A research team has identified exactly where AI-based ECG classifiers break down in real hospital deployments: not retention, but source identification.
The paper introduces an incremental expert bank built on top of frozen ECGFounder features. Each new data source — say, a hospital joining a network — gets its own linear classifier, which prevents earlier models from degrading as new ones are added. On four benchmark datasets (CPSC, PTB-XL, Georgia, and Chapman-Shaoxing), the system achieves a Macro-F1 score of 0.7915 when it knows which source a recording came from, nearly matching a fully offline reference model that scores 0.7885. The catch: real deployments rarely hand you that metadata.
When source labels are unavailable, the system drops to 0.7756 using an MLP router, recovering only slightly to 0.7782 with a top-2 fusion trick that blends the two most probable expert predictions. That 0.0026 gap between hard routing and fusion is statistically indistinguishable from noise. The authors call autonomous source inference — figuring out where a recording came from without being told — the main remaining bottleneck, and the numbers back that up.
Continual learning for medical AI is a crowded research space, but this paper does something useful: it cleanly separates two problems that often get conflated. Keeping old models intact is largely solved here; knowing which model to invoke is not. Worth noting — the method still retains frozen training features for router updates, so anyone calling this memory-free should read the fine print.