Robust Transfer Learning for Paper ECG Recognition

Yinghao Xie; Zhenbang Dai; Haojun Wang; Jinyu Cai; Fabio Bonassi; Hongwu Chen; Johan Sundström; Jiawei Li

cs.LGSep 30, 2026

Robust Transfer Learning for Paper ECG Recognition

Authors: Yinghao Xie, Zhenbang Dai, Haojun Wang, Jinyu Cai, Fabio Bonassi, Hongwu Chen, Johan Sundström, Jiawei Li, +1 more

Organizations: Uppsala University, Sweden · Nanjing Medical University, China · The First Affiliated Hospital with Nanjing Medical University, China

Abstract

Paper ECG recognition is challenging because real-world ECG images vary in layout, physical artifacts, and label availability. We introduce RobECG-CL, a rank-aware contrastive learning framework for robust paper ECG representation learning. Starting from standard 12-lead ECG recordings, we construct progressively degraded paper ECG views with heterogeneous layouts and train the model to balance same-recording invariance with degradation-aware ordering. Across synthetic stress tests on CODE-II and EchoNext, RobECG-CL improves robustness under severe degradation and few-shot transfer, outperforming contrastive learning baselines and surpassing the waveform-based foundation model, ECG-FM, in the 1% labeled setting. On 312 samples of hospital data with 37 labels, RobECG-CL achieves the best macro AUROC.

Figures & tables

Figure 1: Progressive degradation. For each ECG recording, we first render a clean paper ECG using a randomly sampled layout, and then generate progressively degraded views by applying cumulative visual corruptions, including rotation, geometric perturbation, blur, color shift, noise, and artifacts.

PTB-XL $\rightarrow$ downstream
Method	Clean Paper ECG		Severity 1		Severity 2		Severity 3		Severity 4		Avg.
Method	1%	10%	1%	10%	1%	10%	1%	10%	1%	10%	1%	10%
RobECG-CL	0.766 $\pm$ 0.006	0.798 $\pm$ 0.003	0.766 $\pm$ 0.005	0.797 $\pm$ 0.003	0.763 $\pm$ 0.005	0.793 $\pm$ 0.003	0.758 $\pm$ 0.006	0.786 $\pm$ 0.003	0.742 $\pm$ 0.005	0.768 $\pm$ 0.003	0.759 $\pm$ 0.005	0.788 $\pm$ 0.003
MoCo	0.756 $\pm$ 0.007	0.806 $\pm$ 0.002	0.740 $\pm$ 0.006	0.798 $\pm$ 0.002	0.717 $\pm$ 0.009	0.776 $\pm$ 0.004	0.682 $\pm$ 0.013	0.745 $\pm$ 0.004	0.639 $\pm$ 0.016	0.705 $\pm$ 0.006	0.710 $\pm$ 0.007	0.768 $\pm$ 0.003
DINO	0.707 $\pm$ 0.013	0.782 $\pm$ 0.003	0.706 $\pm$ 0.014	0.780 $\pm$ 0.003	0.684 $\pm$ 0.019	0.766 $\pm$ 0.003	0.651 $\pm$ 0.019	0.741 $\pm$ 0.004	0.614 $\pm$ 0.017	0.704 $\pm$ 0.005	0.675 $\pm$ 0.015	0.756 $\pm$ 0.003
BYOL	0.739 $\pm$ 0.006	0.803 $\pm$ 0.001	0.724 $\pm$ 0.007	0.793 $\pm$ 0.001	0.698 $\pm$ 0.009	0.772 $\pm$ 0.003	0.663 $\pm$ 0.011	0.743 $\pm$ 0.003	0.623 $\pm$ 0.013	0.708 $\pm$ 0.003	0.693 $\pm$ 0.008	0.766 $\pm$ 0.002

Table 1: AUROC across severity levels on CODE-II sex classification. Higher severity levels denote progressively stronger degradation. Values are mean

\pm

standard deviation over five runs.

