Classification of “Athlete's Heart” using machine learning of conventional 12-lead ECG: male elite 3,000-m runner data in the CHIEF study

Background Conventional electrocardiographic (ECG) interpretive algorithms often struggle to distinguish physiological cardiovascular adaptations in athletes from cardiac pathology. This study utilized machine learning (ML) to estimate the likelihood of elite 3,000-meter running performance using resting ECG and biological markers within a large military cohort. Methods We analyzed data from 2,296 physically active military males in the Cardiorespiratory Fitness and Health in Eastern Armed Forces (CHIEF) Heart Study. Three ML classifiers—logistic regression (LR), multilayer perceptron (MLP), and support vector machine (SVM)—were trained using 26 ECG features (axis, duration and voltage of P, QRS and T waves in each lead and supine heart rate) and 6 biological features (age, body weight, body height, waist circumference, blood pressure and sitting pulse rate) to classify “elite” runners (defined as the top 5% and 10% performance groups). The whole data were randomly grouped by a 3:1 ratio into a training/validation set ( N = 1,607) and a test set ( N = 689). Results For the top 10% runners, LR demonstrated the highest discriminative power (AUC: 80.81%), followed by MLP (78.43%). In the top 5% group, MLP performed best with an AUC of 74.42%. When specificity was fixed at approximately 60%–70%, the sensitivity of the optimal models for both groups exceeded 81%. Conclusions ML can effectively classify elite endurance capacity using non-invasive resting ECG markers. These findings highlight the potential for updating automated ECG interpretive algorithms to better recognize “athlete's heart.” Furthermore, this approach may serve as a cost-effective preliminary screening tool for identifying elite athletic potential, although further validation in female and more diverse populations is warranted.