Abstract 9962: Longitudinal AI-Enhanced ECG for Dynamic and Competing Risk Prediction: Multicentre Development and Validation Across Fifteen Endpoints with Comparison to Single-ECG Modelling and PREVENT for Primary Cardiovascular Prevention
Abstract
Background: Artificial intelligence-enhanced electrocardiography (AI-ECG) has shown promise for deployable cardiovascular risk prediction. However, the additive value of modelling serial ECGs over single snapshots, across wide-ranging outcomes and broad populations, remains underexplored and incompletely characterised. Methods: We developed the Longitudinal AI-ECG Risk Estimator (LAIRE) model family for dynamic time-to-event prediction utilising up to ten ECGs with age and sex and AIRE+, a single-ECG comparator using only the last snapshot. LAIRE was derived from 766,713 ECG sequences in 189,539 patients from Beth Israel Deaconess Medical Center secondary-care records (50:10:40 patient-wise split) and linked to death registries and 13 ICD-coded disease and survival endpoints. We further developed LAIRE-CR-CVD, a competing-risks model which jointly predicted a cardiovascular disease composite (MI + stroke + HF + CV death) against non-CV death. We evaluated discrimination, calibration, subgroup heterogeneity and within-patient prediction trajectories on the BIDMC test split (n=310K) and two external validation cohorts - Massachusetts General Hospital (MGH, n=4.7M, 1.3M patients) and Emory University Hospital (EUH, n=890K, 350K patients). LAIRE-CR-CVD was compared to PREVENT, PCE and SCORE2 in the outpatient primary-prevention BIDMC subcohort. Results: LAIRE outperformed AIRE+ in 11/13 single-event outcomes across all three cohorts (ΔC range 1.2-3.6%; all p<0.001) with pooled uplifts including mortality +2.5%, stroke +3.6%, 3P-MACE +2.5%, and complete heart block +2.9%. Longitudinal benefit was non-significant externally for type 2 diabetes and hypertension. Benefit grew monotonically with ECG history depth and was largest in patients aged ≥75 years. Prediction variability was 48% lower with LAIRE than AIRE+ and itself prognostic - those who died showed 4.8x higher between-ECG variability than controls in BIDMC (p<0.001). In the PREVENT-eligible cohort, LAIRE-CR-CVD exceeded PREVENT 10-year discrimination (Wolbers C 0.73 vs 0.67, ΔC +6.1% [5.2-7.0%]), with integrated discrimination improvement +0.045 and superior decision-curve net benefit at every threshold, including the 7.5% guideline statin-eligibility cutoff. Furthermore, LAIRE performed well in the 70% of outpatients ineligible for PREVENT calculation due to missing laboratory data. Conclusions: Longitudinal AI-ECG provides robust improvements over single-ECG modelling across cardiovascular and survival endpoints, with benefit scaling with patient ECG history depth and marked in older patients. For outpatient primary prevention, LAIRE-CR-CVD provides meaningful improvement over PREVENT within the derivation set - extending guideline-grade risk stratification to settings where routine ECGs may augment preventive decision-making, including those with missing laboratory data.