A new self‑supervised framework, called Deviation‑Aware Pretraining, can learn useful patterns from nearly one million unlabeled 12‑lead ECGs and then use those patterns to detect heart disease.

The approach has two parts. The first, Deviation Localization Embedding (DLE), scans each heartbeat waveform and spot‑lights any parts that deviate from a normal pattern. The second, Association Reconstruction Embedding (ARE), looks at how those deviations appear across the different leads and learns to reconstruct the overall signal from that information.

When the authors tested the system on several public ECG datasets, the combined DLE‑ARE representations improved classification performance by up to 6.5 % in AUROC and 12.8 % in F1 score compared with earlier techniques, even though the feature extractor was not further fine‑tuned for the specific datasets.

The authors also evaluated the method on a large observational cohort that contains many different patient backgrounds. The absolute predictive numbers were lower there, likely because of the cohort’s heterogeneity, but the relative gains over baseline models were still consistent.

These results suggest that a self‑trained model can reduce the need for large labeled data sets and may help clinicians identify cardiovascular conditions earlier in a variety of healthcare settings.