Clinical Decision Support Dr Jeremy Rogers MD MRCGP Senior Clinical Fellow in Health Informatics Northwest Institute of Bio-Health Informatics Talk Outline Why we need it What does ‘decision support’ mean ? Work so far Why we don’t use it Talk Outline Why we need it What does ‘decision support’ mean ? Work so far Why we don’t use it Drivers for decision support ► Growth of medical knowledge ► Approx 100 articles were published in 1966 from RCTs; ► Over 10,000 annually by 1995 (Chassin, 1998) ► ‘The scarcely tolerable burden of information that is imposed taxes the memory but not the intellect’ (GMC 1993) ► Pressures to use knowledge ► Evidence based medicine ► National service frameworks ► Clinical Governance ► Cost – e.g. $5.5M in 37 Days for one patient at Duke ► ‘Post genomic’ individualised medicine Drivers for decision support ► Public recognition of medical error ► IOM “To err is human” (2000) & “Crossing the quality chasm” (2001) ► More people die from medical errors than from breast cancer or AIDS or motor vehicle accidents ► Jessica Santillan case 17 year old who had a heart and lung transplant from a donor with an incompatible blood group in Feb 2003 at Duke, and died after a re-do 13 days later Committee on Quality of Health Care in America US Institute of Medicine : Quality Chasm Report, 2001 (The American) health care delivery system is in need of fundamental change The current care systems cannot do the job Trying harder will not work Changing systems of care will Talk Outline Why we need it What does ‘decision support’ mean ? Work so far Why we don’t use it Kinds of decision Diagnosis Intervention Prognosis Kinds of support ► Active vs Passive support ► Making specific suggestions – one off, or continuing ? ► Critiqueing recorded actions – screw-up detection ► Tweaking / filtering information display ► Intelligent image processing ► Reminders ? Alerts ? ► Decision support, or decision making ? ► Do we expect human to learn from device ? Drowning in data The case for DS in display filtering EPR - Dr Kildare - 26th Oct 2000 John Doe 36 yrs Engineer Married, 2 children Encounters 12.10.96 Coryza: chest NAD: reassure 13.10.96 URTI: wheezy: amoxycillin 20.10.96 Anxiety: child admitted to H: reassure 24.10.96 PEFR : 300 : 10.11.96 PEFR : 400: CXR requested 12.11.96 CXR Basal Consolidation: : erythromycin 27.11.96 : Chest clear : 07.03.97 Depression: death in family: paroxetine 19.04.97 Gastoenteritis: : reassure 01.06.97 : : rpt Rx paroxetine 18.10.97 Sick note : : 03.03.98 Viral URTI: PEFR 350: salbutamol 04.03.98 WCC NAD : : 30.06.98 PMR report : BP, ECG NAD : 15.09.98 Eczema : : hydrocortisone 05.11.98 Depression : : paroxetine 03.01.99 Fibrositis: trigger spot lwr back: ibuprofen 17.02.99 Allergic Asthma: PEFR 300: salbutamol 21.03.99 Chest Inf: L base: erythromycin 07.10.99 Med4: anxious : 26.01.00 Asthma Review: :Repeat Rx Salbutamol Active Problems Current Medication Asthma Salbutamol Hydrocortisone Letters Results Appt This Visit Code PEFR Asthma C/o Low Mood Notes 550 l /min Chest NAD. No Problems. Declined antidepressant BP Action Salbutamol inh 2 puff qds 1op Influvac im BN #035679A4 PEFR WCC Drowning in data The case for DS in display filtering EPR - Dr Kildare - 26th Oct 2000 John Doe 36 yrs Engineer Married, 2 children Encounters 12.10.96 Coryza: chest NAD: reassure 13.10.96 URTI: wheezy: amoxycillin 20.10.96 Anxiety: child admitted to H: reassure 24.10.96 PEFR : 300 : 10.11.96 PEFR : 400: CXR requested 12.11.96 CXR Basal Consolidation: : erythromycin 27.11.96 : Chest clear : 07.03.97 Depression: death in family: paroxetine 19.04.97 Gastoenteritis: : reassure 01.06.97 : : rpt Rx paroxetine 18.10.97 Sick note : : 03.03.98 Viral URTI: PEFR 350: salbutamol 04.03.98 WCC NAD : : 30.06.98 PMR report : BP, ECG NAD : 15.09.98 Eczema : : hydrocortisone 05.11.98 Depression : : paroxetine 03.01.99 Fibrositis: trigger spot lwr back: ibuprofen 17.02.99 Allergic Asthma: PEFR 300: salbutamol 21.03.99 Chest Inf: L base: erythromycin 07.10.99 Med4: anxious : 26.01.00 Asthma Review: :Repeat Rx Salbutamol Active Problems Current Medication Asthma Salbutamol Hydrocortisone Letters Results Appt This Visit Code PEFR Asthma C/o Low Mood Notes 550 l /min Chest NAD. No Problems. Declined antidepressant BP Action Salbutamol inh 2 puff qds 1op Influvac im BN #035679A4 PEFR WCC Goal of support ► Influence outcome ► Good things more likely; bad things less likely ► Outcomes… ► Fatal events are only the tip of the iceberg ►Easiest to measure, and most dramatic, but…. ► Non fatal events ►Side effects ►Sub-optimal treatment ►Inappropriate treatment ► Non harmful events ►Inefficiency & Confusion ►Inappropriate resource consumption ►Bed stay ►Repeated re-investigation Kinds of DS technology ► Statistical ► ‘93.467% of the time, things that quack and have webbed feet are ducks” ► Model-based ► ‘It’s definitely a duck because you told me its mother was a duck’ ► Neural Networks ► ‘Of all the things you’ve shown me so far, it looks most like the ones you said were ducks.’ Talk Outline Why we need it What does ‘decision support’ mean ? Work so far Why we don’t use it The Story so far… “Three decades of research into computer aids for medical decision making have resulted in thousands of systems and a growing number of successful clinical trials…” BMJ 1997;315:891 (4 October) Projects past and present ► Acute Abdominal Pain (1972) ► Mycin (1977) ► ProFORMA ► Internist/QMR (1980s) ► Protégé ► DXPlain ► PRODIGY (1997-) ► ILIAD ► Prescribing Indicators ► Sophie ► Isobel (2000) ► Medical Logic Modules ► NHS Direct (2000) (Arden Syntax) (1989) ► Knowledge Coupling (PKC.com) 1970s & 1980s Behold, the Oracle… Mycin – Shortliffe 1970s, Stanford > (mycin) What is the AEROBICITY of ORGANISM-1? aerobic ------ PATIENT-1 ------ Is Sylvia Fischer a compromised host? yes Patient's name: Sylvia Fischer Is Sylvia Fischer a burn patient? If so, mild or serious? why Sex: female It is known that: Age: 27 1) THE SITE OF THE CULTURE IS BLOOD ------ CULTURE-1 ------ 2) THE GRAM OF THE ORGANISM IS NEG From what site was the specimen for CULTURE-1 taken? blood 3) THE MORPHOLOGY OF THE ORGANISM IS ROD How many days ago was this culture (CULTURE-1) obtained? 3 Therefore, ------ ORGANISM-1 ------ Rule 52: Enter the identity (genus) of ORGANISM-1: unknown The gram stain of ORGANISM-1: ? A GRAM must be of type (MEMBER ACID-FAST POS NEG) The gram stain of ORGANISM-1: neg If 1) THE BURN OF THE PATIENT IS SERIOUS Then there is weakly suggestive evidence (0.4) that 1) THE IDENTITY OF THE ORGANISM IS PSEUDOMONAS Is Sylvia Fischer a burn patient? If so, mild or serious? serious Is ORGANISM-1 a rod or coccus (etc.): rod Findings for ORGANISM-1: What is the AEROBICITY of ORGANISM-1? why IDENTITY: ENTEROBACTERIACEAE (0.800) PSEUDOMONAS (0.760) It is known that: 1) THE GRAM OF THE ORGANISM IS NEG Is there another ORGANISM? (Y or N) Y 2) THE MORPHOLOGY OF THE ORGANISM IS ROD ------ ORGANISM-2 ------ Therefore, Enter the identity (genus) of ORGANISM-2: unknown Rule 107: The gram stain of ORGANISM-2: (neg .8 pos .2) If Is ORGANISM-2 a rod or coccus (etc.): rod 1) THE AEROBICITY OF THE ORGANISM IS AEROBIC What is the AEROBICITY of ORGANISM-2? anaerobic Then there is suggestive evidence (0.8) that 1) THE IDENTITY OF THE ORGANISM IS