In medicine, predictive models derived from artificial intelligence (AI) and machine learning have some related weaknesses that call for monitoring and correction. Model performance can degrade over time as clinical practice, information systems and patient populations change — a problem known as model drift.
Another potential weakness is poor generalizability across different clinical, geographic or sociodemographic patient groups. Left undetected, these problems threaten patient outcomes and access to care.
With the aid of a new, four-year $1.6 million research grant from the National Library of Medicine, part of the National Institutes of Health (NIH), Sharon Davis, PhD, Assistant Professor of Biomedical Informatics, is studying solutions to address these related weaknesses.
Davis and colleagues previously developed data-driven methods that monitor deployed models, detect drift and recommend updates. The approach has helped maintain a national Department of Veterans Affairs (VA) risk model for acute kidney injury. But like most model maintenance methods, it tracks performance across the patient population as a whole.
The new project asks how gaps in accuracy between patient groups shift over time, and how to catch and correct them. Using VA and Vanderbilt Health data, the team will study prediction models for complications after surgery, hospital-acquired acute kidney injury, and postpartum hemorrhage.
- Track how performance gaps between patient groups change over time, in traditional machine learning models and in large language models used for prediction.
- Develop monitoring and updating methods that restore accuracy both overall and within subgroups.
- Use synthetic patient data to support monitoring for small groups, such as patients at rural hospitals.
- Account for the way an effective AI tool changes the very outcomes it predicts, a shift that can be mistaken for drift.
That last line of work will be tested on a postpartum hemorrhage prediction tool now in use at Vanderbilt Health.
“A model that performs well on the day it’s deployed won’t necessarily keep performing well, or perform equally well for every group of patients,” Davis said. “We want to give health systems practical ways to keep these tools accurate and dependable for everyone they serve.”
The research is supported by NIH grant R01LM015342.
