Medical Technology · eu
When Should Medical AI Be Updated? AEGIS Tests a Post-Market Monitoring Framework with Sepsis and Brain Tumor Models
A new version failing to qualify does not mean the old version can safely remain in use. A Karolinska research team separates update reviews from risk alerts in an effort to turn continuous oversight of medical AI into an actionable process; the results remain limited to a proof of concept.
An artificial intelligence system that helps identify sepsis or delineate brain tumors on magnetic resonance imaging may gradually lose accuracy as patient populations, equipment, and clinical workflows change. Hospitals then face more than the question of whether to adopt a new version. There is another, more pressing question: does the version currently in use already require intervention?
The AEGIS framework proposed by a team at Sweden’s Karolinska Institutet attempts to establish a clear decision pathway for each of these questions. The study was published in *npj Digital Medicine* on October 13, and the institute presented the findings on October 22; the same team had already released a preprint in March. The study focuses on demonstrating how to carry out post-market monitoring and update decisions for medical AI. It has not yet shown that this process can improve patient outcomes.
AEGIS divides the work into three parts: integrating new data and retraining, monitoring model performance, and deciding the next step according to predefined conditions. Drawing on the US Food and Drug Administration’s Predetermined Change Control Plan and European Union regulations related to medical devices and AI, the study attempts to translate oversight requirements into traceable operating procedures. However, the US and European systems are not identical, and adopting the framework does not amount to obtaining regulatory authorization.
Within the framework, candidate models are assessed in order of priority and assigned one of four decisions: deployment permitted, deployment conditionally permitted, referral for clinical review, or deployment rejected. These are all internal governance decisions. A separate ALARM alert monitors deployed models. The system can therefore reject a new version while also issuing a warning about the old version, preventing a blocked update from being mistaken for evidence that the current situation remains safe.
The sepsis demonstration used historical intensive care unit data from PhysioNet Challenge 2019 and a random forest classifier in 11 rounds of testing, progressively introducing data corruption scenarios. The early preprint recorded eight decisions permitting deployment and one decision in each of the other three categories. It also reported a scenario in which the candidate model could not be deployed while the current model triggered an alert. A separate brain tumor demonstration used BraTS 2023/24 magnetic resonance imaging data and the nnU-Net ResENC-M model in 12 rounds of testing. According to the team’s documentation, this demonstration disabled the two intermediate decision categories, testing only whether to permit or reject deployment.
The published paper also tested the ability to detect distribution shifts using data from different hospitals, reporting an increase in ROC-AUC from 0.82 to 1.00, with detection triggered at a data batch size of around 500 patients. No false alarms occurred in these experiments. These figures measure the ability to identify changes in data. They cannot be interpreted as the clinical accuracy of sepsis diagnosis or brain tumor segmentation, nor do they guarantee that other hospitals will obtain the same results.
At present, AEGIS’s drift scoring primarily captures changes in the distribution of input data. It does not yet directly address concept drift, in which the relationship between the same inputs and clinical outcomes has changed. The team has released an implementation guide, documentation templates, and demonstration code, but clinical validation and regulatory assessment for each intended use are still required before the framework is put into patient care. The next step—field testing with clinical partners and regulators—may help answer whether the process can identify problems promptly and support appropriate action.