A Clinician-Supervised Pipeline for Detecting and Communicating Drug-Drug Interactions in Orthopedic Surgery...

Drug-drug interactions (DDIs) are a major cause of preventable harm at hospital discharge. We developed a hybrid workflow combining automated extraction of DDIs from MIMIC-IV orthopedic surgery discharge prescriptions, semantic filtering, clinician oversight, and large language model (LLM) generation of patient-friendly notices.

From 18,275 discharge notes, 17,636 medication lists were extracted and normalized using RxNorm (T121). Strong DDI signals were identified using conservative TwoSIDES thresholds and ranked with a composite score. MedDRA codes were filtered by semantic types T184, T048, and T033, then mapped to English/French labels using UMLS and a bilingual dictionary.

A total of 121 unique interactions were retained. Mistral generated French patient-friendly explanations, which clinicians reviewed and selected for inclusion in notices. Nine of the top 20 interactions were opioid-related, including confusional state, suicidality, and misuse.

This workflow supports clinically aligned DDI detection and patient communication while preserving clinician control.