Test-Time Adaptation for ECG Classification via SQI-Gated Self-Training and Beat-Rhythm Consistency

Test-Time Adaptation for ECG Classification via SQI-Gated Self-Training and Beat-Rhythm Consistency

arXiv:2608.23347v1 [cs.LG] 24 Aug 2026

Wenhan Jiang1 , Zhipeng Deng1 , Jiale Zhou1 , Haolin Wang2 , Yafei Ou3 , and Yefeng Zheng1 Westlake University, Hangzhou, China Hokkaido University, Sapporo, Japan 3 RIKEN, Japan {jiangwenhan, dengzhipeng, zhoujiale, zhengyefeng}@westlake.edu.cn; yafei.ou@riken.jp 1

2

Abstract. Deep learning models for electrocardiogram (ECG) classification often suffer from significant performance degradation when deployed in unseen domains due to shifts in acquisition devices and patient populations. Test-time adaptation (TTA) offers a practical solution by adapting models using only unlabeled data at inference time. However, existing TTA methods often underperform on ECG tasks, since naive online updates ignore the hierarchical beat-rhythm structure of cardiac cycles and are vulnerable to signal artifacts, which leads to unstable adaptation and model drift. We propose BeatRhythm-TTA, an ECG-tailored TTA framework that explicitly accounts for ECG’s noisy observations and structured beat-rhythm semantics under domain shift. First, to handle pervasive ECG artifacts, we introduce a Signal Quality Index (SQI)gated adaptation scheme that selectively filters out low-quality signals to prevent harmful updates. Second, to leverage ECG’s beat-rhythm semantics, we enforce dual-level consistency so the model preserves beat morphology and rhythm dynamics while adapting to shifted acquisition conditions. Extensive experiments on multi-label ECG diagnosis across three adaptation protocols, using PTB-XL as the source domain and CPSC2018/Georgia as two target domains, demonstrate the effectiveness of our method, yielding an average +2.70% relative improvement in Macro-F1 over the best competing method. The code is available at https://github.com/Vincent-Wenhan/BeatRhythm-TTA. Keywords: Test-Time Adaptation · ECG Classification · Domain Shift

1

Introduction

Although deep learning has achieved strong performance on electrocardiogram (ECG) interpretation [15,21], generalizing pretrained models to unseen clinical environments remains challenging due to distribution shifts introduced by heterogeneous hardware and patient cohorts [7,18]. Re-training or fine-tuning for

2

W. Jiang et al. Training Stage

Source Only

Testing Stage

Source-Trained Model

ECGs with Labels

Test ECG from Unseen Domains

General TTA Source-Trained Model

z

Source-Trained Model

AF

Non-AF

BeatRhythm-TTA Source-Trained Model

Train

Non-AF

AF

Beat-Rhythm Semantics

Fig. 1. Motivation of the proposed BeatRhythm-TTA. Comparison of TTA strategies. Generic TTA methods are brittle on ECG, while BeatRhythm-TTA leverages beat-rhythm semantics for robust adaptation. (AF denotes atrial fibrillation, a cardiac rhythm disorder.)

each new target domain is often impractical in clinical practice as it requires substantial expert annotation and repeated engineering effort [18]. This motivates adopting test-time adaptation (TTA) from the broader machine learning literature [25,26], a paradigm that adapts a deployed model using only unlabeled test data, without requiring target labels or retraining from scratch. Most TTA methods have been developed and validated primarily on vision benchmarks [25,26]. These methods are usually based on normalization calibration (e.g., AdaBN [13], MemBN [9]), entropy minimization (e.g., TENT [25], SAR [17], DeYO [12]), or confidence- or energy-based objectives (e.g., COME [30], TEA [28]). Recent work has extended TTA beyond vision, showing promising performance in speech, audio, and AI-for-science under distribution shifts [10,11,29,31]. A closely related paradigm is source-free domain adaptation (SFDA) [8,14,27], which similarly addresses domain shifts without source data but is typically formulated as an offline adaptation stage before deployment. Despite progress in other modalities, ECG-specific TTA remains underexplored, as directly transferring existing methods raises modality-specific challenges. First, adapting models on continuous ECG streams corrupted by transient artifacts (e.g., baseline wander, motion artifacts) amplifies spurious correlations and triggers drift. Second, many existing TTA objectives (e.g., global entropy minimization) ignore the hierarchical beat-rhythm semantics that underlie ECG diagnosis [2,23]. To this end, we propose BeatRhythm-TTA, a semantically grounded TTA framework comprising two complementary components. First, SQI-gated adaptation selectively scales or blocks test-time updates based on a signal quality index (SQI) to avoid harmful adaptation on noisy segments. Second, dual-level consistency enforces beat- and rhythm-level agreement to preserve physiologically motivated semantics at two temporal scales [6]: local beat morphology (e.g., ST-segment changes [23]) and global rhythm irregularities (e.g., atrial fibrillation [2]). By coupling reliability-aware gating with physiologically grounded consistency, BeatRhythm-TTA reduces drift under real-world domain shifts. Ex-

