Evaluation of Machine Learning Models and Feature Representations for 12-Lead Electrocardiogram Classification in School-age Children

In Japan, electrocardiogram (ECG) screening is routinely conducted in school-age children to detect early signs of cardiac disease and reduce sudden death. However, interpreting pediatric ECGs requires substantial clinical expertise, leading to variability and potential inefficiency. This study aimed to develop machine learning (ML) models classifying ECGs as normal or abnormal based on national screening guidelines, with an emphasis on maintaining high sensitivity while improving specificity. We analyzed 1,809 ECG recordings from 1,062 children aged 6 to 18, annotated by pediatric cardiologists. From each ECG, 255 measurement-based, 720 frequency-domain (FFT), and 732 time-domain (including wavelet-derived) features were extracted. Ten ML algorithms were trained on various feature combinations, with hyperparameters optimized using Bayesian optimization. The best-performing model, CatBoost using only measurement-based features, achieved a specificity of 0.859 at a fixed recall of 0.960, outperforming the conventional algorithm (specificity: 0.605). SHAP (SHapley Additive exPlanations) analysis was used to visualize feature contributions, with T-wave amplitude in lead V4 identified as the most influential. These results demonstrate the potential of ML models to reduce false positives while maintaining diagnostic recall in pediatric ECG screening. Future work should focus on condition-specific explanations and optimization under stricter recall thresholds.