1. Enhancing Clinical Reasoning Using Probabilistically Generated Virtual Patient Cases Jeroen Donkers, Bep Boode, Danëlle Verstegen, Bas de Leng, Jean van Berlo Maastricht University

2. Contents  Clinical reasoning and uncertainty  Bayesian belief networks  From belief networks to virtual patients  Example: Diabetes  Conclusions

3. Clinical Reasoning and Uncertainty  During clinical reasoning, a doctor has to deal with:  Indirect observations and measurements  E.g. Vague complaints, measurement errors, etc.  Partly understood processes  Influence of drinking red wine on diabetes?  Genetical, epigenetical (and metagenetical) variations  How will this particular patient react to that medication?  Incomplete and contradictory medical literature  E.g., only data available for men >80 year and obese children.  Population characteristics  Maastricht versus London  All these factors lead to stochastic variables in the reasoning process

4. Clinical Reasoning and Uncertainty  Still, a doctor has to make quick decisions:  ordering the items in a differential diagnosis  Requesting the most informative lab test  Choosing the most effective therapy  Doctors (and all other humans) use subjective probabilities  Unconsiously  Mostly experience-based  Biased by many factors – availability, representativeness, etc.  Only partly supported by scientific and epidemiological data (evidence based)  Errors in subjective probabilities may lead to wrong decisions (Elstein & Schwarz, 2002, BMJ)

5. Improving Subjective Probabilities  Problem-based learning and evidence-based medicine are important educational means to improve clinical reasoning – increasing the use of objective data (Elstein & Schwarz)  However, the set of patient cases presented is often biased  Both can be supported by virtual patients  Support theory suggests that extensive case descriptions (e.g. VPs) can bias subjective probabilities (Bergus et al.)  Can we improve subjective probabilities by populations of VPs?

6. Populations of VPs  Could:  Lead to formation of better subjective probabilities by preventing some biases in probability estimation  Should:  Be large enough to enable learning  Contain small cases to prevent support bias  Represent realistic patient population  How to generate?  Using a mathematical simulation model (e.g. Archimedes)  Using probabilistic models

7. Bayesian Belief Networks (BBN)  Judea Pearl, 1988  Compact graphical model of a joint probability distribution  Structure: models conditional independencies  Basis: P(A|BC) = P(A|B)  A is independent from C, given B  Parameters: conditional probabilities  Graphic representation enables knowledge elicitation  Supports prognosis as well as diagnostic reasoning  P(disease | symptoms) and P(effects | disease)  Allows fast computations (querying)  Robust for small errors in probabilities  Can generate populations (by sampling)

8. Example: Diabetes Mellitus

9. From BBNs to VPs  Step 1: generate a sample, using proper population settings  Step 2: instantiate sampled data by selecting realistic values for qualitative categories

10. From BBNs to VPs  These case descriptions can be transformed into VPs:  Transform into MVP compliant package  Directly import into VP system  However, a more preferable option is to let the VP system generate a new instance on demand  Option a: integrate the BBN into the VP system  Option b: compile the BNN into a VP with stochastical nodes

11. Pilot study  To measure the effect of VP populations we have planned the following controlled experiment:  3 groups students: A B C  2 VP populations: one sampled and one hand-picked  Group A (control): no intervention  Group B: hand-picked VP population  Group C: BBN-generated VP population  VP populations are pre-generated to enforce comparability  VP cases are presented using SurveyMonkey  MCQ to stimulate and test probability formation

12. Conclusions  Subjective probabilities play a major role in clinical reasoning, but can easily be biased  VP populations could lead to improved subjective probabilities  BNNs can generate VP populations and are less difficult to build than simulation models  However, it is still delicate and laborious  Connections to VP system have yet to be developed  Pilot will reveal the effectiveness of the approach