A composite AI-enabled ECG system combining independently trained models for left ventricular systolic dysfunction (LVSD) and diastolic dysfunction (LVDD) detected prevalent structural heart disease (SHD) with sensitivity of 71.8–76.1% and specificity of 70.1–88.3% across 82,368 participants in Korean and US clinical datasets. Among at-risk individuals in the Korean cohort and UK Biobank, AI-ECG positivity predicted incident SHD with hazard ratios of 3.75 and 2.75, respectively, and C statistics of 0.69–0.78 — suggesting meaningful discriminative power beyond chance.
Structural heart disease — encompassing reduced ejection fraction, significant valvular disease, LV hypertrophy, and pulmonary hypertension — silently precedes most heart failure hospitalizations, yet echocardiography is too resource-intensive for population-wide deployment. This composite AI-ECG approach represents a genuinely scalable alternative, leveraging existing 12-lead ECG infrastructure already embedded in clinical workflows globally. The finding that the composite captured signals across SHD phenotypes beyond its training targets is particularly noteworthy, hinting at generalized cardiac remodeling signatures in electrical activity.
However, important caveats apply. C statistics in the 0.69–0.78 range indicate moderate — not exceptional — discrimination. Specificity in the US cohort dropped to 70.1%, raising false-positive burden concerns. Crucially, this is a preprint posted on medRxiv and has not yet been peer-reviewed; findings, effect sizes, and conclusions may change substantially upon independent scrutiny. No randomized evidence yet demonstrates that ECG-based risk stratification improves clinical outcomes — a gap the authors themselves acknowledge requires prospective evaluation.
