Artificial Intelligence-Enabled Prediction of Heart Failure Risk from Single-Lead Electrocardiograms

Abstract

Importance: Despite the availability of disease-modifying therapies, scalable strategies for heart failure (HF) risk stratification remain elusive. Portable devices capable of recording single-lead electrocardiograms (ECGs) may enable large-scale community-based risk assessment. Objective: To evaluate whether an artificial intelligence (AI) algorithm can predict HF risk from noisy single-lead ECGs. Design, Setting, and Participants: A retrospective cohort study of individuals without HF at baseline was conducted among individuals with conventionally obtained outpatient ECGs in the integrated Yale New Haven Health System (YNHHS) and prospective population-based cohorts of the UK Biobank (UKB) and the Brazilian Longitudinal Study of Adult Health (ELSA-Brasil). Data analysis was performed from September 2023 to February 2025. Exposure: AI-ECG-defined risk of left ventricular systolic dysfunction (LVSD). Main Outcomes and Measures: Among individuals with ECGs, lead I ECGs were isolated and a noise-adapted AI-ECG model (to simulate ECG signals from wearable devices) trained to identify LVSD was deployed. The association of the model probability with new-onset HF, defined as the first HF hospitalization, was evaluated. The discrimination of AI-ECG was compared against 2 risk scores for new-onset HF (Pooled Cohort Equations to Prevent Heart Failure [PCP-HF] and Predicting Risk of Cardiovascular Disease Events [PREVENT] equations) using the Harrel C statistic, integrated discrimination improvement, and net reclassification improvement. Results: There were 192667 YNHHS patients (median [IQR] age, 56 [41-69] years; 111181 women [57.7%]), 42141 UKB participants (median [IQR] age, 65 [59-71] years; 21795 women [51.7%]), and 13454 ELSA-Brasil participants (median [IQR] age, 51 [45-58] years; 7348 women [54.6%]) with baseline ECGs. A total of 3697 (1.9%) developed HF in YNHHS over a median (IQR) of 4.6 (2.8-6.6) years, 46 (0.1%) in UKB over a median (IQR) of 3.1 (2.1-4.5) years, and 31 (0.2%) in ELSA-Brasil over a median (IQR) of 4.2 (3.7-4.5) years. A positive AI-ECG screening result for LVSD was associated with a 3- to 7-fold higher risk for HF, and each 0.1 increment in the model probability was associated with a 27% to 65% higher hazard across cohorts, independent of age, sex, comorbidities, and competing risk of death. AI-ECG's discrimination for new-onset HF was 0.723 (95% CI, 0.694-0.752) in YNHHS, 0.736 (95% CI, 0.606-0.867) in UKB, and 0.828 (95% CI, 0.692-0.964) in ELSA-Brasil. Across cohorts, incorporating AI-ECG predictions alongside PCP-HF and PREVENT equations was associated with a higher Harrel C statistic (difference in addition to PCP-HF, 0.080-0.107; difference in addition to PREVENT, 0.069-0.094). AI-ECG had an integrated discrimination improvement of 0.091 to 0.205 vs PCP-HF and 0.068 to 0.192 vs PREVENT; it had a net reclassification improvement of 18.2% to 47.2% vs PCP-HF and 11.8% to 47.5% vs PREVENT. Conclusions and Relevance: Across multinational cohorts, a noise-adapted AI-ECG model estimated HF risk using lead I ECGs, suggesting a potential HF risk-stratification strategy requiring prospective study using wearable and portable ECG devices.