Continual ECG Deployment Improves Expert Retention, Autonomous Inference Lags

This research addresses the challenge of continual learning in multi-source Electrocardiogram (ECG) deployment, particularly when raw ECG data from earlier sources cannot be retained or replayed. The proposed method, referred to as `ours`, constructs an incremental bank of experts. It freezes a pre-trained backbone and assigns a balanced-softmax linear expert to each new data source. A key component is a lightweight router, trained only on retained training features and domain labels from previously observed sources. Instead of committing to a single routed expert, the system employs a validation-calibrated margin rule to fuse the two most likely experts. This approach demonstrates strong source-aware expert retention, achieving performance comparable to an offline independent-head reference. However, the study highlights that autonomous source inference—identifying the correct source without explicit metadata—remains the primary bottleneck. While the method avoids replaying raw ECGs, it does retain frozen training features for router updates, meaning it is not entirely memory-free.