RxGuard: Knowledge-Guided Safety Guardrails for Medication Recommendation

Abstract

Medication combination recommendation predicts a set of medications for a patient visit from longitudinal electronic health records (EHRs) and is safety-critical, since feasibility depends on set-level constraints such as drug--drug interactions (DDIs). Despite substantial progress, existing approaches suffer from three limitations: (1) safety is often imposed as a training-time soft regularizer, while inference relies on thresholding or top-k selection over independent drug scores, providing no executable per-instance feasibility control; (2) external knowledge is mainly used for representation enrichment rather than being compiled into inference-ready decision artifacts that directly support constraint execution and evidence retrieval; and (3) explanations are typically inclusion-centric, lacking structured, auditable rationales for constraint-triggered exclusions. To address these gaps, we propose RxGuard (Rx denotes prescription), a novel decision-oriented framework that formulates medication combination recommendation as utility maximization under explicit feasibility constraints. Technically, RxGuard learns trajectory-aware medication utilities and a candidate set from EHRs, compiles external knowledge into an executable feasibility interface and an Evidence KG for audit retrieval, and performs guarded set selection at inference time to enforce feasibility by construction while producing decision-consistent audit records with grounded rationales for both inclusions and exclusions. Experiments on MIMIC-III and MIMIC-IV demonstrate that RxGuard improves both recommendation performance and safety, while providing consistent, decision-aligned auditing.