Predicting Heart Failure From 12-Lead ECGs Using AI: A HeartShare/AMP-HF Pooled Cohort Analysis

Background

Artificial intelligence applied to electrocardiograms (ECG-AI) offers a scalable approach to identify individuals at risk for heart failure (HF) and guide preventive interventions.

Objective

To assess whether ECG-AI designed to detect systolic and diastolic dysfunction enhances the prediction of incident HF over clinical risk estimation using the PREVENT-HF equation.

Methods

Baseline clinical and ECG data were pooled from FHS, MESA, and CHS. Participants with data sufficient for both ECG-AI and PREVENT-HF assessment were included. Analyses were performed on NHLBI BioDataCatalyst during July-September 2025. Risk of incident HF was estimated using previously validated ECG-AI algorithms that detect systolic (ECG-AI LEF) and diastolic (ECG-AI DD) dysfunction. Discrimination and reclassification were evaluated using Harrell’s C-statistic and net reclassification improvement (NRI).

Results

Of 14,126 participants, positive screening rates were 2.9% for ECG-AI LEF, 11.1% for ECG-AI DD, 11.9% for the composite ECG-AI model, 25.1% for PREVENT-HF score ≥10%, and 5.8% for PREVENT-HF score ≥20%. Incident HF or death occurred 7.7% and 15.1% of participants, respectively. Participants with positive composite ECG-AI screens at baseline had 10-to 20-fold higher risk of developing HF compared to those with negative screens. At 1, 3, 5, and 10 years, the addition of ECG-AI to PREVENT-HF yielded one-directional NRIs ranging from 0.086 to 0.125 at a PREVENT-HF threshold of 10%, and 0.327 to 0.403 at a threshold of 20%.

Conclusions

The addition of ECG-AI to PREVENT-HF improved discrimination of near-term HF risk. ECG-AI may enable population-level HF risk stratification and facilitate targeted prevention strategies.

@2026 anumana, Inc. All rights reserved.