Machine Learning-Based Automated Method for Effective De-noising of Magnetocardiography Signals Using Independent Component Analysis

This study aims to develop an automated method for de-noising cardiac signals using independent component analysis (ICA) on a 37-channel magnetocardiography (MCG) system. The traditional approach of applying ICA involves manual visual inspection to determine the retention or removal of independent component (IC) related to signal or noise, which is time-consuming and lacks assurance in preserving essential attributes of signal components during the de-noising process. To address these challenges, we propose a novel approach. A feature set comprising spectral, statistical, and nonlinear time series properties is computed from the ICs of thirty subjects. These features are then evaluated by a few machine learning (ML) models to optimally select ICs for de-noising cardiac time series. It is found that ICs evaluated by a gradient boosting decision tree (GBDT) classifier could accomplish the task of efficiently selecting components to de-noise MCG with an accuracy of 93%. The performance of the proposed method is qualitatively and quantitatively compared against conventional methods for noise elimination and preserving signal features. The proposed method has extensive application in de-noising multichannel MCG signals where the characteristics of the noise are not clearly known and for routine diagnostic assessments of subjects with cardiac anomalies in hospital settings.