PTB-XL $\rightarrow$ downstream
Method	Clean Paper ECG		Severity 1		Severity 2		Severity 3		Severity 4		Avg.
Method	1%	10%	1%	10%	1%	10%	1%	10%	1%	10%	1%	10%
RobECG-CL	0.692 $\pm$ 0.006	0.738 $\pm$ 0.002	0.691 $\pm$ 0.006	0.735 $\pm$ 0.002	0.689 $\pm$ 0.006	0.734 $\pm$ 0.002	0.684 $\pm$ 0.006	0.729 $\pm$ 0.002	0.673 $\pm$ 0.006	0.715 $\pm$ 0.002	0.686 $\pm$ 0.006	0.730 $\pm$ 0.002
MoCo	0.722 $\pm$ 0.002	0.768 $\pm$ 0.004	0.709 $\pm$ 0.004	0.756 $\pm$ 0.005	0.681 $\pm$ 0.007	0.726 $\pm$ 0.007	0.642 $\pm$ 0.010	0.690 $\pm$ 0.006	0.600 $\pm$ 0.010	0.644 $\pm$ 0.007	0.671 $\pm$ 0.005	0.717 $\pm$ 0.005
DINO	0.709 $\pm$ 0.005	0.766 $\pm$ 0.004	0.695 $\pm$ 0.006	0.758 $\pm$ 0.004	0.670 $\pm$ 0.011	0.743 $\pm$ 0.005	0.631 $\pm$ 0.012	0.708 $\pm$ 0.005	0.594 $\pm$ 0.014	0.667 $\pm$ 0.004	0.660 $\pm$ 0.009	0.728 $\pm$ 0.004
BYOL	0.711 $\pm$ 0.006	0.764 $\pm$ 0.004	0.689 $\pm$ 0.007	0.749 $\pm$ 0.004	0.663 $\pm$ 0.008	0.723 $\pm$ 0.004	0.626 $\pm$ 0.010	0.694 $\pm$ 0.005	0.586 $\pm$ 0.012	0.657 $\pm$ 0.006	0.655 $\pm$ 0.006	0.717 $\pm$ 0.004

Table 2: AUROC across severity levels on EchoNext SHD classification. Higher severity levels denote progressively stronger degradation. Values are mean

\pm

standard deviation over five runs.

Method	1AVB		AF		CRBBB		IRBBB		ISC		LAE		LVH		NORM		OMI		PAC		PVC		SB		Avg.
Method	AUC	F1	AUC	F1	AUC	F1	AUC	F1	AUC	F1	AUC	F1	AUC	F1	AUC	F1	AUC	F1	AUC	F1	AUC	F1	AUC	F1	AUC	F1
RobECG-CL	0.581	0.212	0.753	0.278	0.928	0.691	0.626	0.140	0.653	0.271	0.830	0.429	0.952	0.615	0.841	0.299	0.557	0.177	0.699	0.219	0.842	0.590	0.912	0.635	0.765	0.380
MoCo	0.541	0.171	0.752	0.393	0.877	0.597	0.645	0.192	0.736	0.346	0.777	0.226	0.944	0.516	0.755	0.246	0.628	0.186	0.764	0.156	0.899	0.548	0.892	0.558	0.767	0.345
DINO	0.499	0.171	0.778	0.295	0.880	0.571	0.604	0.213	0.750	0.293	0.702	0.222	0.877	0.324	0.839	0.293	0.628	0.212	0.799	0.203	0.903	0.635	0.848	0.467	0.759	0.325
BYOL	0.500	0.134	0.617	0.273	0.848	0.522	0.574	0.139	0.657	0.259	0.745	0.203	0.960	0.636	0.792	0.250	0.575	0.115	0.729	0.216	0.884	0.577	0.889	0.454	0.731	0.315
SimCLR	0.510	0.148	0.618	0.256	0.722	0.278	0.540	0.136	0.788	0.351	0.790	0.189	0.928	0.415	0.778	0.233	0.598	0.182	0.691	0.216	0.813	0.463	0.840	0.500	0.718	0.281

Table 3: Per-label AUROC and F1 on the 12 diagnostic labels with more than 20 samples in a real-world paper ECG dataset. Results are pooled out-of-fold performance under 5-fold evaluation. Full results on 37 labels are provided in Supplementary Table 7 .

