directing questions; in the "Efficiency" with showing signs demanding an attention; in the "Adequacy" - with showing possible mistakes in previous assessment of treatment adequacy. In all the subsystems algorithms make their choice of signs, necessary for stimulation of the physician, on the basis of decisions of final tasks of the subsystem accepted by SAGe. Algorithms developed for SAGe possess an apparatus of self-tuning for a concrete user though now in the Working Version, this part is absent. Visualization of information in each of the subsystems is organized in such a way to stimulate the physician's intellectual activity. Stimulating motivation for work with the program is necessary for every computer system, though for the Dr.Watson type system it is especially important. So, as to SAGe a specialist must turn to the system of decision making support for help in the most complex question, while the system makes some vague hints instead of giving actual answers. We hope that in due course this will be natural for a physician, however at the start, a mighty stimulating motive for such an activity is necessary. Such a stimulating motive for the system as a whole is the entire first subsystem "Integral assessment" and namely: making the case history of a concrete patient and a case history archive, a possibility integral information and a scope of all patients at once. to receive an For separate subsystems as a motivation can be considered: - for the "Diagnostics" - obtaining a standard assessment of a patient's state severity necessary for a physician; - for the "Efficiency" -a possibility of comparison of changes of various parameters, and evaluation of the therapy invasiveness; - for the "Adequacy" - creating an archive of the physician's own reflections and his problems arising from treatment of the patient. Nevertheless the physician can ignore one or some subsystems. Information about his activity entered into SAGe, is accessible only for himself. The input of SAGe is protected with a parole, and the entry of information into SAGe, with the help of the physician's assistant, takes place in a special subsystem of input. The first results of work reveal a common favorable attitude to SAGe. Among the intellectual subsystems, the greatest success belongs to the "Treatment efficiency", then the "Diagnostics", and finally the "Treatment adequacy". However, during the work there appears greater interest in the two last subsystems. For a while these assessments and dynamics of their changes correspond to our prognoses. However an exhaustive analysis of work with SAGe is still ahead. The task of the following stage of SAGe elaboration is a realization of algorithms W-technologies in complete amount with their self-tuning for users and constructing a communicational subsystem . The communicational subsystem must connect SAGe with informational flows in hospitals, with available informational and reference systems. 5. Summary. The systems of Dr.Watson Type are not, of course, a all the problems of making decision systems. universal solution of Moreover, if in a concrete WTS a sufficiently strong direct motivation for using this system is not made, indirect help in decision making can turn out to be absolutely unclaimed. However, a look at the problem under solution through the user solving this problem can turn out to be very fruitful. Even in a standard approach to decision support it is relevant to use statistics of the user's work. Taking into account subjective distortions of information will help to make a correct decision and, besides, knowledge of his own conclusions indicates a possible direction of the subconscious misrepresentation [19]. On the other hand, in the systems of Dr.Watson Type there can be a direct supposition of a decision as a stimulation of the user's reflections. It is only important here that this decision should not be perceived as a guidance to action. It can be provided, for example, with an intended hyper diagnostics as in the subsystem "Treatment efficiency" of the system - SAGe. The central idea of WTS sufficiently broadly interpreted. WTS know a genuine decision, however they use this knowledge only for creating optimum conditions for independent creativity of the user. On the basis of this idea, there can be constructed not only the systems of decision making support. There are already concrete projects of creation of teaching games, special systems for working out collective decision, organizing business-like games. References 1. Milne R Business benefits of expert system application. IFS Conference on AI in Manufacturing, London. 1990. 2. Glazunov V. N. The quest for principle of technical system action . Moscow. 1990. 3. Heathfield H.A., Wyatt J. Philosophies for the Design and Development of Clinical Decision-Support Systems. Meth. Inform. Med. , 32, 1993, p. 1-8. 4. Shortliff E.H. Knowledge-based systems in medicine. Medical Informatics Europe 1991, Vienna, Austria, 1991,p.5-10 5. Miller R.A., Maserie F.E. The Demise of the "Greek Oracle" Model for Medical Diagnostic Systems. Meth. Info. Med., 1990, v.29, 6. Clarke K., O'Moore R., Smeets R., Talmon J.,Brender J., McNair P., Grimson J., and Barber B. A Methodology for Evaluation of Knowledge-Based Systems. In: Medical Informatics Europe 1991, Vienna, Austria, 1991, p.361-366. 7. Milne R. Second Generation Expert Systems: The Applications Gap . Expert Systems & Their Application, Avignon , France, 1991, p. 259 - 264. 8. Gorbov F.D. About operator's tolerant to handicap. In: Industrial psychology. Moscow. 1964 9.Goldberg S.L. Diagnostics on the basis of the informative space of the antisyndromes //Problems of Control and Information Theory.Budapest.-1984.-Vol.13-pp.401-411. 10. Krupin E.N., Charnis M.Y., Telesheva T.V., Leontiev S.L. Goldberg S.I. Computer Aided Expert System of Diagnostics and Tactics in a Urgent Nevrologic Pathology. In: Medical Informatics Europe 1991. Vienna, 1991. p 251-255. 11. Goldberg S, Meshalkin L. Assisting Dr.WT - Systems. In East - West Conference on Artificial Intelligence: From Theory to Practice, Moscow, Russia, 1993, p.207-209. 12. Goldberg S.I., Meshalkin L.D. A New Class of AI-Systems ( DrWT-Systems ). Technical Cybernetics. 5 , 1992, p. 217-223 13. Yates H.J. Frustration and conflict . London. 1962 14. Altshuler G.S. Creation as exact science. Moscow . 1979 15. Granovskia R. Elements of practical psychology. Leningrad . 1988 16. Goldberg S.I., Lomovskikh V.E., Makhanek A.O., Sklyar M.S. Expert system "DINAR-2" - methodological basis for the pediatric emergency aid organization in a large region. In: Medical Informatics Europe 1991, Vienna, Austria, 1991, p.270-274. 17. Goldberg S.I. , Tsibulkin E.K., Makhanek A.O., Sklyar M.S. Expert system of the regional pediatric reanimation consultative center - DINAR2. Intensive Care Theatre and Oncology. AHM .1993 p. 33 - 35. 19.S.I.Goldberg, A.O.Makhanek, I.D.Novikov. Dispatching-Consultative Expert Systems. In: Herald of VIMI. Vol. 1 , 1991 . p 46-52