What is AI model drift in healthcare?

AI model drift in healthcare is a change in how an AI tool performs after deployment, usually a decline, caused by shifts in the data it sees, the patients it serves, clinical practice, or the tool itself. A model that was accurate at validation can become less accurate, less fair, or miscalibrated without any obvious failure.

Common types of drift

• Data drift. The inputs change. A new lab assay, a coding update, a new EHR template, or a different documentation habit can shift what the model receives.
• Population drift. The patients change, for example after a hospital opens a new site, adds a service line, or sees a new disease wave.
• Concept drift. The relationship between inputs and outcomes changes, as when a new treatment alters who develops a complication.
• Update-driven change. The vendor retrains or updates the model, or an underlying large language model changes version.

For AI agents, drift can also be behavioral: the same agent, with a new model version or new tools, may take different actions than it did during testing.

Why it matters

Joint Commission and CHAI guidance on the Responsible Use of AI in Healthcare explains the need for ongoing monitoring directly: "AI algorithms may have the capacity to learn and adapt over time, data inputs can change or drift over time, and AI tools and their underlying algorithms may be updated periodically. This means AI tool outcomes and performance can change." It also lists "major performance degradation after an update" as an AI safety event hospitals should capture.

The NIST AI Risk Management Framework calls for AI system functionality and behavior to be "monitored when in production" (MEASURE 2.4) and for post-deployment monitoring plans that include incident response, recovery, and change management (MANAGE 4.1).

On the device side, FDA's final guidance on Predetermined Change Control Plans lets manufacturers of AI-enabled devices describe planned modifications, and how they will be validated, in advance.

Federal survey data show monitoring is common but not universal. ASTP/ONC Data Brief No. 80 reports that in 2024, 79% of hospitals using predictive AI conducted post-implementation evaluation or monitoring.

How hospitals monitor for drift

The RUAIH guidance recommends risk-based monitoring: tools that inform or drive clinical decisions should be checked more often than administrative tools. Common practices include tracking input distributions, comparing outputs against known outcomes, checking performance across patient subgroups, setting alert thresholds, and requiring a review after every vendor update. Monitoring only works if someone owns each tool and has the authority to restrict or pause it when a threshold is crossed. Baselines matter as well: recording performance at go-live, broken out by relevant patient subgroups, gives later checks something concrete to compare against.

How Skovos handles this

Skovos focuses on the response side for AI agents. Its audit trail shows what each agent did over time, which helps reviewers spot behavioral change, and the hospital can restrict permissions or recall an agent while it is investigated.

Frequently asked questions

What is the difference between data drift and concept drift?

Data drift is a change in the inputs. Concept drift is a change in the relationship between inputs and the outcome being predicted. Either can degrade performance.

How often should a hospital check a model for drift?

It depends on risk. The Joint Commission and CHAI guidance suggests more frequent checks for tools close to clinical decisions and periodic checks for administrative tools, plus a check after any update.

Can a vendor handle drift monitoring for us?

Vendors can provide dashboards, but the guidance stresses that local monitoring is critical because tools are often developed and validated elsewhere. Monitoring responsibilities should be set in the contract.

Related reading

• AI model drift detection in healthcare
• AI safety monitoring for health systems
• How to stop an AI agent in a hospital

Sources

• Joint Commission and CHAI, Guidance on the Responsible Use of AI in Healthcare: https://digitalassets.jointcommission.org/api/public/content/dcfcf4f1a0cc45cdb526b3cb034c68c2
• NIST, AI Risk Management Framework (AI RMF 1.0): https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf
• FDA, Marketing Submission Recommendations for a Predetermined Change Control Plan for AI-Enabled Device Software Functions: https://www.fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence
• ASTP/ONC Data Brief No. 80, September 2025: https://healthit.gov/data/data-briefs/hospital-trends-use-evaluation-and-governance-predictive-ai-2023-2024/