An Adaptive Multi-Agent Framework for Medication Safety: Evidence-Grounded Assessment, Medication Resolution, and Uncertainty Handling (Preprint)

BACKGROUND Medication safety assessment is complicated by ambiguous medication names, potential drug-related risks, and the need to distinguish reliable evidence from uncertain information. AI-based systems may assist with medication information processing, but their use requires careful handling of medication identity, evidence quality, and uncertainty. OBJECTIVE To develop and conduct a preliminary technical evaluation of an adaptive multi-agent framework for medication safety that combines medication identity resolution, evidence-grounded risk assessment, and explicit uncertainty handling. METHODS The framework was implemented using LangGraph and organized into ten modules for query processing, medication resolution, evidence retrieval, risk assessment, and response handling. Medication resolution used resources including RxNorm, openFDA, DailyMed, and curated medication data. A deterministic safety assessment component and an uncertainty-handling component were used to distinguish potential clinical risk from technical uncertainty. Evaluation included author-labeled offline test queries, live medication-identity checks, acceptance scenarios, and automated test suites. The evaluation focused on software behavior and technical performance rather than clinical outcomes. RESULTS In 42 author-labeled offline queries, medication entity resolution achieved a precision of 0.9167, recall of 1.0000, and F1 score of 0.9565. Routing accuracy was 42/42, and the false-positive rate for ordinary-word inputs was 0/22. Live identity matching succeeded in 14/14 checks, and safe handling of ambiguous or misspelled medication products was observed in 3/3 tested cases. All 18 acceptance scenarios passed. The reported test suites included 452 offline tests and 72 live tests. These results represent preliminary technical evaluations on small test sets and do not establish clinical effectiveness or patient safety. CONCLUSIONS The framework demonstrates the technical feasibility of combining multi-agent orchestration, medication identity resolution, deterministic risk assessment, and explicit uncertainty handling in a medication-safety application. The preliminary results support further evaluation using independently developed test sets, comparisons with baseline systems, expert clinical review, and appropriately designed clinical validation before any use in patient care. CLINICALTRIAL Not applicable. This study did not involve a clinical trial or prospective intervention with human participants.