In this study, we introduce two foundation models: DeepECG-SL and DeepECG-SSL. Both rigorously evaluated on four critical clinical tasks: ECG interpretation (encompassing 77 diagnoses), LVEF prediction, 5-year incident atrial fibrillation prediction, and LQTS detection with subtype classification. Both models achieved state-of-the-art performance on our internal dataset and demonstrated robust generalizability across ten geographically diverse external datasets (373,865 ECGs in four public datasets and 447,538 ECGs in six private healthcare centers) totaling 821,403 ECGs. In this comprehensive validation across 10 medical centers and 821,403 ECGs—one of the largest in ECG AI—both models demonstrated robust clinical performance, maintaining AUROC >0.90 for 76% of diagnostic categories despite diverse patient populations and institutional labeling practices. Our findings demonstrate distinct clinical advantages for each model: DeepECG-SSL excelled in detecting emerging biomarkers with limited training data (LVEF classification, LQTS detection), while the more efficient DeepECG-SL (60-fold smaller, 29-fold faster inference) matched performance on traditional ECG diagnoses. Both models showed minimal performance degradation across external validations (AUROC decline 0.02-0.05), supporting their reliability across diverse clinical settings. To facilitate clinical adoption, we provide comprehensive open-source access to model weights, preprocessing tools, multilingual diagnostic extraction, and validation code. This standardized framework facilitates cost-effective local fine-tuning and reproducible implementation.

Recent advances in ECG foundation models includes ECGFounder²¹, a supervised algorithm trained on 10 million ECGs, ECG-FM²² a self-supervised model, and KED²³ (Knowledged-enhanced ECG) a contrastive learning-based model. All have demonstrated progress in automated ECG interpretation. ECGFounder achieved high internal performance (AUROC ≥ 0.95 in 82 out of 150 diagnostic labels) but required an order of magnitude more data than our study. ECG-FM, which shares our WCR-based¹⁴ SSL training method, SSL training strategy, employed 5-second ECG segments to predict 13 diagnostic labels—doubling the training data but omitting temporally dependent patterns vital for detecting conditions such as second-degree AV blocks. The KED framework used contrastive learning to align ECG signals with ECG text reports. It demonstrated impressive diagnostic performance on single-label classification tasks, with AUROC often exceeding 0.90, without requiring additional training on multiple external datasets. used contrastive learning to align ECG signals with text reports, enabling potentially infinite textual labels; however, it is primarily suited to single-label tasks, may struggle with multi-label scenarios, and did not evaluate or release its model for conditions not explicitly mentioned in ECG reports (e.g., left ventricular dysfunction). DeepECG qualifies as a foundation model because it is trained on a large, heterogeneous dataset encompassing a broad range of clinical conditions, contains an extensive number of parameters, and demonstrates strong performance across multiple external datasets. Moreover, it exhibits label-efficient adaptability, making it suitable for various downstream tasks without full re-training. This aligns with the Stanford HAI definition²⁴, which emphasizes comprehensive pretraining on broad data to enable general-purpose capabilities.

Our models demonstrated consistently high AUROC (> 0.95) for common diagnoses such as sinus rhythm, atrial fibrillation, right bundle branch block, and ventricular pacing. Performance remained robust even for more complex labels like Wolff-Parkinson-White, though it varied in rarer conditions (such as Brugada), where limited sample sizes likely contributed to wider confidence intervals. Notably, digital biomarker predictions such as LVEF, 5-year incident atrial fibrillation risk, and LQTS—traditionally requiring echocardiography or genetic testing—were accurately derived from routine ECGs, highlighting the potential for earlier, non-invasive detection in clinical practice. For instance, both DeepECG-SL and DeepECG-SSL surpassed Hou et al.’s²⁵ ECG-based LVEF < 50 prediction benchmark, while also delivering lower mean absolute errors (MAE) in LVEF regression. Similarly, both models improved upon our team’s prior iAF5²⁶ detection model (mean AUROC difference: 0.015 at ECG-level) on the same dataset we previously reported for validation. On LQTS detection, Jiang et al.⁶ reported marginally higher AUROCs on MHI-ds, but the differences from our model were not statistically significant (overlapping confidence intervals). Their model was trained on a unique Canadian registry (4,521 ECGs), whereas ours used a smaller MHI-ds subset (2,741 LQTS ECGs), limiting direct comparability. Access to their full training data might further improve our performance.

