An arXiv preprint reports a 2.70% relative average gain in Macro-F1 for BeatRhythm-TTA over the best competing method across three adaptation protocols. Macro-F1 is a score that gives each diagnosis class equal weight, rather than letting classes with more records dominate.
The work tackles a specific problem in ECG machine learning: performance can degrade when acquisition devices and patient populations shift from one domain to another. Its research question is whether test-time adaptation can mitigate that drop, allowing an ECG classifier to adjust when it encounters a new target domain.
How the adaptation works
BeatRhythm-TTA combines SQI-gated teacher-driven pseudo-labeling with two consistency checks between clean and augmented views. SQI refers to signal-quality information; the checks operate at both the rhythm and beat levels.
The benchmark used PTB-XL as the source domain, with 21,837 records, and CPSC2018 and Georgia as target domains, with 6,877 and 10,344 records respectively. The records were mapped to a shared six-class multi-label diagnosis space after unshared labels and empty samples were discarded.
Researchers scored the adapted models on target test sets with Macro-F1 and Macro-AUC, another class-balanced measure built from per-class area-under-the-curve values. The reported figures were averaged over three runs. The evaluation covered offline TTA, continual online TTA and independent online TTA.
Reported scores across protocols
In the offline setting, the reported averaged Macro-F1 figures ran from 60.64% to 62.42%, with the higher value reported for BeatRhythm-TTA. The corresponding Macro-AUC figures ran from 88.13% to 88.84%.
Under continual online TTA, the method's reported average was 61.74% for Macro-F1 and 88.77% for Macro-AUC. The results text reported improvements of 1.12% for Macro-F1 and 0.65% for Macro-AUC.
Under independent online TTA, the method's reported averages were 62.63% for Macro-F1 and 88.45% for Macro-AUC. The results text reported a 1.97% Macro-F1 improvement in that setting.
What the results leave unanswered
An offline ablation, or component-by-component test of the method, reported lower performance for configurations without SQI gating. The configuration without consistency regularization was reported with lower Macro-F1 and Macro-AUC, while the beat-level consistency component was associated with a clear additional gain.
The authors conclude that reliability-aware updates and semantically grounded beat-rhythm constraints are critical for stable ECG adaptation under real-world shifts. That interpretation remains tied to the benchmark: the study used the three named datasets and a shared six-class label space, and it reports three-run averages without confidence intervals or other uncertainty estimates.
The offline design also adapts on the target test set and evaluates on that same set, so the reported offline scores do not test post-adaptation performance on a separate target sample. More broadly, the evidence is limited to the benchmark domains and the six-class label space used here; it does not establish how the method would perform on other datasets, label spaces or settings.
The preprint's abstract states that code is available at https://github.com/Vincent-Wenhan/BeatRhythm-TTA. Whether the reported gains hold beyond the tested domains and shared six-class label space remains open.
Paper data and sources
Original title: Test-Time Adaptation for ECG Classification via SQI-Gated Self-Training and Beat-Rhythm Consistency
Authors: Wenhan Jiang, Zhipeng Deng, Jiale Zhou et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-24
DOI: Not available
Original paper · Full text