An ECG biomarker for sudden cardiac death discovered with deep learning
TL;DR
A deep learning model trained on hundreds of thousands of ECGs identified a novel electrocardiogram biomarker that predicts sudden cardiac death, flagging a high-risk patient group largely missed by current clinical standards.
Problem / question
The current standard for predicting sudden cardiac death risk, left ventricular ejection fraction (LVEF), misses most patients who die suddenly and falsely flags many who never need the implanted defibrillators they receive.
Methods
Researchers trained a 64-layer ResNet deep learning model on 262,554 ECGs from 75,157 Swedish patients to predict sudden cardiac death within one year. They validated the model on a 40% hold-out set of 119,541 ECGs, a US dataset of 251,858 ECGs, and a Taiwanese case-control registry of 4,268 patients. They also used a Variational Autoencoder (VAE) to generate synthetic ECGs, morphing low-risk beats into high-risk ones to visualize the specific waveform changes the model detected.
Key findings
The model achieved an AUC of 0.872 in the Swedish hold-out set, identifying a high-risk group (2.2% of the sample) with a 7.0% annual sudden cardiac death rate. Notably, 86.1% of these high-risk patients were not flagged by the standard LVEF metric. High-risk patients who received defibrillators had a 54.4% lower mortality rate than expected. The generative model revealed a previously undescribed biomarker: a slurred downstroke at the end of the QRS complex in lead aVL, which independently predicted sudden cardiac death.
Why it matters
This approach identifies a large, previously unsuspected population of high-risk patients who could potentially be saved by implantable defibrillators, while also providing a concrete, visible ECG biomarker that may point to underlying mechanisms like myocardial fibrosis.
Limitations
The mortality benefit of defibrillators was estimated from observational data, which is subject to unmeasured confounding. Additionally, death certificates used for training may misclassify some non-arrhythmic deaths, and the external validation datasets had structural differences, such as lacking cause-of-death data in the US cohort.
Takeaway
Deep learning can discover novel, visually interpretable ECG biomarkers that outperform the current clinical standard for predicting sudden cardiac death.