Model Monitoring After Clinical Deployment
Holding
Clinical AI monitoring must prove that the deployed system remains within its intended use and performance assumptions. Uptime monitoring is not enough.
Authority
FDA quality expectations, AI risk management practices, and medical device lifecycle principles all require continuing control after release. For AI systems, the controlled object is not only code. It is code, model, data pipeline, user workflow, and clinical environment.
Issue
A model may deteriorate because a scanner changes, laboratory assays shift, patient mix changes, staff behavior evolves, or an EHR field is repurposed. None of these events necessarily produces a software error, but each may affect clinical safety.
Resolution
The monitoring plan should track input distribution, missingness, population segments, prediction confidence, override rate, downstream outcome proxies, latency, and user behavior. Thresholds should trigger review, not automatic retraining, unless that retraining is itself controlled.
Evidence Package
The record should include monitoring specifications, baseline distributions, alert thresholds, review minutes, corrective actions, model suspension criteria, and change-control links. A model that cannot be monitored should not be treated as operational clinical infrastructure.