Figure 2: UMAP visualization of representations for sex prediction with 10% labeled training data. RobECG-CL learns degradation-robust representations while preserving the relative structure among degraded views. The severely corrupted views are not forced to be close to the clean view.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Clean Paper ECG MAE		Severity 1 MAE		Severity 2 MAE		Severity 3 MAE		Severity 4 MAE		Avg. MAE		Avg. RMSE		Avg. $R^{2}$
Method	1%	10%	1%	10%	1%	10%	1%	10%	1%	10%	1%	10%	1%	10%	1%	10%
RobECG-CL	12.66 $\pm$ 0.22	11.75 $\pm$ 0.15	12.67 $\pm$ 0.19	11.78 $\pm$ 0.17	12.76 $\pm$ 0.21	11.87 $\pm$ 0.18	12.90 $\pm$ 0.23	12.07 $\pm$ 0.18	13.16 $\pm$ 0.25	12.45 $\pm$ 0.19	12.83 $\pm$ 0.21	11.98 $\pm$ 0.17	16.00 $\pm$ 0.31	15.07 $\pm$ 0.23	0.102 $\pm$ 0.035	0.204 $\pm$ 0.024
MoCo	13.10 $\pm$ 0.10	12.18 $\pm$ 0.08	13.00 $\pm$ 0.09	12.26 $\pm$ 0.07	13.25 $\pm$ 0.13	12.50 $\pm$ 0.05	13.61 $\pm$ 0.22	12.87 $\pm$ 0.06	13.88 $\pm$ 0.22	13.22 $\pm$ 0.06	13.37 $\pm$ 0.14	12.60 $\pm$ 0.05	16.48 $\pm$ 0.16	15.66 $\pm$ 0.07	0.047 $\pm$ 0.019	0.140 $\pm$ 0.007
DINO	13.28 $\pm$ 0.20	12.10 $\pm$ 0.05	13.27 $\pm$ 0.22	12.14 $\pm$ 0.05	13.42 $\pm$ 0.23	12.38 $\pm$ 0.05	13.66 $\pm$ 0.15	12.76 $\pm$ 0.06	13.90 $\pm$ 0.10	13.17 $\pm$ 0.07	13.51 $\pm$ 0.17	12.51 $\pm$ 0.05	16.64 $\pm$ 0.18	15.47 $\pm$ 0.06	0.028 $\pm$ 0.021	0.160 $\pm$ 0.006
BYOL	13.14 $\pm$ 0.12	12.12 $\pm$ 0.05	13.25 $\pm$ 0.15	12.26 $\pm$ 0.04	13.46 $\pm$ 0.20	12.50 $\pm$ 0.03	13.69 $\pm$ 0.19	12.87 $\pm$ 0.02	13.88 $\pm$ 0.13	13.22 $\pm$ 0.02	13.48 $\pm$ 0.15	12.59 $\pm$ 0.01	16.66 $\pm$ 0.16	15.65 $\pm$ 0.03	0.027 $\pm$ 0.019	0.141 $\pm$ 0.003
SimCLR	13.24 $\pm$ 0.09	12.60 $\pm$ 0.08	13.28 $\pm$ 0.06	12.65 $\pm$ 0.09	13.52 $\pm$ 0.07	12.75 $\pm$ 0.07	13.72 $\pm$ 0.09	13.02 $\pm$ 0.03	13.86 $\pm$ 0.13	13.31 $\pm$ 0.06	13.52 $\pm$ 0.08	12.87 $\pm$ 0.05	16.68 $\pm$ 0.10	15.94 $\pm$ 0.06	0.025 $\pm$ 0.011	0.108 $\pm$ 0.007

Appendix

Table 4: Transfer performance across degradation severity levels on CODE-II age prediction. Higher severity levels denote progressively stronger degradation. Values are mean

\pm

standard deviation over five runs.

PTB-XL $\rightarrow$ downstream
Method	Clean Paper ECG		Severity 1		Severity 2		Severity 3		Severity 4		Avg.
Method	1%	10%	1%	10%	1%	10%	1%	10%	1%	10%	1%	10%
RobECG-CL	0.695 $\pm$ 0.006	0.739 $\pm$ 0.007	0.695 $\pm$ 0.005	0.736 $\pm$ 0.006	0.692 $\pm$ 0.006	0.731 $\pm$ 0.006	0.688 $\pm$ 0.006	0.725 $\pm$ 0.005	0.671 $\pm$ 0.005	0.705 $\pm$ 0.005	0.688 $\pm$ 0.005	0.727 $\pm$ 0.006
MoCo	0.696 $\pm$ 0.014	0.755 $\pm$ 0.005	0.678 $\pm$ 0.012	0.744 $\pm$ 0.005	0.650 $\pm$ 0.012	0.717 $\pm$ 0.006	0.611 $\pm$ 0.014	0.680 $\pm$ 0.006	0.568 $\pm$ 0.015	0.637 $\pm$ 0.005	0.645 $\pm$ 0.011	0.710 $\pm$ 0.004
DINO	0.648 $\pm$ 0.014	0.722 $\pm$ 0.007	0.647 $\pm$ 0.013	0.720 $\pm$ 0.005	0.622 $\pm$ 0.016	0.707 $\pm$ 0.003	0.583 $\pm$ 0.018	0.680 $\pm$ 0.003	0.541 $\pm$ 0.016	0.640 $\pm$ 0.003	0.612 $\pm$ 0.013	0.696 $\pm$ 0.004
BYOL	0.665 $\pm$ 0.008	0.745 $\pm$ 0.004	0.650 $\pm$ 0.006	0.736 $\pm$ 0.004	0.624 $\pm$ 0.006	0.713 $\pm$ 0.004	0.590 $\pm$ 0.009	0.680 $\pm$ 0.003	0.549 $\pm$ 0.011	0.640 $\pm$ 0.003	0.619 $\pm$ 0.005	0.706 $\pm$ 0.003