We addressed four critical barriers to real-world AI adoption—fairness, privacy, resource efficiency, and explainability—linking technical rigor to clinical trust. Fairness evaluations revealed minimal disparities in TPR and FPR across age and sex, mitigating biases against underdiagnosed populations. DeepECG-SSL’s strong performance across diverse centers suggests self-supervised learning can reduce label-related biases. To gauge privacy risks, we used MIA and t-SNE clustering, showing how specialized centers like MHI could be re-identified due to distinct disease distributions. Safeguards such as differential privacy²⁷ may reduce this risk but can slightly impact AUROC (drops of 0.03–0.05). Resource analysis confirmed DeepECG-SL’s suitability for low-resource environments, with faster inference, while DeepECG-SSL’s larger size excelled in data-scarce tasks, even on CPU-based setups. Finally, explainability via LIME showed both models rely on clinically meaningful waveform features, with DeepECG-SSL focusing on narrower segments—an approach akin to expert cardiologists’ attention. This transparency supports trustworthy integration into routine ECG workflows, augmenting rather than disrupting clinical practice. In terms of workflow integration, we envision running these models alongside traditional ECG systems to complement rather than replace existing interpretations. This “dual reporting” approach can bolster diagnostic confidence, refine triage priorities, and ultimately reduce unnecessary downstream testing and is currently being tested²⁸ in a randomized controlled trial (NCT 06462989).

Our open-source foundation model targets a critical unmet need by improving diagnostic access in resource-limited settings. Traditional ECG-AI solutions often require large, site-specific labeled data or proprietary software²⁹, which hinders practical deployment in smaller hospitals or under-resourced regions. In contrast, DeepECG’s label-efficient training and robust performance—even for low-data digital biomarker tasks—allow rapid local adaptation with minimal computational overhead. Notably, no comparable model is readily available for straightforward container-based deployment: our Docker solution can output predictions in standard EMR systems with minimal additional infrastructure. By enabling earlier detection of conditions like LQTS or low LVEF—often underdiagnosed where advanced cardiac testing is scarce—this approach holds promise for more equitable and cost-effective cardiovascular care worldwide.

Despite promising performance, several critical limitations warrant consideration. While UK Biobank results suggest minimal demographic bias, our predominantly North American validation cohorts necessitate broader evaluation across diverse populations, particularly regarding racial, socioeconomic, and comorbidity dimensions. The proposed “dual-reporting” approach requires careful examination of clinical workflow integration, including protocols for AI-human discrepancies, alert fatigue mitigation, and liability considerations. Although LIME-based explanations provide interpretability, their clinical utility and alignment with established ECG criteria need rigorous validation through structured cardiologist feedback. Label variability in EHR-dependent datasets remains a concern, though robust external performance partially mitigates this issue. Moreover, we did not directly compare model performance with an expert committee, but our reported accuracy aligns with or surpasses that of prior AI studies tested against cardiologist-level performance³⁰. Our framework’s demonstrated efficiency— achieving AUROC 0.97 with just 100,000 ECGs—suggests sufficient performance with modest dataset sizes, contrasting with more data-intensive domains. However, practical implementation faces substantial hurdles, including regulatory approval pathways, EMR integration costs, and computational resource requirements. Future work should prioritize prospective validation studies, systematic failure mode analysis (particularly for rare pathologies), and exploration of multimodal data integration, while maintaining focus on real-world clinical utility. While our accuracy aligns with or exceeds previous AI studies, direct comparison with expert committees would provide valuable validation of clinical readiness. Lastly, all p-values do not take into account multiple testing.

We present two extensively validated ECG foundation models across multiple clinical settings. DeepECG-SSL excels in emerging biomarker detection with limited training data (including LQTS subtyping and LVEF classification), while the more efficient DeepECG-SL matches performance on traditional diagnoses despite being substantially smaller. Both models demonstrate robust external generalization across most diagnostic categories, minimal demographic bias, and clinically aligned feature detection. Our open-source pipeline enables reproducible deployment and efficient local adaptation through standardized preprocessing and Dockerized implementation. An ongoing randomized trial evaluates clinical integration alongside existing ECG systems, with future work focused on expert committee validation and expanded clinical applications. These contributions establish a comprehensive framework for responsible, clinically validated AI in cardiovascular diagnostics