Arrhythmias & electrophysiology
Foundation models for electrocardiogram interpretation: clinical implications
An erratum to this article was published (Eur Heart J, 15 May 2026 ↗)
Authors' conclusion
This study establishes SSL as a promising paradigm for ECG analysis, particularly in settings with limited annotated data, enhancing accessibility, generalizability, and fairness in AI-driven cardiac diagnostics.
Study context
- Population
- 12-lead electrocardiogram (ECGs); 77 cardiac conditions; 7 private multilingual healthcare systems and 4 public ECG repositoriesover 1 million training ECGs
- Comparison
- DeepECG-SSL vs DeepECG-SL ECG foundation models
- Outcome
- Micro-averaged AUROC across 77 cardiac conditions for ECG interpretation
Methodological detailsmulticentre, external validation, Open-source models; trained on over 1 million ECGs with automated free-text extraction from ECG reports
- Multicentre
- External validation
- Open-source models; trained on over 1 million ECGs with automated free-text extraction from ECG reports
Performance
As reported in the abstract
Micro-averaged AUROC, 77 conditions
| Test or model | AUROC internal (MHI-ds) | AUROC external public | AUROC external private |
|---|---|---|---|
| DeepECG-SSL | 0.990 (95% CI: 0.990, 0.990) | 0.981 (95% CI: 0.981, 0.981) | 0.983 (95% CI: 0.983, 0.983) |
| DeepECG-SL | 0.992 (95% CI: 0.992, 0.992) | 0.980 (95% CI: 0.980, 0.980) | 0.983 (95% CI: 0.983, 0.984) |
AUROC internal (MHI-ds)
DeepECG-SSL0.990 (95% CI: 0.990, 0.990)
DeepECG-SL0.992 (95% CI: 0.992, 0.992)
AUROC external public
DeepECG-SSL0.981 (95% CI: 0.981, 0.981)
DeepECG-SL0.980 (95% CI: 0.980, 0.980)
AUROC external private
DeepECG-SSL0.983 (95% CI: 0.983, 0.983)
DeepECG-SL0.983 (95% CI: 0.983, 0.984)
Further findings
| Outcome | DeepECG-SSL | DeepECG-SL | P value |
|---|---|---|---|
| LQTS genotype classification (n = 127 ECGs) | P = 0.026 | ||
| AUROC | 0.931 | 0.850 | |
| 5-year AF risk prediction (n = 132 050 ECGs) | P < 0.001 | ||
| AUROC | 0.742 | 0.734 | |
| LVEF ≤40% classification (n = 25 252 ECGs) | P < 0.001 | ||
| AUROC | 0.926 | 0.917 | |
| LQTS detection (n = 934 ECGs)Comparable performance | P = 0.117 | ||
| AUROC | 0.767 | 0.735 |
Fairness by age and sex
- Minimal disparities for both models (TPR and FPR difference)
- <0.1
Original publication
Foundation models for electrocardiogram interpretation: clinical implications. Eur Heart J. Published 22 January 2026. doi:10.
Composed by AI from the published abstract and checked against it. How Briefs are made →