Machine learning in diagnosing coronary artery disease via optical pumped magnetometer magnetocardiography: a prospective cohort study

<i>Objective.</i>The potential of optical pumped magnetometer magnetocardiography (OPM-MCG) for diagnosing coronary artery disease (CAD) has been initially shown, yet lacks large-scale prospective research.<i>Approach.</i>Using invasive coronary angiography (ICA) as a reference, we constructed three feature sets for the development of machine learning (ML) models: a 'Heart' feature set consisting only of OPM-MCG features, a 'Clinical' feature set, and a 'Heart + Clinical' combined feature set. We assessed the performance of 11 ML models with 10-fold cross-validation and conducted a feature importance analysis.<i>Main results and Significance</i>. Among 1513 participants (mean age 58.2 ± 12.0 years, 75.5% male), 1194 (78.92%) tested positive for ICA. Significant differences were observed in 'Heart' and 'Clinical' features between ICA-positive and negative groups. ML models using only 'Heart' features (AUC 0.84-0.88) outperformed those using only 'Clinical' features (AUC 0.62-0.75). Combining both feature types improved diagnostic accuracy (AUC 0.75-0.90). Feature importance analysis highlighted that 'Significant change of Ar-PN' in OPM-MCG was key for ICA diagnosis (47.8%), along with 'Abnormal Sp-TT', 'Significant change of Ps-PN', and 'Abnormal Mg-TT'. OPM-MCG has high performance in diagnosing CAD, and the significant change of Ar-PN is the most important feature. Cat Boost and random forest are more suitable for OPM-MCG to build ML diagnostic models for CAD.