Development and external validation of AI-ECG models in athlete pre-participation screening: Performance, limitations, and clinical implications

BACKGROUND: Pre-participation screening (PPS) in competitive athletes aims to identify cardiovascular diseases associated with sudden cardiac death (SCD). Although the 12‑lead electrocardiogram (ECG) represents the cornerstone of PPS, structural abnormalities may demonstrate limited or incomplete electrical expression, particularly in asymptomatic athletes with physiological remodeling. Artificial intelligence (AI)-enabled ECG models have shown promising performance in hospital-based populations, but their transportability to low-prevalence athlete screening environments remains uncertain. OBJECTIVES: To develop and externally validate a deep learning (DL)-based AI-ECG ensemble model for detecting imaging-confirmed structural heart disease in competitive athletes undergoing PPS. METHODS: A convolutional neural network (CNN) ensemble was trained using hospital-derived ECG images from Beth Israel Deaconess Medical Center (BIDMC, Boston, USA) and externally validated in the Italian Team for Athlete CARDiac evaluation and AI-based Risk prediction (ITACARD-AI) registry. Separate CNNs were developed for valvular heart disease (VHD) and cardiomyopathies (CM) and combined using XGBoost meta-learning. Model performance was assessed using area under the receiver operating characteristic curve (AUROC), subgroup analyses, and threshold-based evaluation. RESULTS: The ITACARD-AI cohort included 1115 competitive athletes (mean age 26 ± 13 years; 70% male), including 48 athletes (4.3%) with VHD and 30 (2.7%) with CM. External validation demonstrated substantial performance degradation compared with hospital-based internal validation. AUROC values decreased to 0.70 (95% CI 0.64-0.75) for VHD and 0.69 (95% CI 0.60-0.78) for CM, indicating only modest discrimination in the screening population. Threshold analyses showed high negative predictive values (∼99%) but persistently low positive predictive values (≤8%), reflecting limited disease enrichment and strong prevalence dependence. CONCLUSIONS: Hospital-trained AI-ECG models demonstrated limited transportability to real-world athlete screening populations. The marked reduction in external performance suggests that low disease prevalence, heterogeneous physiological remodeling, and incomplete ECG expression of structural abnormalities may substantially limit class separability in PPS environments. Although AI-ECG may provide cautious adjunctive support within physician-led workflows, these findings highlight the intrinsic challenges of applying ECG-based AI to low-prevalence sports cardiology screening populations.