Key result
An image-based convolutional neural network accurately identified six clinical rhythm and conduction disorders from printed ECG images, achieving an average AUROC of 0.99 on a held-out test set.
Why the study?
AI for automated ECG diagnosis is limited by reliance on signal-based data that are not routinely available in remote settings.
Does an image-based convolutional neural network accurately diagnose rhythm and conduction disorders from printed ECGs compared to signal-based models?
Observational (n=1,506,112)
Yes
Does an image-based convolutional neural network accurately diagnose rhythm and conduction disorders from printed ECGs compared to signal-based models?
Effect estimate: AUROC 0.99
An AI model trained on ECG images can accurately diagnose multiple rhythm and conduction disorders, performing comparably to signal-based models and enabling broader application of automated ECG interpretation in settings without access to raw signal data.
May enable ECG AI in signal-scarce settings; leaves open prospective clinical validation before adoption.
Aims The application of artificial intelligence (AI) for automated diagnosis of electrocardiograms (ECGs) can improve access to high quality diagnostic care in remote settings but is limited by the reliance on signal-based data that are not routinely available. We sought to develop a multilabel automated diagnosis model for electrocardiographic images, more suitable for broader use. Methods and Results A total of 2,228,236 12-lead ECGs from 811 municipalities in Minas Gerais, Brazil were sampled into 90% (training):5%(validation):5%(held-out test), and were transformed to ECG images with varying lead locations and formats. We trained a convolutional neural network (CNN) using an EfficientNet-based architecture on ECG images to identify 6 physician-defined clinical labels spanning rhythm and conduction disorders, as well as a hidden label for gender. We trained another CNN for signal-based classification. The image-based model performed well on the held-out test set (average AUROC 0.99, AUPRC 0.68). This was replicated in a distinct test set from Brazil validated by at least two cardiologists (average AUROC 0.99, AUPRC 0.86) as well as an external validation set of 21,785 ECGs from Germany (average AUROC 0.97, AUPRC 0.73), with performance superior to signal-based models. Expert review of 120 out of 120 high confidence false positive predictions on the held-out and external validation sets were confirmed to be true positives with incorrect labels. The model learned generalizable features, confirmed using Gradient-weighted Class Activation Mapping (Grad-CAM). Conclusion We developed an externally validated model that extends the automated diagnosis of key rhythm and conduction disorders to printed ECGs as well as to the identification of hidden features, allowing the application of AI to ECGs captured across broad settings.
No takes yet. Share an insight, caveat, or question.
Sangha et al. (2021) conducted an observational in Rhythm and conduction disorders (n=1,506,112). Image-based convolutional neural network (EfficientNet-B3) vs. Signal-based CNN model was evaluated on Average Area Under the Receiver Operating Characteristic (AUROC) curve for 6 clinical labels on the held-out test set (AUROC 0.99). An image-based convolutional neural network accurately identified six clinical rhythm and conduction disorders from printed ECG images, achieving an average AUROC of 0.99 on a held-out test set.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: