Who Notices When a Clinical Model Drifts? (Life Science Daily News)
Life Science Daily News published my article on clinical model drift. Most tools get one exam, then go live. The quiet failures worry me more than the loud ones.


On October 21, 2026, Life Science Daily News published my article, “Who Notices When a Clinical Model Drifts?”
Most clinical prediction tools get one big exam. Someone shows a committee a curve, and the tool goes live. After that the exam rarely gets repeated, so model drift can go unnoticed.
The emergency department keeps changing anyway. A triage protocol gets rewritten. The lab switches assays. Flu season fills the hallway with stretchers, and the electronic health record gets an upgrade over a long weekend. The model does not know any of that happened. It keeps scoring patients the way it learned to score them.
The best-known example came early in the COVID-19 pandemic. The University of Michigan had deployed a widely used proprietary sepsis-alerting model, and in April 2020 the hospital’s clinical AI governance committee switched it off. The patients coming through the doors had changed enough to alter the relationship between fever and bacterial sepsis, and the alerts turned spurious.
A later analysis across 24 hospitals in four health systems put numbers on it. In the three weeks after each system’s first COVID-19 case, daily sepsis alerts rose 43 percent while total hospital census fell 35 percent. The share of patients generating an alert went from 9 percent to 21 percent. The authors were careful: part of the jump reflected real change in who was in the building. Even accurate alerts can swamp a unit with no spare hands. Nurses reported the overalerting within weeks. The quiet versions worry me more.
Researchers at Vanderbilt built seven models for hospital-acquired acute kidney injury, then followed them across nine later years. Discrimination held up for every model. Calibration did not. The models increasingly overpredicted risk. The area-under-the-curve slide can still look fine while the probability attached to each patient has quietly changed shape.
There is a stranger version of the problem. After a tool works and outcomes improve, the model can look less accurate on paper, because the flagged patients now have the bad outcome less often. A governance group that reads every dip as failure could end up retiring the tools that helped most.
Michigan could switch its sepsis model off within weeks because nurses spoke up and a committee had the authority to act. In the full article I offer questions for the next governance meeting, including who owns post-go-live performance by name, and whether you are tracking calibration at the thresholds you actually use. Ask whether your department could do the same for each model it runs today, including the ones nobody has complained about yet.
This post is educational. It is not legal, compliance, or clinical advice. Your own governance group decides.
Photo: ThisIsEngineering on Pexels
