Deep Learning Analysis of Single-Lead ECGs enables Pragmatic Heart Failure Risk Assessment in the General Population
Meraj Neyazi, Jan P Bremer, Jan Brederecke, Marius S Knorr, Ferdinand Seum, Carla Reinbold, Stefan Gross, Dora Csengeri, Stefan Blankenberg, Marcus Dörr, Marcus Vollmer, Renate Bonin-SchnabelAbstract
Effective heart failure (HF) prevention requires early identification of high-risk individuals, yet population-wide stratification remains difficult. We evaluated whether deep learning using single-lead (lead I) electrocardiograms, obtainable from medical systems and wearables, enables population-scale risk assessment. We developed AI-HF to estimate incident clinical HF risk using UK Biobank (UKB) data, validating in the prospective SHIP-START and SHIP-TREND cohorts.
The analysis included 31,740 UKB participants (median age 64, 5.2y follow-up, 243 events), 3,025 SHIP-START participants (age 50, 15y follow-up, 166 events), and 1,342 SHIP-TREND participants (age 51, 9y follow-up, 84 events). Participants with prevalent HF were excluded. Performance was evaluated at a harmonized 5-year prediction horizon. C-indices for incident clinical HF were 0.693 (95% CI 0.654–0.732) in UKB, 0.715 (0.652–0.777) in SHIP-START, and 0.791 (0.749–0.833) in SHIP-TREND. Hazard ratios per standard deviation increase in AI-HF output were 1.67 (1.56–1.79), 1.43 (1.25–1.65), and 1.46 (1.34–1.59), respectively (all p<0.001). Adding biometric variables improved discrimination modestly (C-indices: 0.714, 0.718, 0.77).
Across cohorts, AI-HF identified individuals at elevated 5-year incident clinical HF risk using single-lead ECGs. Given the ubiquity of wearables, this method may enable population-scale assessment to support targeted prevention and early intervention.