ENTEROBACTERIACEAE Findings for ORGANISM-2: IDENTITY: BACTEROIDES (0.720) PSEUDOMONAS (0.646) Abdominal Pain: De Dombal (1972) A multicentre study of computer aided diagnosis for patients with acute abdominal pain was performed in eight centres with over 250 participating doctors and 16,737 patients. Performance in diagnosis and decision making was compared over two periods: a test period (when a small computer system was provided to aid diagnosis) and a baseline period (before the system was installed). The two periods were well matched for type of case and rate of accrual. The system proved reliable and was used in 75.1% of possible cases. User reaction was broadly favourable. Abdominal Pain: De Dombal During the test period improvements were noted in diagnosis, decision making, and patient outcome. Initial diagnostic accuracy rose from 45.6% to 65.3%. The negative laparotomy rate fell by almost half, as did the perforation rate among patients with appendicitis (from 23.7% to 11.5%). The bad management error rate fell from 0.9% to 0.2%, and the observed mortality fell by 22.0%. The savings made were estimated as amounting to 278 laparotomies and 8,516 bed nights during the trial period-equivalent throughout the National Health Service to annual savings in resources worth over 20m pounds and direct cost savings of over 5m pounds. Computer aided diagnosis is a useful system for improving diagnosis and encouraging better clinical practice. Br Med J (Clin Res Ed) 1986 Sep 27;293(6550):800-4 Medical Logic Modules (Arden Syntax) maintenance: title: ;; filename: template;; version: 1.00;; institution: ;; author: ;; specialist: ;; date: 1993-01-01;; validation: testing;; library: purpose: ;; explanation: ;; keywords: ;; citations: ;; knowledge: type: data-driven;; data: ;; evoke: ;; logic: ;; action: ;; end: An MLM… maintenance: title: Check for adequacy of therapeutic anticoagulation with warfarin;; filename: warfarin_anticoagulation;; version: 1.07;; institution: Columbia-Presbyterian Medical Center;; author: Randolph C. Barrows, Jr., MD (barrows@cucis.cis.columbia.edu);; specialist: ;; date: 1994-04-28;; validation: testing;; library: purpose: To warn the health care provider that a patient maintained on warfarin is NOT in a therapeutic range for low-intensity or full-intensity anticoagulation. Low-intensity anticoagulation is defined as a prothrombin INR in the range of 2.00 - 3.00 (roughly corresponding to a PT in the range of 1.2-1.5 times control). Full-intensity anticoagulation is defined as an INR in the rage of 3.00 - 4.50 (roughly corresponding to a PT in the rage of 1.5 2.0 times control).;; explanation: ;; keywords: ;; citations: Scientific American Medicine;; …and (some of) its logic /* the INR-containing procedures */ storage_of_INR := EVENT { '32506~service event', ‘2256~presbyterian coagulation profile'; '32506~service event', ‘2302~stat coagulation profile' }; /* See if patient has a warfarin order. Probably need to add 31058 Bishydroxycoumarin Preparations Here I only want header table info, no components. Is it ok to say null components? */ (start_time, status, order_key, frequency):= READ LAST { 'dam'="PDQORD1", display_header'="TRSKF",'display_comp'=""; ; '28612~CPMC Drug: Coumadin 10 Mg Tab', '28613~CPMC Drug: Coumadin 2 Mg Tab', '28614~CPMC Drug: Coumadin 2.5 Mg Tab', '28615~CPMC Drug: Coumadin 5 Mg Tab', '29932~CPMC Drug: Ud Coumadin 10 Mg Tab', '29933~CPMC Drug: Ud Coumadin 2 Mg Tab', '29934~CPMC Drug: Ud Coumadin 2.5 Mg Tab', '29935~CPMC Drug: Ud Coumadin 5 Mg Tab', '33033~CPMC Drug: Coumadin 7.5 Mg Tab' }; Knowledge Couplers: PKC.com Larry Weed MD Some CPOE Success