Summary

The paper An Electrocardiogram Foundation Model Built on over 10 Million Recordings by Jun Li et al. describes a purported best-in-class electrocardiogram classification model, capabable of annotating around 20 different types of arrhythmias and cardiac abnormalities with very high area under the curve (AUROC) of around 0.95 or higher. Notably, the model was validated on completely different datasets than the one it was trained on, rather than doing an usual test/train split. It also was trained to detect axis deviations on Lead I alone (mimicking a single lead wearable device).

For Further Reading

• Positive Unlabeled (PU) learning
• DeLong method to evaluate AUROC variability.