A SHAP-Informed Formal Feature Attribution Framework for Drug-Drug Interaction Risk in Large-Scale Claims Data

Pairwise drug-drug interaction databases flag co-prescribed pairs, but they under-weight multi-drug combinations that drive adverse drug events in older adults. We studied that gap in Virginia All-Payer Claims Database records (2016-2019) for adults aged 65-114 years with non-opioid emergency department visits (n = 1182 cases; 16,105 matched controls across three geriatric age bands). Features came from pharmacy and medical claims in a short pre-index window optimized for acute ADE timing (21-30 days by age band; 21 days for ages 65-84). We trained gradient-boosting models separately among patients with similar pre-index healthcare contact volume (2016-2018 training; 2019 holdout), then used Formal Feature Attribution to score drug pairs and triplets and Intervention Rate ranks to order deprescribing review. On the 2019 holdout, geriatric AUPRC was 0.101-0.335 (PR lift 1.6×-4.2×). FFA flagged 115 synergistic pairs and 312 high-confidence triplets (e.g., furosemide + hydrochlorothiazide + lisinopril; digoxin + furosemide + amiodarone, IE = 8.7). Top Intervention Rate drugs included simvastatin, furosemide, and alprazolam. Moderate preventive Z-code monitoring (Q2) was protective (OR = 0.25; 95% CI 0.18-0.34) versus no monitoring, while fragmented high-intensity monitoring (Q4) was not. The framework prioritizes medication combinations for pharmacist review in claims data; it does not replace pharmacokinetic confirmation or prove that changing a drug caused fewer ED visits.