Image-based deep learning for emergency electrocardiogram classification

Researchers developed a deep learning system based on images for the classification of emergency electrocardiograms (ECGs). The system was tested on the InCor-EMG dataset containing 18,519 emergency ECGs in 12 categories labeled by 19 cardiologists. The final ConvNeXt model achieved an F1-score of 0.807 (95% CI, 0.788-0.825), which is comparable to cardiologist performance (0.820). The system demonstrated better results than commercial Mortara Veritas software in most categories. Model performance was more strongly associated with inter-reader agreement than with training sample size. The system performed reliably on both scanned and photographed ECGs. These findings support the use of ECG image classification when digital waveforms are unavailable.