Automatic classification of magnetocardiograms with the machine learning approach

Magnetocardiographic mapping (MCG) measures magnetic fields generated by the electrophysiological activity of the heart, and is a new imaging technology developed for the rapid, non-invasive detection of ventricular repolarization (VR) abnormalities. MCG data are usually mapped, simultaneously or sequentially, from 33-60 locations above the frontal torso, with SQUID-based device. Previous research has shown that, compared to standard ECG, multichannel MCG provides non-invasive evaluation of cardiac electrogenesis, with similar investigation time, but higher spatial and temporal resolution. The diagnostic potentiality of MCG spreads from three-dimensional electroanatomical localization of arrhythmias, to the identification of VR abnormalities in patients with myocardial ischemia and non-diagnostic ECG (1,2). The analysis of VR from MCG can be done visually and/or quantitatively. Quantitative VR parameters can be calculated from the ST interval and/or the T wave (3-7). Interactive computer-aided analysis of MCG parameters, especially of the ST interval, can be biased by the signal to-noise ratio (SNR) and by examiner’s experience. Therefore automatic analysis procedures would be desirable to speed-up the procedure and to minimize human input. The aim of this study was to validate automatic classification of Magnetocardiograms with the Machine Learning (ML) approach, developed with the NSF SBIR project (Phase I) and described by Embrechts (8), by comparison with computer-aided interactive analysis of MCG, independently performed by two expert cardiologists. As the ST-segment has usually a low SNR, whereas the T-wave is most likely to show primary abnormalities due to ischemia and has a high SNR, ML was applied to the magnetic field data of the T-wave only. This study confirms that MCG can be performed in an unshielded hospital room fully equipped for intensive cardiac care and interventional cardiology, with quality good enough to detect ventricular repolarization abnormalities. Magnetocardiographic imaging is rapid, safe, and accurate for the detection of myocardial ischemia, even in subjects with a normal or non-specific 12-lead ECG. Automatic classification of rest MCG provides quick detection of electrophysiological alterations associated with ischemic heart diseases, with sensitivity ranging between 60 and 70%, specificity of 85 % and predictive accuracy better than 70%, thus better than that of rest ECG. Interestingly such results, which were observed in an unselected population, improved when patients successfully treated with PTCA before MCG were excluded from the statistic evaluation. The predictive accuracy of the ML method was comparable with that obtained with time-consuming interactive computer-aided analysis by expert cardiologists, or even better when patients with successfully treated coronary occlusion were excluded. Further work is deserved to improve the predictive accuracy by incorporating domain knowledge into the machine learning process and to differentiate magnetocardiographic abnormalities due to different cardiac diseases by solving a non-ordinal multi-class classification problem.