AI-Powered ECG Model Shows Promise in Heart Failure Risk Prediction: JAMA

A new study published in the Journal of American Medical Association showed that a noise-adapted AI-ECG model successfully estimated heart failure risk using only lead I ECGs across diverse multinational cohorts. This suggests a potential strategy for heart failure (HF) risk stratification that could be applied using wearable and portable ECG devices, warranting further prospective studies.

Single-lead electrocardiograms (ECGs) may be recorded using portable equipment, which might allow for extensive community-based risk assessment. Thus, to determine if an artificial intelligence (AI) system can predict heart failure risk from noisy single-lead electrocardiograms, Lovedeep Dhingra and colleagues carried out this investigation.

A retrospective cohort analysis assessed persons without heart failure at baseline from UK Biobank, YNHHS, and ELSA-Brasil, utilizing outpatient ECG data. The data were evaluated between September 2023 and February 2025, with the major exposure being the AI-ECG-predicted risk of left ventricular systolic dysfunction (LVSD).

Lead I ECGs were separated to mimic wearable device signals, and a noise-adaptive AI-ECG model trained to identify LVSD was used. The model's relationship with new-onset heart failure (first HF hospitalization) was investigated. Its predictive ability was compared to the PREVENT and PCP-HF risk scores using the integrated discrimination improvement, Harrell C statistic, and net reclassification improvement.

Baseline ECGs were obtained from 192 667 YNHHS patients, 42 141 UKB participants, and 13 454 ELSA-Brasil participants. 31 (0.2%) in ELSA-Brasil, 46 (0.1%) in UKB, and 3697 (1.9%) in YNHHS experienced heart failure over a median (IQR) of 4.2 (3.7-4.5) years, 3.1 (2.1-4.5) years, and 4.6 (2.8-6.6) years, respectively.

Regardless of age, sex, comorbidities, or competing risk of mortality, a positive AI-ECG screening result for LVSD was linked to a 3- to 7-fold increased risk for HF, and every 0.1 increase in the model likelihood was linked to a 27% to 65% greater hazard across cohorts. The discrimination of AI-ECG for new-onset HF was 0.828 in ELSA-Brasil, 0.723 in YNHHS, and 0.736 in UKB.

Across cohorts, integrating AI-ECG predictions beside PCP-HF and PREVENT equations resulted in a higher Harrel C statistic. AI-ECG improved integrated discrimination by 0.091 to 0.205 vs PCP-HF and 0.068 to 0.192 vs PREVENT, as well as net reclassification by 18.2% to 47.2% vs PCP-HF and 11.8% to 47.5% vs PREVENT.

Overall, a noise-adapted AI-ECG model predicted HF risk using lead I ECGs across global cohorts, indicating a viable HF risk-stratification technique that needs to be studied prospectively employing wearable and portable ECG devices.

Source:

Dhingra, L. S., Aminorroaya, A., Pedroso, A. F., Khunte, A., Sangha, V., McIntyre, D., Chow, C. K., Asselbergs, F. W., Brant, L. C. C., Barreto, S. M., Ribeiro, A. L. P., Krumholz, H. M., Oikonomou, E. K., & Khera, R. (2025). Artificial intelligence-enabled prediction of heart failure risk from single-lead electrocardiograms. JAMA Cardiology. https://doi.org/10.1001/jamacardio.2025.0492