Continual ECG Deployment: Expert Retention vs Source Inference

A paper by Yufan Lu, Xinhui Liu, Chenyang Xu, Yuxi Zhou, Hao Wang and Shenda Hong, submitted 2 Jul 2026 to BIBM2026, proposes a continual ECG deployment method that keeps pretrained backbone features but adds per-source linear experts and a lightweight router. The approach builds an incremental expert bank on frozen 1024-dimensional ECGFounder features and avoids replaying raw ECG signals.

How does the method work?

The method freezes a 1024-dimensional ECGFounder backbone and adds a balanced-softmax linear expert for each arriving data source, while fitting a lightweight router on retained training features and domain labels. The system retains only frozen training features for router updates, uses a validation-calibrated margin rule to fuse the two most likely experts rather than committing to a single routed expert, and explicitly does not replay raw ECGs.

Implementation details in the abstract note three core pieces: a frozen ECGFounder feature extractor (1024 dimensions), an incremental expert bank of balanced-softmax linear classifiers, and a router trained on retained features and labels. The paper labels the setup Raw-ECG-Replay-Free because raw ECGs are not replayed, but it also states the approach is not memory-free because frozen training features are retained for router updates.

How well did it perform?

On four multi-source ECG collections, source-aware expert selection reached a Macro-F1 of 0.7915 ±0.0036, while a matched offline independent-head reference reached 0.7885 ±0.0009, showing strong source-aware expert retention. Without source IDs, an MLP router reached 0.7756 ±0.0027 Macro-F1 and top-2 margin fusion reached 0.7782 ±0.0022.

The paper evaluates on CPSC, PTB-XL, Georgia, and Chapman-Shaoxing datasets and reports that the top-2 fusion gain over hard MLP routing is +0.0026, with a paired bootstrap 95% confidence interval that includes zero. Across three different domain orders the top-2-to-oracle gap remains between 0.0111 and 0.0133, which the authors identify as evidence that autonomous source inference is the main remaining bottleneck once experts are retained.

Why it matters

The results distinguish two problems that are often conflated in continual multi-source medical deployments: retaining source-specific expertise and inferring the source when metadata are unavailable. The experiments show expert retention can match an offline independent-head reference on Macro-F1 (0.7915 vs 0.7885), while autonomous source inference — routing without source IDs — still lags (MLP router 0.7756). That separation guides where engineering effort should go: keep experts stable, and focus research on more reliable source inference or router strategies.

What to watch

Watch for follow-ups that reduce the 0.0111–0.0133 top-2-to-oracle gap or that eliminate the need to retain frozen training features, since the paper notes the method is not memory-free. Also look for comparisons that test other router architectures or different calibration rules for fusing experts on the same CPSC, PTB-XL, Georgia and Chapman-Shaoxing benchmarks.

Details worth carrying forward: the system uses balanced-softmax linear experts, preserves no raw ECG replay, and relies on a validation-calibrated margin rule to fuse top-2 experts. The authors submit the work as "Separating Expert Retention from Autonomous Source Inference in Raw-ECG-Replay-Free Continual ECG Deployment" (arXiv:2607.01674).