Abstract 16130: Interpretable Electrocardiographic Signatures Underlying AI-Derived Accelerated and Resilient Cardiovascular Aging
Abstract
Background
AI-ECG models can estimate biological age from routinely acquired ECGs and identify individuals at elevated cardiovascular risk, suggesting that ECGs contain measurable signatures of cardiovascular aging. Linking AI-derived aging phenotypes to specific morphologic and conduction features may establish their biological validity and improve clinical interpretability.
Methods
Using an AI-ECG aging model trained on 5,014,359 ECGs from a large, diverse US health system from 2000 to 2024, we defined 3 ECG-level cardiovascular aging phenotypes by comparing chronological age at ECG acquisition with AI-predicted age: normal (within ±1 Mean Absolute Error [MAE]), accelerated (>1 MAE older), and resilient (>1 MAE younger). To identify features independently associated with these phenotypes, we evaluated diagnostic ECG labels using multivariable logistic regression, adjusting for chronological age and EHR-recorded sex. For the subset of ECGs in sinus rhythm, we applied adjusted linear regression to continuous waveform intervals. Finally, we matched accelerated agers 1:1 to two normal-ager cohorts, by chronological age and by AI-predicted biological age, and used the standardized mean difference (SMD) in each continuous ECG feature to determine which cohort they more closely resembled.
Results
An independent test set included 1,206,874 ECGs from 374,223 patients (47.3% female; mean chronologic age 61.7 [SD 18.5]). The AI-ECG age model had an MAE of 9.35 years, classifying 59.9% ECGs as normal agers, 19.5% as accelerated agers, and 20.6% as resilient agers. Accelerated and resilient aging phenotypes showed distinct feature labels [Fig A]. For those in sinus rhythm, accelerated agers showed more prolonged AV conduction, longer ventricular depolarization, and QRS left axis deviation compared to normal agers, with resilient agers demonstrating the opposite [Fig B]. Across all continuous ECG features, accelerated-aging ECGs more closely resembled the ECGs matched to AI-predicted age than those matched to chronological age [Fig C].
Conclusions
AI-derived aging phenotypes correspond to coherent, interpretable electrophysiologic patterns of cardiovascular aging, with accelerated agers exhibiting ECG morphologies resembling those who share their biological age rather than chronological age. These signatures may enable identification of novel, clinically interpretable electrocardiographic markers of cardiovascular aging.