BeatRhythm-TTA

3

Clean View Teacher Model

PseudoLabels

EMA Beat View Beat Beat Segmentation Beat Segmentation Segmentation R T P Q

Beat SQI

SQI Calculation

Beat SQI Beat SQI

Sample SQI

Confidence Filter

Dual-Level Consistency

Consistency Loss

Pseudo-Label Loss

S

BN Others

Augmented View Augmented View

Student Model

SQI Calculation

Dual-Level Consistency Rhythm Consistency

Concentration

Beat View

Global Consistency Loss

Clean Feature Sharpness

Beat SQI

Pull

Augmented Feature

Wander

Beat Consistency

Pull

Pull

Beat SQI

Local Consistency Loss

Raw ECG Input

Fig. 2. BeatRhythm-TTA Framework. Given an unlabeled target ECG, adaptation is driven by two complementary modules. (1) SQI-gated pseudo-labeling: Blocks self-training updates based on teacher confidence and beat-wise signal quality, where EMA (exponential moving average) maintains a slowly updated teacher for more stable pseudo-labels. (2) Dual-level consistency: Enforces rhythm- and beat-level agreement between clean and augmented views to stabilize adaptation.

tensive experiments on multi-label ECG classification across three adaptation protocols, using PTB-XL as the source domain and CPSC2018/Georgia as two target domains, validated the effectiveness of our approach. It outperforms the strongest competing methods with an average +2.70% relative improvement in Macro-F1 across protocols.

2

Method

2.1

Preliminaries and Overview

Problem definition. Let f (·; θ) be an ECG classifier that maps a C-lead segment x ∈ RC×T to logits z ∈ RK , with multi-label targets y ∈ {0, 1}K , where K denotes the number of diseases to classify. We first train θS on the labeled S source domain DS = {(xSi , yiS )}N i=1 :
θS = arg min E(xS ,yS )∼DS Lsup f (xS ; θ), y S . (1) θ

T

At deployment, unlabeled target samples {xTi }N i=1 arrive from an unseen domain DT . Test-time adaptation (TTA) updates the model using only xT by minimizing

4

W. Jiang et al.

a self-supervised loss:
θT = arg min ExT ∼DT Ltta f (xT ; θ), xT . θ

(2)

Offline and online TTA settings. We consider both offline and online testtime adaptation settings. Offline TTA assumes the entire target test distribution T {xTi }N i=1 is available for adaptation and we optimize Ltta on the unlabeled target test set and then evaluate on the same set. Online TTA processes test samples sequentially: as each sample arrives, the model is updated on it and predicts on-the-fly. We consider three protocols: (i) offline, (ii) continual online, and (iii) independent online TTA. Continual online carries updates across samples (θi → θi+1 ), while independent online resets to θS for each sample. Method Overview. The overview of our method is shown in Fig. 2. In real ECG recordings, signal quality can vary due to artifacts, and diagnostic semantics appear at both the beat and rhythm levels. Motivated by these properties, we propose BeatRhythm-TTA, an ECG-tailored framework with two complementary modules: an SQI-gated pseudo-labeling module that computes a beat-wise SQI to gate teacher-driven self-training, and a dual-level consistency module that enforces rhythm- and beat-level agreement across clean and augmented views. 2.2

