ai
Deep-Learning Model Predicts Sudden Cardiac Death From ECGs With Varying Accuracy Across Sweden, US, and Taiwan
A model trained on Swedish ECG records achieved an AUC of 0.872 for sudden cardiac death prediction. It maintained strong performance on U.S. and Taiwan data without retraining and identified a small high-risk group with elevated event rates.
ER
medinform.jmir.orgAudio version
Tap play to generate a narrated version.
A deep-learning model trained on every ECG waveform recorded in one Swedish region predicted sudden cardiac death with an AUC of 0.872. The analysis used a 40 percent lockbox containing 113,072 ECGs from 35,417 patients who had no defibrillators at the time of recording.
Post by @EricTopol on X
The model outperformed the AHA/ACC 10-year cardiovascular risk score, which reached an AUC of 0.697, and a SEER deep-learning heart-disease model, which reached an AUC of 0.655.
In the Swedish highest-risk group, defined as the top 2 percent of scores, 86.1 percent of patients would not have been flagged by reduced left ventricular ejection fraction alone. Patients in that group who later received defibrillators experienced a 54.4 percent lower mortality rate than expected, while an additional 3.8 percent per year had documented ventricular fibrillation or ventricular tachycardia.
The Sweden results followed provisional acceptance at Nature after the model was retrained on 60 percent of accessible patients.
The identical model, without modification, was tested on 251,858 ECGs from 139,613 patients under age 80 at Sharp HealthCare in San Diego recorded in 2021 and 2022. It produced a zero-shot AUC of 0.822 for predicting ventricular fibrillation or ventricular tachycardia, with the top 2.2 percent risk group showing a 29.1 percent incidence of those events.
A further validation at National Taiwan University Hospital examined 257 emergency-department cardiac-arrest patients and 4,011 controls.
The model distinguished future arrhythmic arrests from controls with a zero-shot AUC of 0.767. The manuscript presents results from three independent cohorts covering population-based Swedish data, U.S. health-system records, and a Taiwan hospital registry.
