Causal-Driven Machine Learning Approaches For Explainable Decision Making In Medical Diagnosis Support Applications

Abstract

Machine learning models for medical diagnosis assistance are becoming more commonly used alongside explainability approaches like SHAP and LIME, but the feature attribution approaches in these techniques identify those variables most statistically correlated with a prediction, which might include spurious correlations, for instance a comorbidity that happens to be correlated with a disease without causation. In this work we propose the Causal-Driven Explainable Diagnosis (CDED) approach that starts with learning a patient-cohort structural causal graph on symptoms, lab results, comorbidities, and diagnoses with constraint-based causal discovery, then feature attributions are computed as causal effects, and finally counterfactual explanations are generated that describe how much change would have been required to alter the diagnosis decision. There is a clinician-in-the-loop validation step that records disagreements and flagged spurious explanations into a feedback buffer for periodic causal graph refinement. The proposed approach is evaluated on clinical data simulated from the MIMIC-III critical care dataset in four different scenarios of increasing complexity: standard scenarios, confounded scenarios, rare disease scenarios, and conflicting symptom scenarios. As against the black-box classifier which offers no explanation and the association-based SHAP approach to explainable classification, the proposed CDED methodology exhibits 81.5 percent accuracy rate in diagnosing conflicting symptom cases, in comparison to 58.6 percent and 63.4 percent of the two baselines respectively, as well as delivering considerably higher explanation fidelity and clinician trust scores. This highlights the advantage of utilizing causal effects, in addition to statistical association, in generating explanations for improving the quality of diagnostic predictions.