Stories ► ► ► ► ► ► Barnes-Jewish Hospital, St. Louis, Missouri 130 potentially dangerous drug interactions identified two-thirds of those involving the drug cisapride averted Brigham and Women’s Hospital, Boston 81% decline in medical errors after implementation CPOE 64% of decline due to first, and simplest, version of the technology, which included features such as predetermined lists of medications and doses, display of patient data, basic drug dosage, interaction, and duplication checking. Montefiore Medical Center, New York City 50% decrease in medication errors following CPOE Time from placing an order to its arrival in pharmacy reduced to two hours. Ohio State University Medical Center, Columbus, Ohio Length of stay decreased by two days following CPOE Pharmacy orders turnaround reduced by two hours Pharmacy charges per admission reduced by $910 University Community Hospital, Tampa, Florida 77% reduction in all adverse drug events, and 85% in severe ADEs Cost of drugs for one family reduced by more than $200,000 per year. Children’s Hospital of Pittsburgh 50% reduction in harmful error Virtual elimination of weight-related adverse drug events Complete eradication of transcription/handwriting errors 50% reduction in medication delivery times. Other successes… ► Strong evidence suggests that some CDSSs can improve physician performance. Additional well-designed studies are needed to assess their effects and cost-effectiveness, especially on patient outcomes (Johnston 1994) ► Mothers receiving computer-generated reminders had 25% higher on-time immunization rate for their infants (Alemi, 1996) ► Decision support system was safe and effective and improved the quality of initiation and control of warfarin treatment by trainee doctors (BMJ 1997;314:1252) ► Computerized physician order-entry reduced adverse drug events by 55% (Bates, 1998) ► 9% of redundant lab tests at a hospital could be eliminated using a computerized system (Bates, 1998) ► 74% of the studies of preventive healthcare reminder systems and 60% of the evaluations of drug dosing models reported a positive impact (Trowbridge & Weingarten, AHRQ, 2001) ..and some failures ► (PRODIGY) - No effect was found … on the management of asthma or angina in adults in primary care BMJ 2002; 325: 941-944 ► ..decision support system did not confer any benefit in absolute risk reduction or blood pressure control BMJ 2000;320:686-690 ► Computerised decision support systems have great potential for primary care but have largely failed to live up to their promise BMJ 1999;319:1281 My own failure: Prescribing Indicators ► General Practice Repeat Prescribing ► Patients get more drug without seeing doctor ►typically, enough for 1-3 months ► 35% of population at any one time on repeat Rx ► Medication Review ► Accepted part of good clinical practice ► Requirement in NSF for Older People ► But: signing authorities is daily batch process ►>30 scrips per GP per day ►No time for careful review What is ‘Medication Review’ ? Indicators of ‘quality’ prescribing ► Cantrill et al: 13 indicators: ► Dose too high or too low? ► Course too long ? ► Expensive or useless drug ? ► Interaction with another drug ? ► Contraindicated ? ► By brand ? ► REASON FOR USE DOCUMENTED ? ► Manual system: impractical ► Our project: (2000-2002) ► computerise the indicators Complex implementation.. Patient ID: Medication: Problem List: 4578 DITA906 DISR10514B 183... (Oedema) 1B17.. (Depressed) G5732. (Paroxysmal Atrial fibrillation) G73z0. (Intermittent claudication) H3.... (Chronic obstructive pulm.dis.) 137S.. (Ex smoker) 246... (O/E - blood pressure reading) 442... (Thyroid