Electrocardiogram-based deep learning to predict left ventricular systolic dysfunction in paediatric and adult congenital heart disease in the USA: a multicentre modelling study.

Summary

In >49,000 patients and >200,000 ECG–echocardiogram pairs spanning paediatric and adult congenital heart disease, a CNN-based AI-ECG achieved AUROC 0.95 (internal) and 0.96 (external) for detecting LVEF ≤40%. AI-ECG high-risk classification predicted future LV dysfunction and mortality across lesion subgroups; salient ECG features localized to precordial QRS/T territories.

Key Findings

• CNN trained on paired ECG–echo data achieved AUROC 0.95 (AUPRC 0.33) internally and AUROC 0.96 (AUPRC 0.25) externally for detecting LVEF ≤40%.
• AI-ECG high-risk classification in patients with LVEF >40% predicted future LV dysfunction (HR 12.1) and higher mortality risk.
• Saliency analyses emphasized precordial QRS and T-wave territories; model performance was consistent across congenital lesion subtypes and paced rhythms.

Clinical Implications

AI-ECG could prioritize echocardiography, enable remote surveillance, and support earlier intervention in congenital heart disease populations at risk for LVSD.

Limitations

• Observational model development without randomized clinical utility testing
• Lesion-specific biases and site-level acquisition differences may affect generalizability

Future Directions

Prospective impact trials to test AI-ECG–guided care pathways, integration into remote monitoring, and calibration across devices and congenital lesion subgroups.