Appendix

Table 5: AUPRC across degradation severity levels on CODE-II sex classification. Higher severity levels denote progressively stronger degradation. Values are mean

\pm

standard deviation over five runs.

PTB-XL $\rightarrow$ downstream
Method	Clean Paper ECG		Severity 1		Severity 2		Severity 3		Severity 4		Avg.
Method	1%	10%	1%	10%	1%	10%	1%	10%	1%	10%	1%	10%
RobECG-CL	0.624 $\pm$ 0.008	0.683 $\pm$ 0.003	0.622 $\pm$ 0.009	0.679 $\pm$ 0.003	0.621 $\pm$ 0.009	0.679 $\pm$ 0.004	0.617 $\pm$ 0.010	0.673 $\pm$ 0.003	0.601 $\pm$ 0.009	0.654 $\pm$ 0.004	0.617 $\pm$ 0.009	0.674 $\pm$ 0.003
MoCo	0.663 $\pm$ 0.004	0.715 $\pm$ 0.005	0.644 $\pm$ 0.005	0.698 $\pm$ 0.005	0.611 $\pm$ 0.007	0.665 $\pm$ 0.007	0.571 $\pm$ 0.014	0.626 $\pm$ 0.006	0.528 $\pm$ 0.011	0.572 $\pm$ 0.006	0.609 $\pm$ 0.005	0.661 $\pm$ 0.005
DINO	0.650 $\pm$ 0.006	0.712 $\pm$ 0.003	0.628 $\pm$ 0.007	0.701 $\pm$ 0.003	0.600 $\pm$ 0.013	0.679 $\pm$ 0.003	0.555 $\pm$ 0.017	0.636 $\pm$ 0.005	0.520 $\pm$ 0.016	0.594 $\pm$ 0.005	0.595 $\pm$ 0.010	0.669 $\pm$ 0.003
BYOL	0.654 $\pm$ 0.005	0.709 $\pm$ 0.003	0.624 $\pm$ 0.008	0.690 $\pm$ 0.003	0.597 $\pm$ 0.009	0.665 $\pm$ 0.004	0.558 $\pm$ 0.014	0.631 $\pm$ 0.005	0.516 $\pm$ 0.013	0.588 $\pm$ 0.005	0.593 $\pm$ 0.007	0.660 $\pm$ 0.003

Appendix

Table 6: AUPRC across degradation severity levels on EchoNext structural heart disease classification. Higher severity levels denote progressively stronger degradation. Values are mean

\pm

standard deviation over five runs. For baseline methods, severity-specific AUPRC is not available in the source table and is marked as missing.

Label	RobECG-CL			MoCo			DINO			BYOL			SimCLR			PULSE			GEM
Label	AUROC	AUPRC	F1	AUROC	AUPRC	F1	AUROC	AUPRC	F1	AUROC	AUPRC	F1	AUROC	AUPRC	F1	AUROC	AUPRC	F1	AUROC	AUPRC	F1
1AVB (36)	0.581	0.149	0.212	0.541	0.158	0.171	0.499	0.124	0.171	0.500	0.167	0.134	0.510	0.129	0.148	0.531	0.136	0.131	0.532	0.176	0.140
AF (28)	0.753	0.278	0.278	0.752	0.325	0.393	0.778	0.309	0.295	0.617	0.205	0.273	0.618	0.220	0.256	0.750	0.264	0.275	0.818	0.385	0.244
CRBBB (36)	0.928	0.724	0.691	0.877	0.656	0.597	0.880	0.581	0.571	0.848	0.563	0.522	0.722	0.248	0.278	0.895	0.676	0.597	0.908	0.693	0.677
IRBBB (24)	0.626	0.157	0.140	0.645	0.259	0.192	0.604	0.209	0.213	0.574	0.139	0.139	0.540	0.128	0.136	0.569	0.159	0.162	0.607	0.198	0.194
ISC (41)	0.653	0.240	0.271	0.736	0.249	0.346	0.750	0.317	0.293	0.657	0.211	0.259	0.788	0.428	0.351	0.688	0.246	0.205	0.726	0.264	0.270