hormone tests) 44P... (Serum cholesterol) 7L172. (Blood withdrawal for testing) Ontology ID Product Rubric 345031(oral dig) DITA905 Digoxin 125 mcg tab 345031 DITA906 Digoxin 250 mcg tab 345031 DITA908 Digoxin 62.5 mcg tab 9099269 Systemic Digoxin G57.. Cardiac dysrhythmias G573. Atrial fibrillation and flutter G5730 Atrial fibrillation G5731 Atrial flutter G5732 Paroxysmal atrial fibrillation G573z Atrial fibrillation and flutter NOS IDENT “9099269” MAIN digoxin PROPERTIES HAS_DRUG_FEATURE physiological action WHICH_IS process ACTS_ON heart Indication Code Rubric Atrial fibrillation 14AN. H/O atrial fibrillation 3272. ECG: atrial fibrillation 3273. ECG: atrial flutter 7936A IV pacer control of A Fib G573. Atrial fibrillation / flutter 305084 Digoxin Liquid HAS_DRUG_FEATURE indication FOR treating ACTS_ON supraventricular arrhythmia HAS_DRUG_FEATURE indication FOR treating ACTS_ON atrial fibrillation HAS_DRUG_FEATURE information source IS_PART_OF interaction appendix 329308 Digoxin elixir 345031 Oral Digoxin tablet 305075 Digoxin injection 305093 Digoxin Paed inj ..and disappointing results ► Machine says there is no recorded indication in 33% of prescribing events ► BUT high false positive rate: 62% ► => it is wrong, most of the time ► Why ? Of all alerts where machine says ‘no indication’… BNF Omits 5% Idiosyncratic record 27% Human could infer 3% Mapping error 27% No Record 38% Problems with the oracle ► Painful data acquisition ► Exhaustive ► Includes exhaustive negative findings ►(which clinicians traditionally largely omit) ► Slow to use ► Poor support for clinical workflow ► Clinician is passive ► Infrequent recognised need 1990s – More modest aspirations ► Narrow Domain systems ► ECG interpretations ► Arterial blood gas interpretation ► Predicting drug-drug interaction ► Alerts and Reminders ► Out of range test flagging ► But plans for the oracle are resurfacing in expectation of imminent EPR Talk Outline Why we need it What does ‘decision support’ mean ? Work so far Why we don’t use it You can lead a horse to water… “Three decades of research into computer aids for medical decision making have resulted in thousands of systems and a growing number of successful clinical trials…” “Yet only a handful of applications are in everyday use” BMJ 1997;315:891 (4 October) Decision Support Systems in Use Today (2003) QMR PUFF HELP Diagnostic decision-support system for internists 1972 routine use Pulm onary function tests 1977 ? Know ledge-based HIS 1980 ?routine use ACORN Coronary care adm ission 1987 decom m issioned DXplain Liporap MDDB Epileptologists' Assistant Cancer, Me? Hepaxpert I, II, III Interpretation of acid-base disorders Managed Second Surgical Opinion System Colorado Medicaid Utilization Review System Geriatric Discharge Planning System Microbiology/ Pharmacy Expert System PEIRS NéoGanesh POEMS SETH Jeremiah Clinical Event Monitor VIE-PNN CEMS GermAlert Germwatcher Orthoplanner RaPiD DoseChecker Coulter® FACULTYT M SahmAlert Reportable Diseases TxDENT RetroGram Automedon ERA Therapy Edge ATHENA Clinical decision support 1987 routine use Dyslipoproteinaem ia phenotyping 1987 ?routine use Diagnosis of dysm orphic syndrom es 1988 ?routine use Nurse progress note assistant 1989 decom m issioned Patient cancer advice 1989 ? Hepatitis serology 1989 routine use acid-base disorders 1989 ?routine use Managed care 1989 ? Prescription quality review 1990 ? Patient discharge planning 1990 ? Drug sensitivity 1991 ?routine use Pathology reports 1991 decom m issioned Ventilator m anager 1992 2001 Post-operative care 1992 ? Clinical toxicology 1992 ?routine use Orthodontic treatm ent planner 