Silent cardiac dysfunction — particularly the notoriously difficult-to-diagnose heart failure with preserved ejection fraction — claims lives precisely because routine screening tools miss it until late-stage damage accumulates. A tool that could flag high-risk patients from a standard or even single-lead ECG, before symptoms crystallize, would fundamentally shift the prevention window for millions of adults.
Investigators trained two deep-learning models on a dataset of over one million ECGs drawn from more than 165,000 patients at Atrium Health Wake Forest Baptist, then validated them externally on nearly 43,000 patients at the University of Tennessee Health Science Center. Both a conventional 12-lead and a single-lead (Lead-I) configuration were tested. The models classified ECGs into four clinically meaningful categories: reduced ejection fraction (EF below 40%), midrange ejection fraction (40–50%), HFpEF, and healthy controls. For the 12-lead model, area-under-the-curve performance reached 0.90 for reduced EF and 0.80 for HFpEF in holdout data, with the single-lead version performing only marginally lower — a finding with significant implications for wearable and remote monitoring applications. A boosting layer incorporating clinical risk factors did not meaningfully improve on ECG-AI alone, suggesting the waveform itself carries most of the diagnostic signal.
This work sits within a rapidly maturing field where AI-augmented electrocardiography has already demonstrated value for atrial fibrillation, low ejection fraction, and even metabolic conditions. What distinguishes this study is simultaneous multi-class classification across the full ejection fraction spectrum, including the historically elusive HFpEF phenotype. The AUC for HFpEF (0.80 in primary cohort, 0.73 externally) is modest but meaningful given that no reliable non-echocardiographic screen currently exists for that condition. Key limitations include retrospective design, predominantly adult datasets despite a pediatric validation arm, and the reality that AUC metrics do not directly translate to clinical net benefit without prospective deployment studies. This is a strong proof-of-concept finding that warrants prospective randomized evaluation before clinical adoption.