Explainable Risk Stratification of Polypharmacy-Related Adverse Outcomes in the Elderly via Temporal Feature Engineering and Rule-Augmented Gradient Boosting

Abstract

Polypharmacy in community-dwelling older adults is associated with a sharp rise in adverse drug events, falls, emergency department visits, and avoidable hospitalization, yet most existing risk stratification tools rely on static cross-sectional inputs that overlook the temporal dynamics of prescription churn and drug-drug interaction onset. This paper introduces a rule-augmented gradient boosting framework that combines time-windowed prescription features with curated pharmacological knowledge to stratify elderly patients into low, medium, and high adverse-outcome risk strata. Records from the FDA Adverse Event Reporting System (FAERS) are linked at a patient-archetype level with MIMIC-IV medication trajectories and a NHANES elderly subset, yielding 184,627 patient-episodes after harmonization and deduplication. We construct 142 temporal features-rolling prescription counts, drug-class entropy, anticholinergic burden ramp-up, concurrent-prescription synchronicity, and CYP-substrate co-exposure densities-and inject Beers Criteria and STOPP/START expert rules as soft constraints through a rule-aware loss term within the LightGBM objective. SHAP-based attribution provides clinician-readable explanations at both global and patient-instance levels. The pipeline attains AUROC = 0.9127, F1 = 0.8261, and Brier score = 0.0814, outperforming logistic regression, vanilla LightGBM, XGBoost, and a temporal convolutional network baseline by between 2.31% and 4.46% absolute AUROC. Ablation shows that temporal features contribute 1.94% AUROC and rule injection contributes 1.27%. The framework supports clinical deployment with sub-50 millisecond per-patient inference and stable performance across age and sex subgroups.