Multiscale Two-Directional Two-Dimensional Principal Component Analysis and Its Application to High-Dimensional Biomedical Signal Classification

Goal: Time–frequency analysis incorporating the wavelet transform followed by the principal component analysis (WT-PCA) has been a powerful approach for the analysis of biomedical signals, such as electromyography (EMG), electroencephalography, electrocardiography, and Doppler ultrasound. Time–frequency coefficients at various scales were usually transformed into a 1-D array using only a single or a few signal channels. The steady improvement of biomedical recording techniques has increasingly permitted the registration of a high number of channels. However, WT-PCA is not applicable to high-dimensional recordings due to the curse of dimensionality and small sample size problem. In this study, we present a multiscale two-directional 2-D principal component analysis method for the efficient and effective extraction of essential feature information from high-dimensional signals. Multiscale matrices constructed in the first step incorporate the spatial correlation and physiological characteristics of subband signals among channels. In the second step, the two-directional 2-D PCA operates on the multiscale matrices to reduce the dimension, rather than vectors in conventional PCA. Results are presented from an experiment to classify 20 hand movements using 89-channel EMG signals recorded in stroke survivors, which illustrates the efficiency and effectiveness of the proposed method for a high-dimensional biomedical signal analysis.