1992 ?routine use Clinical alerts 1992 ?routine use Neo-natal parentral nutrition 1993 ?In use Mental health decision support system 1993 routine use Infection control 1993 ?routine use Infection control 1993 ?routine use Orthodontic treatm ent planner 1994 ?routine use Designs rem ovable partial dentures 1994 ?routine use Drug dose checker 1994 ?routine use Haem atology 1995 ?routine use Drug sensitivity 1995 ?routine use Infection control 1995 ?routine use Screeing dental patients 1997 ?routine use Decision support for drug regim ens for HIV-infected patients 1999 routine use Ventilator m anager 2001 routine use Web-enabled electronic decision support and referrals system for cancer 2001 Under evaluation Web-enabled decision support system for the treatm ent of HIV 2001 routine use DSS for the m anagem ent of hypertension in prim ary care 2002 routine use http://www.openclinical.org/aisinpractice.html Why ? – the domain ► Rigid criteria difficult to apply in chaotic settings ► Medical data doesn't fit quantised definitions ► Even complex decision support algorithms require simplified and standardised inputs by users ► And descriptive data is very hard to quantise ► Rules are situation specific ► localising decisions to available resource is costly ► When are decisions actually made ? ► To be effective, system needs to be physically available in situation where decision is made Why ? - the technology ► Highly mobile workforce vs highly static computers ► Slow computers ► Crude knowledge bases poor performance ► Lack of stats for bayesian approaches ► Crude KR technology for model-based ► Closed software architectures ► Can’t integrate 3rd party DS modules with EPR Why ? – the law ► Medicolegal aspect of EPR ► Confidentiality & Consent ► HIPAA ► Medicolegal aspects of DS technology ► Responsibility for action rests with clinician ► Systems that are as effective as clinician overall no help if behaviour includes obvious clinical howlers ► Burden of recording why did not follow DS advice Why ? – the people ► Poor data quality ► Numerical data easy to obtain ► Much of medicine not numerical ► Inconsistent data entry Data Quality (Frequency of recording per GP per year) READ CODE Sore Throat Symptom Visual Acuity ECG General Ovary/Broad Ligament Op Specific Viral Infections Alcohol Consumption H/O Resp Disease Full Blood Count Practice A 0.6 0.4 2.2 7.8 1.4 0 0 0 Practice B 117 644 300 809 556 106 26 838 Why? – the people ► Poor data quality ► I know what I’m doing ► Numerical data easy to obtain ► Perception of infallibility ► Much of medicine not numerical ► 88% of the time users requested to bypass PRODIGY (Beaumont 1988) ► Inconsistent data entry ► What happened to my clinical ► Reluctance to change clinical practice to fit the tool autonomy ? ► Interface issues BMJ 1999;318:1527-1531 ► Weed’s knowledge couplers ► Users intolerant of less than perfect performance BMJ 2003;326:314 Why ? - money Through more improved choice of initial antibiotics to treat pneumonia, a group of mid-west hospitals decreased complications, mortality rates and hospital days and costs… Improved management of diabetic patients through frequent e-mail communication can produce better outcomes and fewer visits… …but hospital revenues also decreased as patients shifted from higher paying to lower paying DRGs. …but lower physician group revenues under ‘fee for service’ payment. Summary ► Research and commercial products pre-date IOM by almost 30 years ► Widespread adoption has not occurred even where results were positive ► Significant hurdles remain ► Legal ► Technical - EPR is harder than it looks ► Human factors