Electrocardiographic aging based on artificial intelligence and the risk of heart failure and its subtypes: a multi-national study
S Chung, S C You, S Cho, T Hwang, J H Kim, D Kim, T H Kim, J S Uhm, H N Pak, P S Yang, E Lee, Z I Attia, P A Friedman, H T Yu, B JoungAbstract
Background
Prior studies suggest that artificial intelligence (AI) can infer “biological age” from the electrocardiogram (ECG), and that this ECG-derived aging signal may improve cardiovascular risk prediction.
Purpose
To evaluate whether AI-estimated ECG aging is associated with the risk of incident HF overall and with subtype-specific risks for HF with preserved (HFpEF), mildly reduced (HFmrEF), and reduced (HFrEF) ejection fraction.
Methods
A deep learning model was trained using 3,672,020 ECGs from 837,666 participants to estimate AI-ECG age. The model was then applied to four external cohorts from South Korea, the United Kingdom, and the United States (n=112,761; 37,734; 41,552; and 93,090), excluding individuals with a history of HF. AI-ECG-aged status was defined as an age gap (AI-ECG age minus chronological age) >7 years. Incident HF was identified from diagnoses and classified into HFpEF, HFmrEF, and HFrEF; subtype analyses were performed in one South Korean cohort. Associations were assessed using time-to-event models, with competing-risk analyses for cumulative incidence.
Results
During follow-up, a total of 1,528; 202; 63; and 10,328 incident HF events occurred across the four cohorts, respectively. Compared with the normal group, the AI-ECG-aged group had a consistently higher risk of incident HF in each cohort (adjusted hazard ratios [95% CI]: 2.04 [1.80–2.31], 2.11 [1.46–3.05], 2.96 [1.63–5.37], and 1.54 [1.46–1.63]). In subtype analyses, AI-ECG-aged status was associated with increased risks across all HF subtypes (HFpEF 2.46 [2.05–2.96]; HFmrEF 2.27 [1.58–3.26]; HFrEF 2.24 [1.71–2.93]), indicating broadly comparable risk amplification rather than subtype-restricted effects. In prespecified subgroup analyses, associations remained directionally consistent across key demographic and clinical strata, with no clear evidence of heterogeneity; subtype-specific risk elevations were similarly preserved across strata, suggesting that ECG aging captures a generalized HF vulnerability signal not confined to a particular EF phenotype or baseline risk profile.
Conclusion
AI-derived ECG aging was associated with a higher risk of incident HF across multinational cohorts and demonstrated similar magnitude risk increases across EF-defined HF subtypes. ECG-based biological aging may serve as a scalable, noninvasive marker for HF risk stratification, including subtype-oriented prediction.Figure