Explainable Machine Learning for Predicting Adverse Drug Events in Older Adults with Polypharmacy: A Single-Center...

Area of Science:

• Gerontology
• Pharmacology
• Artificial Intelligence

Background:

• Polypharmacy significantly increases adverse drug event (ADE) risk in older adults.
• Accurate risk stratification for ADEs in this population remains a clinical challenge.

Purpose of the Study:

• Develop and evaluate explainable machine learning (ML) models to predict ADEs in older adults with polypharmacy.
• Enhance the interpretability of ML models for clinical application in medication safety.

Main Methods:

• Retrospective cohort study of 7505 older adults (≥65 years) with outpatient care.
• Developed logistic regression, random forest, and Light Gradient-Boosting Machine (LightGBM) models using demographic, comorbidity, medication, and laboratory data.
• Assessed model performance using discrimination, calibration, and classification metrics, with SHapley Additive exPlanations (SHAP) for interpretability.

Main Results:

• Random forest model showed high accuracy (0.810) and specificity (0.828).
• Logistic regression achieved the highest AUROC (0.705) and sensitivity (0.589).
• SHAP analysis identified medication count, sodium level, diabetes mellitus, and comorbidity burden as key predictors of ADE risk.

Conclusions:

• Explainable ML models demonstrate moderate yet clinically relevant performance in predicting ADEs for older adults on polypharmacy.
• These AI-driven approaches can support interpretable medication safety risk stratification in clinical practice.