Artificial intelligence-derived ECG age and biological cardiovascular aging in competitive athletes: determinants independent of training exposure

Background Artificial intelligence (AI)–derived ECG age is a novel biomarker of biological cardiovascular aging, with discrepancies between predicted and chronological age strongly associated with morbidity and mortality. Whether long-term high-intensity athletic training attenuates or accelerates AI-ECG–estimated aging remains unknown. Purpose To evaluate the association between AI-derived ECG age and chronological age in competitive athletes and to examine the influence of training load, cumulative training years, sleep, and psychological distress on this divergence. Methods This cross-sectional study included adult competitive athletes from multiple training centers in Brazil. Participants completed a standardized questionnaire on demographics, training history, lifestyle, and psychological distress (DASS-21). Resting 12-lead ECGs were recorded under standardized conditions, transmitted to a centralized telehealth center, and processed using a validated deep learning model to estimate biological cardiovascular age. The primary outcome was ΔECG-age (ECG-predicted minus chronological age). Associations with ΔECG-age were assessed using linear regression models. Variables with clinical relevance were included in multivariable analysis, adjusting for age, sex, weekly training volume, cumulative training years, sleep duration, and DASS-21 scores. Standardized beta coefficients were reported. Results Overall, 574 competitive athletes were included (median age 26 [interquartile range 22-31] years, 62.9% were men, 45.8% were white and 32.3% mixed race). The most frequent sports modalities were wrestling 102 (15.3%), athletics 78 (11.7%), pentathlon 56 (8.4%) and volleyball 54 (8.1%). The median ΔECG-age was 2.5 [interquartile range 2.0-6.4]. In univariable analysis, cumulative years in sport were associated with lower ΔECG-age (β = –0.28, p < 0.001). However, after multivariable adjustment for age, sex, weekly training volume, sleep duration, and psychological distress, only chronological age remained independently associated with ΔECG-age (β = –0.59, p < 0.001). Training load, cumulative exposure, sleep, and DASS-21 scores were not independently associated with accelerated AI-derived cardiovascular aging. Conclusions In competitive athletes, AI-derived ECG age appears primarily driven by chronological aging rather than cumulative training exposure. The absence of independent associations with training load or lifestyle factors suggests that long-term high-intensity athletic training does not accelerate electrical cardiovascular aging as estimated by deep learning models. These findings support the robustness of AI-ECG age as a biomarker of biological cardiovascular aging in highly trained populations.