Appendix

Table 7: Per-label AUROC, AUPRC, and F1 on all 37 diagnostic labels in the real-world paper ECG set collected from The First Affiliated Hospital with Nanjing Medical University. Results are pooled out-of-fold performance under strict patient-aware 5-fold evaluation.

Figure 3: Diagnostic-label centroids in the semantic UMAP spaces of RobECG-CL, BYOL, DINO, MoCo, and SimCLR. Here, $n$ denotes the number of ECGs positive for each label. For each method, the standardized per-label logits are projected into a two-dimensional UMAP space. Each colored point is then calculated as the arithmetic mean of the UMAP coordinates of all ECGs positive for the corresponding label. Edges connect label pairs with the strongest true-label co-occurrence, while the number on each edge indicates the number of ECGs positive for both labels.

Figure 4: Grad-CAM heatmap visualizations on real-world paper ECGs. Each panel shows the attribution map for one diagnostic label from the external hospital dataset. RobECG-CL highlights localized ECG regions across diverse diagnoses, suggesting that the learned representation remains interpretable under real-world paper ECG degradation.

Method	Step Distribution ( $\epsilon$ )	Physical Interpretation	Implementation
Rotation	$\mathcal{N}(0,5)$	Paper misalignment during scanning/photography.	iaa.Affine
Color Temperature	$\mathcal{N}(0,300)$	Variations in white-balance and ambient illumination.	iaa.ChangeColorTemperature
Cropping	$\mathcal{N}(0.002,0.01)$	Partial border loss or automatic edge trimming.	iaa.Crop
Noise	$\mathcal{N}(2,6)$	Pixel-level noise introduced by camera/scanner sensors.	iaa.AdditiveGaussianNoise
Gaussian Blur	$\mathcal{N}(1,3)$	Defocus, motion blur, and image-softening effects.	cv2.GaussianBlur

Appendix

Table 8: Summary of progressive degradation methods, parameter step distributions, physical interpretations, and corresponding implementations.

Panel of Figure 5 — Figure 5: Examples of anonymized paper ECG images from the real-world cohort.

Method	Max epoch	Actual epoch	Pretraining LR	Optimizer	Early stopping strategy
SimCLR	50	38	$3\times 10^{-2}$	SGD	Patience = 5; validation contrastive loss
MoCo	80	26	$3\times 10^{-2}$	SGD	Patience = 5; validation contrastive loss
DINO	50	50	$5\times 10^{-4}$	AdamW	Patience = 5; validation self-distillation loss
BYOL	50	16	$3\times 10^{-2}$	SGD	Patience = 8; validation BYOL loss

Appendix

Table 9: Pretraining settings of baseline methods.

Pareto weight	Clean Paper ECG			Severity 1			Severity 2			Severity 3			Severity 4			Avg.
Pareto weight	AUROC	AUPRC	F1	AUROC	AUPRC	F1	AUROC	AUPRC	F1	AUROC	AUPRC	F1	AUROC	AUPRC	F1	AUROC	AUPRC	F1
$0$	0.835	0.637	0.611	0.836	0.638	0.611	0.831	0.631	0.607	0.820	0.618	0.597	0.803	0.592	0.576	0.825	0.623	0.600
$\mathbf{0.5}$ (RobECG-CL)	0.841	0.649	0.619	0.839	0.649	0.620	0.837	0.644	0.617	0.823	0.624	0.607	0.805	0.599	0.575	0.829	0.633	0.607
$1.0$	0.838	0.644	0.617	0.836	0.641	0.614	0.834	0.638	0.614	0.823	0.621	0.599	0.802	0.590	0.581	0.827	0.627	0.605
$2.0$	0.840	0.648	0.620	0.840	0.646	0.617	0.836	0.640	0.615	0.820	0.617	0.594	0.787	0.570	0.551	0.825	0.624	0.599

Appendix

Table 10: Ablation study.

