Deep learning model discovers a new ECG biomarker for sudden cardiac death by pairing predictive and generative AI

Story Overview

A UC Berkeley-led team trained a predictive deep-learning model on hundreds of thousands of real ECG waveforms tied to Swedish health records, then paired it with a generative model that redraws synthetic traces to reveal exactly which waveform features drove high sudden cardiac death scores. The approach surfaced a previously undescribed slurred downstroke at the end of the R wave in lead aVL—an easily visible QRS change that flagged patients missed by the current left-ventricular ejection fraction standard.