Pharmacy Pilots VIM and MEDLOG for Safer, Traceable AI

• NEJM AI proposes a 'Values in the Model' (VIM) label that summarizes an AI’s decision tendencies (for example, cost‑saving versus maximal benefit) and a complementary 'MEDLOG' audit log to record AI actions. The paper frames governance and calls for pilots, not immediate regulation.

• Due diligence and electronic health record (EHR) integration should seek VIM value profiles for order verification, clinical decision support (CDS) rules, automated dispensing cabinets (ADCs), and inventory forecasting. Pharmacy & Therapeutics (P&T), informatics pharmacists, and contracting can embed explicit tradeoff criteria in requests for proposals.

• A VIM snippet in the pharmacist verification queue clarifies an AI’s bias (e.g., cost‑aware vs interventionist), helping interpret dosing or antibiotic recommendations and reducing rounds‑based back‑and‑forth. P&T should define which products or situations require checking VIM (for example, high‑alert meds or restricted antimicrobials).

• Operationalizing MEDLOG requires mapping logs to EHR order IDs, sterile‑compounding batch numbers, and automated dispensing cabinet (ADC) dispenses so audits trace from recommendation to dispense. Pharmacy informatics should own validation (parallel runs, reconciliations), set drift monitors and rollback triggers, and manage PHI access and retention.

• When AI moves from informational clinical decision support to patient‑specific directives (for example, auto‑ordering), it may qualify as a medical device and change regulatory and liability obligations. Define who signs off on classification and require audit trails that record human overrides to maintain clinician accountability.

Assign VIM and MEDLOG to pilot governance: task Pharmacy Informatics and Pharmacy & Therapeutics to run a single‑service trial assessing verification efficiency and audit readiness.

• VIM is a one‑page, clinician‑facing profile that states an AI’s decision tendencies (e.g., intervention vs conservation, cost vs benefit). It lists developer‑disclosed value tuning, benchmarked results by axis, model version/date, and a short summary intended for CDS and verification screens.

• VIM relies on a vetted library of value‑sensitive clinical vignettes tested across diverse stakeholder panels. Results produce a machine‑readable JSON profile with alignment percentages by stakeholder and axis. Early evidence is preliminary (RAISE poll split 47/53; 14 large language models produced varied responses).

• MEDLOG is an audit trail that records each AI‑assisted clinical action with context: timestamps, model/version, inputs, prompts, recommendations, rationale snippets, user overrides, order IDs, and outcome links. Integrating MEDLOG with the EHR audit trail, automated dispensing cabinet (ADC) events, and incident dashboards enables root‑cause review.

• Pilots should be led locally by Pharmacy & Therapeutics (P&T) and pharmacy informatics with support from clinical leadership, quality, and compliance. Developers would supply VIM data; health systems should host MEDLOG, define retention and PHI access rules, and map logs into existing audit and governance workflows.