Interpretable machine learning for predicting paroxysmal atrial fibrillation from automatically measured ECG features

Abstract Background The intermittent and often asymptomatic nature of paroxysmal atrial fibrillation (PAF) presents significant diagnostic challenges, since repeated ECG screening frequently returns normal results. However, modern electrocardiographs can automatically extract detailed signal features that may be indicative of underlying PAF. Leveraging these parameters with machine learning offers a promising, non-invasive approach for early identification of at-risk individuals. Aim To develop and evaluate interpretable machine learning models for predicting PAF based solely on automatically extracted ECG features, and to identify clinically meaningful predictors. Method We analyzed 350 digital ECGs, recorded in sinus rhythm, from 126 patients with PAF and 224 control subjects. All ECGs were automatically processed to extract feature parameters. Data were split into training and testing sets using a 5:1 ratio. Two machine learning models —XGBoost and Random Forest— were trained and evaluated using ROC curve analysis. Feature importance was assessed to identify key parameters. Results XGBoost outperformed Random Forest, achieving an AUC of 0.90 compared to 0.82. Feature importance analysis from XGBoost highlighted P-wave amplitudes in leads I, V5, and V6 as significant predictors. Conclusions In a dataset of 350 digitaly recorded ECGs, XGBoost demonstrated strong predictive capability for identifying patients with a history of PAF, based solely on automatically derived ECG parameters. The most influential features corresponded with established P-wave indices, highlighting the potential of interpretable machine learning to enable accurate, scalable, and non-invasive PAF screening using standard digital ECG data.