medRxivJClub
PharmaAgent: a real-world evaluation of high- and low-autonomy agentic systems for inpatient vancomycin dosing
medRxiv · · Preprint · Open access
Jiang, Black + 12 more
Abstract ↗AI summary
The abstract is read at the publisher; the summary is JClub's.
High-autonomy PharmaAgent vancomycin dosing at Stanford Hospital reduced mean absolute percentage error to 21.8% versus 37.8% for low autonomy across 403 events.
- Why it matters: Vancomycin’s narrow therapeutic window requires individualized dosing and repeated monitoring, and health systems need consistent support beyond manual physician-to-pharmacist workflows.
- What they did: The LLM-based agentic system was silently deployed for two months at Stanford Hospital, comparing high vs low autonomy with and without institutional guidelines over 403 dosing events in 316 patients.
- The result: Recommendations were within 25% of the pharmacist order in 73.4% vs 50.9% of events (P < 0.001), but pharmacist review of 196 out-of-tolerance recommendations found safety concerns in 65.8%, more often with low autonomy (odds ratio 2.37).
The findingWhy it mattersWhat they didThe result
- Open access