An ECG biomarker for sudden cardiac death discovered with deep learning

Key Points:

• A deep learning model was developed to predict sudden cardiac death (SCD) within one year after an ECG, trained on Swedish health data using death certificates as labels, achieving an AUC of 0.872, outperforming traditional cardiovascular risk scores.
• The model identifies a high-risk group (2.2% of patients) with an annual SCD rate of 7.0%, exceeding rates from randomized defibrillator trials, suggesting these patients could benefit from preventive defibrillator implantation.
• Validation in external datasets from the USA and Taiwan showed robust performance in predicting arrhythmic events (ventricular fibrillation/tachycardia), with zero-shot AUCs of 0.822 and 0.767 respectively, confirming model generalizability across diverse populations and data sources.
• Analysis indicates that high-risk patients with defibrillators have significantly lower sudden cardiac and all-cause mortality, supporting the model’s clinical utility in identifying preventable arrhythmic deaths despite potential confounding in observational data.
• A novel ECG morphology feature—slurred terminal R wave in lead aVL—was discovered using a generative model guided by the predictive model, linked to myocardial fibrosis and conduction abnormalities, offering new insights into the mechanisms underlying SCD and avenues for future research.