SQI-gated Pseudo-labeling

To prevent harmful updates triggered by noisy samples and segments, we perform selective self-training at test time, only updating the model when the current ECG sample is both confident under a teacher model and reliable according to a signal quality index (SQI). Signal quality indices have been widely used to quantify waveform reliability in ECG analysis [3,20]. Unlike prior SQIs that mainly target record-level screening and may require heavier, domain-tuned features, our SQI is a lightweight beat-wise score designed for test-time gating, computed online without labels to control update eligibility and strength. b Given an unlabeled ECG sample x ∈ RC×T , we detect R-peaks {ri }M i=1 using the WFDB toolchain [5], where Mb denotes the number of detected beats in the sample. We select a reference lead based on derivative energy and apply robust normalization with basic physiological sanity checks. We then extract beat-centered windows and scale the window length by the median RR interval to account for heart-rate variation. For each beat window x(i) , we compute three lightweight morphology factors: (i) QRS concentration (conc(i) ): the QRS complex, the sharp waveform around the R-peak in the P-QRS-T beat structure, carries concentrated energy in clean beats [24]; we measure it as the energy ratio of a small center region to the whole beat window; (ii) sharpness (sharp(i) ): true QRS has steep slopes, measured by a highquantile magnitude of the first-order difference normalized by signal root mean square; (iii) baseline wander ratio (bwr(i) ): strong low-frequency drift, measured as the energy ratio of a moving-average baseline estimate to the whole window, indicates poor motion.

BeatRhythm-TTA

5

We combine them to obtain a beat SQI score SQI(i) : SQI(i) = σ(conc(i) ) · σ(sharp(i) ) · σ(−bwr(i) ),

(3)

where σ(·) is the sigmoid function. Finally, we aggregate beat-wise scores into a P sample-level quality weight w = M1b i SQI(i) (mean over beats). We adopt a teacher-student scheme to generate pseudo labels from a clean view and supervise the student on an augmented view as:   Lpl = I[c(x) ≥ τc ] I[w(x) ≥ τq ] ℓbce pθ (x̃), pθ̄ (xc ) , (4) where xc = x is the clean view, x̃ = A(x) is a time-aligned augmentation, and ℓbce denotes binary cross-entropy (BCE) loss. We denote pθ (x) = σ(fθ (x)) as multi-label probabilities, and use a teacher fθ̄ to produce soft pseudo-labels pθ̄ (xc ). We define the confidence score as c(x) = mink pθ̄ (xc )k − 12 , and I[·] denotes the indicator function, so activate pseudo-label updates only when both the confidence and the SQI thresholds are satisfied. The teacher is updated by exponential moving average (EMA) [22,26]: θ̄ ← µθ̄ + (1 − µ)θ. This EMA update stabilizes pseudo-labels and mitigates rapid forgetting during online adaptation. 2.3

Dual-level Consistency

ECG semantics appear at two temporal scales: beat-level morphology and rhythmlevel dynamics across beats [6]. Diagnoses may rely on subtle beat morphology such as ST-segment elevation/depression and T-wave inversion [23] or rhythm irregularity across multiple beats such as atrial fibrillation (AF) [2]. Accordingly, to stabilize adaptation while preserving clinically meaningful morphology, we regularize test-time updates with a dual-level consistency objective that aligns global sample representations and local beat-centric representations between a clean view xc and a time-aligned augmented view x̃. Let F (·) denote intermediate features extracted from a chosen backbone layer. Given the two time-aligned views, we obtain feature of the clean view (Fc = F (xc )), the augmented view (Fa = F (x̃)), and minimize a cosine alignment loss: Lrhythm = 1 − cos(sg(Fc ), Fa ) , (5) where sg(·) denotes stop-gradient operator. Rhythm alignment captures global dynamics but does not explicitly constrain beat morphology. Using the previously detected R-peaks {ri } and the RR-based windowing, we extract beat-centric embeddings from Fc and Fa , yielding paired (i) (i) Mb b sets {bc }M i=1 and {ba }i=1 . We align beat embeddings using cosine distance:   (i) (i) ℓbeat = 1 − cos sg(b(i) ), b , (6) c a and aggregate across beats using the beat-wise SQI weights from Eq. (3), emphasizing reliable beats and suppressing corrupted ones during adaptation: P (i) (i) ω ℓ Lbeat = Pi (i) beat , ω (i) = max(SQI(i) , τfloor ), (7) +ϵ i ω

