Novel electrocardiogram compression and classification algorithms for wearable healthcare system

Abstract

Wearable healthcare systems (WHS) have received much attention recently for their valuable roles in prevention-oriented and personalized healthcare. Electrocardiography (ECG) is one of the most common biomedical signals employed for personal health monitoring. Key issues in ECG-based WHS are artifact removal, efficient compression for storage and transmission, and automatic ECG classification. New methods are developed in this thesis to address these issues. First, a new artifact removal algorithm based on local polynomial regression (LPR) with model averaged and refined variable bandwidth selection (LPR-MA-RVBS) is proposed to remove baseline wandering (BW) and Gaussian noise (GN) in ECG signals. An improved model averaged and refined variable bandwidth selection (MA-RVBS) method is proposed to address the bandwidth selection problem in LPR for estimating BW possibly with discontinuity and suppressing GN. A low-complexity multiplier-less fixed-point implementation of LPR is developed, where constant multiplications are realized as limited additions and shifts. Experimental results show that both BW and GN can be well removed by the proposed artifact removal algorithm. Moreover, the proposed multiplier-less implementation greatly reduces the computational complexity as well as run-time of the proposed algorithm in software implementation. Secondly, we propose a novel local principal component analysis (LPCA) scheme with time-and-level alignment (TLA) and residual coding (RC) for compression of ECG signals. It uses a new TLA algorithm to address the variations in period and level of ECG cycles and LPCA to better model the resulting waveforms which may vary with different body states. It therefore overcomes the sensitivity issue of conventional PCA-based compression algorithms to alignment quality and changes in body states. Moreover, a robust statistics-based method is used for the detection of outlying ECG cycles and a discrete-cosine transform (DCT)-based coding method is developed to compress such outlying ECG cycles. Methods for coding the principal component (PC) scores of the normal ECG cycles and other parameters are also developed. Experimental results show that it can achieve better performance than other state-of-the-art methods in terms of compression ratio, reconstruction error and quality score. Finally, a novel deep feature learning approach and its efficient multiplier-less implementation is proposed for automatic ECG classification for arrhythmia and myocardial infarction (MI). It employs a new adaptive multiple-descriptor decomposition (AMDD) method, with modest compression to minimize information loss, to assist automatic ECG feature extraction in frequency domain. A multi-stage sparse auto-encoder (SAE) with activation thresholding is proposed to learn high-level feature representation, which helps to remove more than half of the weakly expressed neurons at the same classification accuracy. A multiplier-less implementation of the classifier is developed where weight multiplications of the deep learning-based classifier are realized as limited shifts and additions. A new grouped multiplier block (GMB) approach is also proposed to further reduce the additions required by reusing the intermediate partial sums. Experimental results show that the proposed approach gives better performance than other state-of-the-art approaches in terms of sensitivity, specificity, and accuracy. Moreover, the proposed fixed-point approach offers efficient multiplier-less implementation for hardware realization as well as reduced run-time in software implementation.