A convolutional neural network trained on raw 12-lead ECGs from Zhongshan Hospital and validated in two independent cohorts — including Boston's Beth Israel Deaconess Medical Center — generated individualized, day-by-day probabilities of developing heart failure with reduced ejection fraction (HFrEF) over a 5-year window. Across 458,884 patients, the model achieved C-indices of 0.971, 0.945, and 0.855 in the three cohorts respectively, with calibration and decision-curve analyses confirming real-world clinical utility. Attention-based interpretability frameworks identified QRS duration, heart rate, and QT interval as the dominant predictive ECG features.
HFrEF carries roughly 50% five-year mortality once diagnosed, making pre-symptomatic identification enormously valuable. Current screening relies on echocardiography — expensive, operator-dependent, and poorly suited to population-scale deployment. An ECG-based model that slots into routine clinical workflows could flag high-risk patients years before structural deterioration becomes irreversible, enabling earlier initiation of guideline-directed medical therapy such as ARNI, SGLT2 inhibitors, and beta-blockers shown to reduce mortality by 30–40%.
Several caveats warrant caution. Performance dropped meaningfully in the U.S. cohort (C-index 0.855 vs. 0.97 in Chinese centers), hinting at population-specific ECG phenotypes or data-quality differences that could limit generalizability. The study is retrospective and observational — causal benefit of model-guided intervention remains unproven. As a preprint posted on medRxiv and not yet peer-reviewed, these results must be treated as preliminary. Independent prospective validation across broader racial and geographic cohorts is essential before clinical deployment. Still, the scale, multinational design, and interpretability rigor make this one of the more compelling AI-ECG prognostic studies to date.