6

W. Jiang et al.

where τfloor = 0.05 sets a minimum beat weight. We combine the dual-level consistency losses with the SQI-gated pseudo-label loss: L = Lpl + λbeat Lbeat + λrhythm Lrhythm . (8) λbeat and λrhythm are scalar hyperparameters that balance the contributions of beat-level and rhythm-level consistency against the pseudo-labeling loss. By optimizing this objective, the model is updated via SQI-gated self-training, while dual-level consistency regularizes test-time updates to preserve both beat- and rhythm-level semantics.

3

Experiments

3.1

Datasets and Experiment Setup

We evaluated our method on three public 12-lead ECG datasets from the PhysioNet/CinC Challenge 2020 [19]. Specifically, we used PTB-XL (Germany, 21,837 records) as the source domain, and evaluated cross-domain adaptation on CPSC2018 (China, 6,877 records) and Georgia (USA, 10,344 records) as two distinct target domains. All datasets were projected into a shared 6-class label space (atrial fibrillation, first-degree atrioventricular block, left bundle branch block, right bundle branch block, premature atrial contraction, and normal sinus rhythm), discarding unshared labels and empty samples. Evaluation metrics. Macro-F1 (the unweighted mean of per-class F1 scores) and Macro-AUC (the unweighted mean of per-class AUROC) were reported on the target test sets after adaptation. Macro averaging gives equal weight to each class, which is important under class imbalance in multi-label ECG diagnosis. All the results were averaged over 3 runs. Implementation details. Experiments were implemented in PyTorch on an RTX 4090 GPU. For preprocessing, all ECGs were resampled to 500 Hz and padded to a fixed length of 10 s. Following [1,4,16], a 1D ResNet-18 was pretrained on PTB-XL (8:2 train/val split) for 100 epochs using the BCE loss, standard augmentations, and the Adam optimizer (lr = 10−4 , batch size 512). The best validation checkpoint was selected as the source model to initialize TTA. Offline TTA adapted on the target set for 1 epoch (batch size 256). Continual and independent online TTA processed single samples (batch size 1) for 5 and 10 internal steps using the Adam optimizer with lr = 10−6 and 10−4 , respectively. For our method, we disabled dropout and updated only BatchNorm affine parameters. Thresholds in Eq. (4) were set to τc = 0.2 and τq = 0.05. The loss weights in Eq. (8) were set to λbeat = 0.5 and λrhythm = 1.0 for all experiments. 3.2

Results and Analysis

We compared our method against the source-only baseline and representative TTA methods (AdaBN [13], MemBN [9], TENT [25], SAR [17], DeYO [12], COME [30] and TEA [28]). For the offline setting, we additionally evaluated

BeatRhythm-TTA

7

Table 1. Offline TTA results. Bold: best, underlined: second best. All values are reported in percent (%). Method

CPSC2018 AUC F1

Georgia AUC F1

Avg AUC F1

Source Only

87.62

60.84

88.59

60.35

88.11

60.59

AdaBN [13] TENT [25] COME [30] SAR [17] MemBN [9] TEA [28] DeYO [12]

83.25 87.57 87.28 83.43 83.51 87.47 87.55

46.71 60.63 60.14 47.63 47.58 60.36 60.62

87.33 88.57 88.43 87.48 87.58 88.62 88.57

52.29 60.29 59.83 53.40 54.32 60.28 60.30

85.29 88.07 87.86 85.45 85.55 88.04 88.06

49.50 60.46 59.99 50.52 50.95 60.32 60.46

SHOT [14] SF(DA)2 [8] UCon [27]

