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

Area of Science:

• Sports Cardiology
• Artificial Intelligence in Medicine
• Cardiovascular Disease Screening

Background:

• Pre-participation screening (PPS) in athletes aims to detect sudden cardiac death (SCD) risk factors.
• Electrocardiogram (ECG) is crucial for PPS, but may not detect all structural heart abnormalities.
• AI-ECG models show promise in hospitals but their utility in athlete screening is unclear.

Purpose of the Study:

• Develop and validate a deep learning (DL) AI-ECG ensemble model.
• Detect imaging-confirmed structural heart disease in competitive athletes during PPS.

Main Methods:

• A convolutional neural network (CNN) ensemble was trained on hospital ECGs.
• External validation was performed using the Italian Team for Athlete CARDiac evaluation and AI-based Risk prediction (ITACARD-AI) registry.
• Separate CNNs for valvular heart disease (VHD) and cardiomyopathies (CM) were combined using XGBoost meta-learning.

Main Results:

• External validation in 1115 athletes showed performance degradation compared to internal validation.
• Area under the receiver operating characteristic curve (AUROC) was 0.70 for VHD and 0.69 for CM.
• High negative predictive values (~99%) but low positive predictive values (≤8%) were observed.

Conclusions:

• Hospital-trained AI-ECG models have limited transportability to real-world athlete screening.
• Low disease prevalence and heterogeneous athlete physiology challenge AI-ECG application in PPS.
• AI-ECG may offer adjunctive support but faces significant hurdles in low-prevalence sports cardiology screening.