Clinical impact of AI-assisted medication reconciliation at admission: A prospective before-and-after study

Abstract

Background: Medication reconciliation (MedRec) at admission remains vulnerable to unintentional discrepancies that can precipitate preventable harm. Advances in natural language processing (NLP) offer opportunities to assemble more complete preadmission medication lists and direct clinician attention to high-risk mismatches. Objective: To evaluate the clinical and operational impact of an AI-assisted MedRec workflow embedded in the electronic health record (EHR) at hospital admission. Methods: We performed a prospective before-and-after study at a tertiary academic hospital, enrolling consecutive adult medical admissions across two 8-week epochs: baseline standard MedRec and post-deployment AI-assisted MedRec. The intervention ingested heterogeneous sources (prior notes, discharge/clinic summaries, pharmacy fills when available) and generated explainable flags for likely discrepancies between the best possible medication history and draft admission orders. The primary outcome was unintentional medication discrepancies per patient. Secondary outcomes included high-severity potential adverse drug events (pADEs) at admission, time-to-completion of MedRec, and a composite of 30-day ED revisit/readmission. Results: Among 840 admissions (420 per epoch), cohorts were demographically similar. Mean unintentional discrepancies per patient declined from 1.15 to 0.94 (incidence rate ratio 0.82, 95% CI 0.72-0.93). Patients with ≥1 high-severity pADE at admission decreased from 21.7% to 16.0% (odds ratio 0.69, 95% CI 0.51-0.92). Median time-to-completion shortened from 66.9 to 51.1 minutes (ratio 0.76), indicating ~24% faster completion. The 30-day use composite numerically declined from 13.5% to 11.8% but the study was not powered for this endpoint. Interrupted time-series analysis showed a level shift coincident with deployment. Conclusions: AI-assisted MedRec at admission reduced unintentional discrepancies and high-severity pADEs while improving workflow efficiency, without observed safety trade-offs. Pairing pharmacist expertise with targeted, explainable AI support offers a pragmatic path to safer and faster admissions and warrants evaluation in multi-site randomized designs.