Dimensional contraction by principal component analysis as preprocessing for independent component analysis at MCG

Abstract

We propose a noise reduction method for magnetocardiograms (MCGs) based on independent component analysis (ICA). ICA is useful to separate the noise and signal components, but ICA-based automatic noise reduction faces two main difficulties: the dimensional contraction process applied after the principal component analysis (PCA) used for preprocessing, and the component selection applied after ICA. The results of noise reduction vary among people, because these two processes typically depend on personal qualitative evaluations of the obtained components. Therefore, automatic quantitative ICA-based noise reduction is highly desirable. We will focus on the first difficulty, by improving the index used in the dimensional contraction process. The index used for component ordering after PCA affects the accuracy of separation obtained with ICA. The contribution ratio is often used as an index. However, its efficacy is highly dependent on the signal-to-noise ratio (SNR) it unsuitable for automation. We propose a kurtosis-based index, whose efficacy does not depend on SNR. We compare the two decision indexes through simulation. First, we evaluate their preservation rate of the MCG information after dimensional contraction. In addition, we evaluate their effect on the accuracy of the ICA-based noise reduction method. The obtained results show that the kurtosis-based index does preserve the MCG signal information through dimensional contraction, and has a more consistent behavior when the number of components increases. The proposed index performs better than the traditional index, especially in low SNRs. As such, it paves the way for the desired noise reduction process automation.