Development and evaluation of an AI-ECG model for incident heart failure risk prediction: a cross-continent study

Introduction Heart failure (HF) remains a global public health concern with an estimated prevalence of 64 million people worldwide. Accurate risk stratification of HF incidence in patients with prior heart diseases is critical for early intervention. Recent advances in artificial intelligence (AI) have enabled large scale analysis of widely-accessible electrocardiograms (ECGs) to support data-driven risk modelling. While such AI-ECG models have shown promising results, most have only been evaluated on populations drawn from the same cohorts used for model development, which means their generalisability across diverse populations remains unclear in the global context. Purpose We run a cross-sectional, retrospective study with 2 aims: (i) to develop an AI-ECG model that combines ECG with demographics to predict risk of incident HF, and (ii) to assess model generalisability by systematically validating on cross-continental populations. Our study centres on patients with prevalent hypertension and/or myocardial infarction, a clinically important group for studying HF progression. Methods We use the Harvard–Emory dataset (N=168,925) to develop a Cox-based AI-ECG model (AI-ECG-HF) to estimate incident HF risk (Fig.1). For each patient, AI-ECG-HF extracts deep features from 12-lead ECG median waves and integrates age and sex information through an adaptive gating fusion mechanism. This enables individualized re-weighting of the inputs to predict a hazard ratio representing HF risk. Generalisation was assessed through internal 10-fold cross-validation within the Harvard–Emory cohort (N=18,769) and external evaluation across 3 geographically distinct datasets: (a) MIMIC-IV (N=29,197), (b) Alberta Cohort (AB Cohort) (N=97,462), and (c) UK Biobank Cohort (UKB) (N=18,292) (Table 1). We report model performance using Harrell’s C-index. We further compared our AI-ECG-HF against 3 established clinical risk scores for 5-year HF onset: Health ABC, WATCH-DM, and the modified PCP-HF, which rely on conventional clinical, medication, and blood biomarker variables, using UKB data. Results Our model achieved C-indices of 0.765 [95% confidence interval (CI) 0.756-0.773] internally on Harvard-Emory, and maintained robust risk stratification across external centers: 0.720 [95% CI 0.711-0.729] on MIMIC-IV, 0.725 [95% CI 0.721-0.729] on the AB Cohort, and 0.762 [95% CI 0.732-0.792] on the UKB. Furthermore, compared to traditional clinical approaches, our model demonstrated significantly higher ranking performance (two-sided t-tests, p<0.0001), with 6.9% to 22.5% relative improvements in C-index. Conclusion We demonstrated that our AI-ECG model can achieve better HF risk stratification than conventional risk scores and maintain robust performance across diverse populations. Our study highlights the potential of AI-driven, ECG-based risk assessment as a scalable and cost efficient tool to support more advanced proactive and preventive HF care globally.