Deep Learning-Based Identification of Echocardiographic Abnormalities From Electrocardiograms.

Summary

Using 229,439 ECG–echo pairs from 8 centers, CNNs predicted 12 echocardiographic abnormalities with an AUC of 0.80 internally and 0.78 externally. A composite logistic model achieved 73.8% accuracy (sensitivity 81.1%, specificity 60.7%), supporting ECG-first triage for structural and valvular disease.

Key Findings

• Trained on 229,439 ECG–echo pairs across 8 centers; external validation performed in 2 centers.
• Composite abnormality label achieved AUC 0.80 (internal) and 0.78 (external).
• Composite logistic model: accuracy 73.8%, sensitivity 81.1%, specificity 60.7%.

Clinical Implications

Adjunctive ECG-AI could prioritize patients for echocardiography, accelerate heart failure/valvular disease detection, and optimize imaging resources, particularly in low-resource settings.

Limitations

• Retrospective development; clinical impact not tested in prospective workflow
• Moderate specificity may increase downstream echocardiography utilization

Future Directions

Prospective, randomized implementation studies to test patient outcomes, cost-effectiveness, and integration with clinical pathways; calibration across device vendors and health systems.