Development and Multinational Validation of Artificial Intelligence-Enabled ASCVD Risk Stratification Using Electrocardiograms

AimsDespite the availability of clinical risk scores for atherosclerotic cardiovascular disease (ASCVD), their use is limited because the required predictor data are often missing. We developed and validated ECG-ASCVD, a scalable risk prediction paradigm that utilizes ECGs to target ASCVD risk factor assessment. MethodsAdults aged 30-79 who had undergone a clinical ECG were identified in the Yale New Haven Health System (YNNHS) and a state death index. We developed ECG-ASCVD-12, ECG-ASCVD-IMAGE, and ECG-ASCVD-1 to predict time-to-ASCVD from 12-lead ECG signals, ECG images, and lead-1 signals, respectively. Model performance was assessed in held-out individuals without prior ASCVD and in two external prospective cohorts, ELSA-Brasil (ELSA) and the UK Biobank (UKB). We then simulated the deployment of ECG-ASCVD in a random sample of 100,000 adults at YNHHS. ResultsThe development cohort included 363,788 individuals (median age, 57.1 [45.5-67.2] years; 48% Women). The YNHHS, ELSA, and UKB test cohorts included 83,917, 10,934, and 54,166 individuals, respectively. ECG-ASCVD-12 demonstrated generalizable discrimination (C-index: 0.684 to 0.746) and remained independently associated with ASCVD (adjusted hazard ratio: 1.23-1.34 per SD) after adjusting for PREVENT scores (C-index: 0.696-0.782) across the validation cohorts. ECG-ASCVD-IMAGE performed similarly (C-index: 0.673-0.748) while ECG-ASCVD-1 had modestly lower performance (C-index: 0.671-0.735). Simulated deployment suggested that ECG-ASCVD could enable the detection of high ASCVD risk patients who lack the data required for PREVENT. ConclusionWe developed an ECG-ASCVD toolkit and validated it across diverse multinational cohorts. These results highlight the potential utility of resting ECG information for predicting ASCVD risk, enabling targeted screening.