87.35 87.46 87.65

60.20 60.38 60.92

88.47 88.52 88.61

59.76 60.07 60.36

87.91 87.99 88.13

59.98 60.22 60.64

Ours

88.70

62.57

88.97

62.26

88.84

62.42

source-free domain adaptation baselines (SHOT [14], SF(DA)2 [8] and UCon [27]). Tables 1–2 show that our method consistently achieves the best average MacroAUC and Macro-F1 across offline and online protocols. Offline TTA. Under batch adaptation (Table 1), our method improves the averaged Macro-F1 from 60.64% to 62.42%, and Macro-AUC from 88.13% to 88.84%. This suggests that combining SQI-gated updates with dual-level consistency provides reliable optimization signals when the target set can be revisited. Continual online TTA. In the streaming setting without resets (Table 2), updates accumulate and the model can be influenced by occasional low-quality segments. Our method remains robust and yields the best averaged metrics, improving Macro-F1 by +1.12% and Macro-AUC by +0.65%. This highlights the benefit of reliability-aware SQI gating for preventing unstable updates in long test streams under continual online TTA setting. Independent online TTA. Also shown in Table 2, our method achieves the highest average performance in the independent setting, improving Macro-F1 by +1.97%. ECG samples vary widely across individuals in heart rate, morphology, and noise patterns. By combining SQI-gated updates with beat-rhythm consistency, our method adapts to each sample in a personalized and stable manner, yielding the largest gains. Ablation Study. Table 3 validates the necessity of each proposed module under offline TTA. Removing SQI gating degrades performance, confirming that reliability-aware filtering prevents noisy segments from driving harmful adaptation. Removing the consistency regularization further reduces Macro-F1 and MacroAUC, showing that semantic-preserving constraints offer complementary benefits

8

W. Jiang et al.

Table 2. Online TTA results under continual and independent settings. Bold: best, underlined: second best within each setting. All values are reported in percent (%). Method

CPSC2018 AUC F1

Georgia AUC F1

Avg AUC F1

Source Only

87.62

88.59

88.11

60.59

87.03 86.16 87.92 88.12 87.91 88.06 87.96 88.77

60.37 54.92 60.48 60.62 60.55 60.51 60.39 61.74

87.85 87.82 87.93 88.13 85.36 88.07 87.91 88.45

59.06 60.12 60.33 60.66 45.95 60.45 60.18 62.63

60.84

60.35

Continual Online TTA AdaBN [13] TENT [25] COME [30] SAR [17] MemBN [9] TEA [28] DeYO [12] Ours

86.11 84.92 87.29 87.63 87.43 87.52 87.38 88.98

58.81 53.68 60.63 60.86 60.44 60.77 60.55 62.53

87.94 87.40 88.54 88.60 88.38 88.59 88.54 88.55

61.93 56.15 60.33 60.37 60.65 60.25 60.24 60.94

Independent Online TTA AdaBN [13] TENT [25] COME [30] SAR [17] MemBN [9] TEA [28] DeYO [12] Ours

87.25 87.20 87.35 87.64 85.17 87.55 87.35 88.27

59.37 60.37 60.47 60.90 45.42 60.68 60.39 62.97

88.44 88.43 88.50 88.61 85.55 88.59 88.47 88.63

58.75 59.87 60.19 60.41 46.47 60.23 59.96 62.30

Table 3. Ablation study under offline setting. All values are reported in percent (%). SQI

Consistency Rhythm Beat

✓ ✓ ✓

✓ ✓ ✓

✓ ✓

CPSC2018 AUC F1

Georgia AUC F1

Avg AUC F1

87.62 88.45 88.55 88.65 88.70

88.59 88.89 88.91 88.95 88.97

88.11 88.67 88.73 88.80 88.84

60.84 61.73 61.95 62.38 62.57

60.35 61.58 61.82 62.13 62.26

60.59 61.66 61.89 62.26 62.42

beyond gating. Finally, beat-level consistency yields a clear additional gain, supporting the role of beat-aware alignment in ECG adaptation.

4

Conclusion

We presented BeatRhythm-TTA, an ECG-tailored framework combining SQIgated pseudo-labeling with dual-level consistency regularization. Across two target domains, our method consistently outperformed representative TTA and SFDA baselines under offline, continual, and independent online protocols. The results

BeatRhythm-TTA

9

highlight that reliability-aware updates and semantically grounded beat-rhythm constraints are critical for stable ECG adaptation under real-world shifts.

