Explainable Risk Stratification for Polypharmacy-Related Adverse Outcomes in Community-Dwelling Elderly: A Rule-Enhanced Machine Learning Approach

Abstract

Polypharmacy among community-dwelling elderly populations presents substantial clinical challenges, including elevated risks of falls, delirium, and hospital readmission. This study proposes a hybrid rule-enhanced machine learning framework for explainable risk stratification without requiring specialized clinical systems. The methodology integrates rule-based screening using established pharmacological risk dictionaries with gradient boosting algorithms to generate interpretable probability estimates for adverse outcomes. Patient medication lists are standardized to generic nomenclature and mapped to sedative burden scores, anticholinergic indices, and drug-drug interaction matrices. The framework outputs 30-day and 90-day readmission risk probabilities alongside actionable clinical recommendations. Evaluation encompasses high-risk detection recall, false positive rates, SHAP-based feature contribution analysis, and subgroup fairness metrics across vulnerable populations including living-alone, minority, and LGBTQ+ elderly cohorts. Results demonstrate the potential for reproducible, transparent algorithmic approaches to enhance medication safety review in community care settings while supporting health equity objectives.