Automatic classification of magnetocardiograms with Machine Learning approach

Magnetocardiographic (MCG) mapping measures magnetic fields generated by the electrophysio-logical activity of the heart. MCG data are usually acquired at 36 locations above the frontal torso, with SQUID-based device. Quantitative analysis of ventricular repolarization (VR) parameters pro-vides information additional to ECG, useful to detect myocardial ischemia in patients with appar-ently normal ECG. However manual calculation of VR parameters is time consuming and can be biased by the examiner¿s experience. Alternatively, the use of machine learning (ML) has been re-cently proposed to automate interpretation of MCG recordings and to minimize human input for the analysis. The aim of this study was to validate the output of ML analysis by comparison with con-ventional computer-aided MCG analysis performed by two cardiologists. Methods: MCG data were acquired at rest, with a 36-channel MCG system (sensitivity of 20 fT/Hz½) above the frontal torso, for 90 seconds (sampling rate of 1 KHz; bandwidth of DC Hz to 100 Hz). Digital low pass filter at 20 Hz was used before ML was applied. To eliminate stochastic noise components, all sig-nals were averaged using the maximum of the R peak as a trigger point. For automatic classifica-tion, data from a time window between the J point and T peak were used. Direct Kernel partial least square (DK-PLS) was the ML tool used, because it was the model providing the best performance in a preliminary preclinical test. The training data consisted of 73 cases correctly classified by ex-perts. ML testing was done on a set of 77 randomly analysed MCG recordings belonging to 25 sub-jects (13 patients with ischemic heart disease (IHD) defined by coronary angiography, and 12 sub-jects without any evidence of cardiac disease at clinical history, physical examination and echocar-diography). For each case at least 2 MCG datasets, recorded in different sessions were analysed. Results: DK-PLS identified abnormal VR in 9 IHD patients and excluded VR abnormalities in 11 controls (69.2% sensitivity, 91% specificity, 90% positive predictive value, 73.3% negative predic-tive value, 80% predictive accuracy). ML classification of MCG was highly reproducible. No dif-ference was found between MCG classification done by cardiologists and with the ML method. Conclusions: This study confirms that DK-PLS ML, applied on MCG recording at rest, has a pre-dictive accuracy of 80% in detecting electrophysiological alterations associated with IHD. Further work is needed to test the ML capability to differentiate VR alterations due to IHD from those due to non-ischemic cardiomyopathies