References 1. Abbaspourazad, S., Elachqar, O., Miller, A., Emrani, S., Nallasamy, U., Shapiro, I.: Large-scale training of foundation models for wearable biosignals. In: International Conference on Learning Representations (ICLR) (2024) 2. Carrington, M., Providência, R., Chahal, C.A.A., Ricci, F., Epstein, A.E., Gallina, S., Fedorowski, A., Sutton, R., Khanji, M.Y.: Monitoring and diagnosis of intermittent arrhythmias: evidence-based guidance and role of novel monitoring strategies. European Heart Journal Open 2(6), oeac072 (2022) 3. Clifford, G.D., Behar, J., Li, Q., Rezek, I.: Signal quality indices and data fusion for determining clinical acceptability of electrocardiograms. Physiological Measurement 33(9), 1419–1433 (2012) 4. Gao, Y., Kim, S., McIntosh, C.: EchoingECG: An electrocardiogram cross-modal model for echocardiogram tasks. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 175–185. Springer (2025) 5. Goldberger, A.L., Amaral, L.A., Glass, L., Hausdorff, J.M., Ivanov, P.C., Mark, R.G., Mietus, J.E., Moody, G.B., Peng, C.K., Stanley, H.E.: PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals. Circulation 101(23), e215–e220 (2000) 6. Hong, S., Xiao, C., Ma, T., Li, H., Sun, J.: MINA: Multilevel knowledge-guided attention for modeling electrocardiography signals. In: 28th International Joint Conference on Artificial Intelligence, IJCAI 2019. pp. 5888–5894 (2019) 7. Huang, Z., MacLachlan, S., Yu, L., Herbozo Contreras, L.F., Truong, N.D., Ribeiro, A.H., Kavehei, O.: Generalization challenges in electrocardiogram deep learning: insights from dataset characteristics and attention mechanism. Future Cardiology 20(4), 209–220 (2024) 8. Hwang, U., Lee, J., Shin, J., Yoon, S.: SF(DA)2: Source-free domain adaptation through the lens of data augmentation. In: 12th International Conference on Learning Representations, ICLR 2024 (2024) 9. Kang, J., Kim, N., Ok, J., Kwak, S.: MemBN: Robust test-time adaptation via batch norm with statistics memory. In: European Conference on Computer Vision (ECCV). pp. 467–483. Springer (2024) 10. Kim, C., Park, J., Shim, H., Yang, E.: SGEM: Test-time adaptation for automatic speech recognition via sequential-level generalized entropy minimization. In: Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH. vol. 2023, pp. 3367–3371 (2023) 11. Kim, H., Kim, S., Mok, J., Yoon, S.: Battling the non-stationarity in time series forecasting via test-time adaptation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 17868–17876 (2025) 12. Lee, J., Jung, D., Lee, S., Park, J., Shin, J., Hwang, U., Yoon, S.: Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. In: 12th International Conference on Learning Representations, ICLR 2024 (2024) 13. Li, Y., Wang, N., Shi, J., Liu, J., Hou, X.: Revisiting batch normalization for practical domain adaptation. arXiv preprint arXiv:1603.04779 (2016) 14. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International Conference on Machine Learning. pp. 6028–6039. PMLR (2020)