Explore similar work

A Signal-Language Foundation Model for Broad-Spectrum Cardiovascular Assessment from Routine Electrocardiography
May 25, 2026Ziqing Yu, Yuhui Tao, Jiayu Huo +19Electrocardiogram Foundation Models Adult-To-Pediatric Electrocardiogram Transfer
Department of Cardiology, Zhongshan Hospital of Fudan University, Shanghai Institute of Cardiovascular Diseases, National Clinical Research Centre for Interventional Medicine, Shanghai, China · Digital Medical Research Center, School of Basic Medical Sciences, Fudan University, Shanghai, 200032, China. · Shanghai Key Laboratory of Medical Imaging Computing and Computer Assisted Intervention, Shanghai, 200032, China. +8
ELF: A Family of Encoder-Free ECG-Language Models
Jan 5, 2026William Han, Tony Chen, Chaojing Duan +6Electrocardiography Cnn-Lstm
Carnegie Mellon University, USA · Allegheny Health Network, USA · University of California, Los Angeles, USA +1
Pretraining Strategies and Scaling for ECG Foundation Models: A Systematic Study
May 12, 2026M A Al-Masud, Nils StrodthoffElectrocardiogram Foundation Models Electrocardiography
AI4Health Division Carl von Ossietzky Universität Oldenburg Oldenburg, Lower Saxony, Germany

May 25, 2026, cs.AIJ/K move · Enter open · S save

A Signal-Language Foundation Model for Broad-Spectrum Cardiovascular Assessment from Routine Electrocardiography

Ziqing Yu, Yuhui Tao, Jiayu Huo, Lei Pan+18

Department of Cardiology, Zhongshan Hospital of Fudan University, Shanghai Institute of Cardiovascular Diseases, National Clinical Research Centre for Interventional Medicine, Shanghai, China · Digital Medical Research Center, School of Basic Medical Sciences, Fudan University, Shanghai, 200032, China. · Shanghai Key Laboratory of Medical Imaging Computing and Computer Assisted Intervention, Shanghai, 200032, China.+8

Electrocardiogram Foundation Models Adult-To-Pediatric Electrocardiogram Transfer Electrocardiography Cardiac Assessment

Open paper AskPDF arXiv

Electrocardiography (ECG) is central to cardiovascular care, but conventional AI models are often restricted to common arrhythmias and may generalize poorly across populations or clinically subtle diseases. We developed ECG Contrastive Language-Image Pre-training (ECGCLIP), a signal-language contrastive learning framework that aligns ECG waveforms with expert diagnostic reports. ECGCLIP was pre-trained on 2,837,962 ECG studies from 1,324,856 patients and evaluated on a held-out internal test set plus nine independent external cohorts comprising about 1.5 million ECGs. Evaluation covered 89 downstream tasks, including 45 ECG diagnoses, 39 echocardiographic targets, and 5 rare cardiac diseases, using PRAUC as the primary metric. ECGCLIP consistently improved performance over random initialization and Merl-R18 baselines. On the internal test set, ECGCLIP-R34 achieved strong performance for atrial fibrillation (PRAUC 0.900) and ST-segment elevation myocardial infarction (PRAUC 0.383), with robust generalization across all external cohorts. It also improved low-prevalence and diagnostically elusive diseases, including Ebstein anomaly, constrictive pericarditis, dextrocardia, and cardiac amyloidosis, with internal PRAUC values of 0.253, 0.175, 0.121, and 0.201, respectively. ECGCLIP was data efficient, matching or exceeding full-dataset baseline performance with only 10% of training data. Feature visualization and saliency analysis suggested clinically meaningful representations aligned with established electrocardiographic criteria. These findings indicate that large-scale ECG-report contrastive pre-training can expand routine ECG interpretation beyond common arrhythmias toward broad cardiovascular assessment and opportunistic screening of echocardiographic and rare conditions.

Abstract

Explore similar work

A Signal-Language Foundation Model for Broad-Spectrum Cardiovascular Assessment from Routine Electrocardiography

ELF: A Family of Encoder-Free ECG-Language Models

Pretraining Strategies and Scaling for ECG Foundation Models: A Systematic Study