Explainable Predictions of Adverse Drug Events from Electronic Health Records Via Oracle Coaching

Abstract

Information about drug efficacy and safety is limited despite current research on adverse drug events (ADEs). Electronic health records (EHRs) may be an overcoming medium, however the application and evaluation of predictive models for ADE detection based on EHRs focus primarily on predictive performance with little emphasis on explainability and clinical relevance of the obtained predictions. This paper therefore aims to provide new means for obtaining explainable and clinically relevant predictions and medical pathways underlying ADEs, by deriving sets of rules leading to explainable ADE predictions via oracle coaching and indirect rule induction. This is achieved by mapping opaque random forest models to explainable decision trees without compromising predictive performance. The results suggest that the average performance of decision trees with oracle coaching exceeds that of random forests for all considered metrics for the task of ADE detection. Relationships between many patient features present in the rulesets and the ADEs appear to exist, however not conforming to the causal pathways implied by the models - which emphasises the need for explainable predictions.