10

W. Jiang et al.

15. Liu, C., Cheng, S., Ding, W., Arcucci, R.: Spectral cross-domain neural network with soft-adaptive threshold spectral enhancement. IEEE Transactions on Neural Networks and Learning Systems 36(1), 692–703 (2025) 16. Liu, C., Wan, Z., Ouyang, C., Shah, A., Bai, W., Arcucci, R.: Zero-shot ECG classification with multimodal learning and test-time clinical knowledge enhancement. In: International Conference on Machine Learning. pp. 31949–31963. PMLR (2024) 17. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable testtime adaptation in dynamic wild world. In: International Conference on Learning Representations (ICLR) (2023) 18. Ong Ly, C., Unnikrishnan, B., Tadic, T., Patel, T., Duhamel, J., Kandel, S., Moayedi, Y., Brudno, M., Hope, A., Ross, H., et al.: Shortcut learning in medical AI hinders generalization: method for estimating AI model generalization without external data. npj Digital Medicine 7(1), 124 (2024) 19. Perez Alday, E.A., Gu, A., Shah, A.J., Robichaux, C., Wong, A.K.I., Liu, C., Liu, F., Bahrami Rad, A., Elola, A., Seyedi, S., Li, Q., Sharma, A., Clifford, G.D., Reyna, M.A.: Classification of 12-lead ECGs: the PhysioNet/Computing in Cardiology Challenge 2020. Physiological Measurement 41(12), 124003 (2020). https://doi.org/10.1088/1361-6579/abc960 20. Rahman, S., Karmakar, C., Natgunanathan, I., Yearwood, J., Palaniswami, M.: Robustness of electrocardiogram signal quality indices. Journal of the Royal Society Interface 19(189) (2022) 21. Ribeiro, A.H., Ribeiro, M.H., Paixão, G.M., Oliveira, D.M., Gomes, P.R., Canazart, J.A., Ferreira, M.P., Andersson, C.R., Macfarlane, P.W., Meira Jr, W., et al.: Automatic diagnosis of the 12-lead ECG using a deep neural network. Nature Communications 11(1), 1760 (2020) 22. Tarvainen, A., Valpola, H.: Mean teachers are better role models: weight-averaged consistency targets improve semi-supervised deep learning results. Advances in Neural Information Processing Systems 30 (2017) 23. Vogel, B., Claessen, B.E., Arnold, S.V., Chan, D., Cohen, D.J., Giannitsis, E., Gibson, C.M., Goto, S., Katus, H.A., Kerneis, M., et al.: ST-segment elevation myocardial infarction. Nature Reviews Disease Primers 5(1), 39 (2019) 24. Wagner, G.S., Strauss, D.G.: Marriott’s Practical Electrocardiography. Lippincott Williams & Wilkins (2013) 25. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: TENT: Fully test-time adaptation by entropy minimization. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=uXl3bZLkr3c 26. Wang, Q., Fink, O., Van Gool, L., Dai, D.: Continual test-time domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7201–7211 (2022) 27. Xu, G., Guo, H., Yi, L., Ling, C., Wang, B., Yi, G.: Revisiting source-free domain adaptation: a new perspective via uncertainty control. In: The Thirteenth International Conference on Learning Representations (2025), https://openreview.net/ forum?id=nx9Z5Kva96 28. Yuan, Y., Xu, B., Hou, L., Sun, F., Shen, H., Cheng, X.: TEA: Test-time energy adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23901–23911 (2024) 29. Zhang, C., Zheng, H., You, X., Zheng, Y., Gu, Y.: PASS: Test-time prompting to adapt styles and semantic shapes in medical image segmentation. IEEE Transactions on Medical Imaging 44(4), 1853–1865 (2025)

BeatRhythm-TTA

11

30. Zhang, Q., Bian, Y., Kong, X., Zhao, P., Zhang, C.: COME: Test-time adaption by conservatively minimizing entropy. In: International Conference on Learning Representations (ICLR) (2025) 31. Zhou, J., Wang, W., Li, S., Qu, X., Guo, X., Liu, Y., Tang, W., Lin, X., Zheng, Y.: TopoTTA: Topology-enhanced test-time adaptation for tubular structure segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 24123–24134 (2025)

Related documents

• SparrowSNN: A Hardware/software Co-design for Energy Efficient ECG Classification#152990
• Dramatic and Puzzling Changes in All Waveforms: Unraveling an ECG Mystery#7749
• Acquisition and Classification of heart rate variability using time-frequency representation.#485099
• The Electrocardiogram-Based Phase Adjustment Task (PAT-ECG): An Open-Access Laboratory Test of Cardiac Interoceptive Accuracy.#1166904
• DeepArrhythmia: Segment-Contextualized ECG Arrhythmia Classification via Selective Evidence Acquisition#1117419