SVD Based Pattern Analysis for Quasiperiodic Signal Processing -- ECG Data Compression and Noise Reduction

• Main Authors: Jyh-Jong Wei, 魏志中
• Other Authors: Gwo-Jen Jan
• Format: Others
• Language: en_US
• Published: 2002
• Online Access: http://ndltd.ncl.edu.tw/handle/04476274982877606132

博士 === 國立臺灣大學 === 電機工程學研究所 === 90 === The goal of this dissertation is to propose a pattern analysis method using singular value decomposition (SVD) for quasiperiodic signal processing. The SVD based technique is employed as a robust tool for the ECG data compression and noise reduction. The approach of the pattern analysis is exploited as an innovative method to decompose ECG waveforms as a linear combination of a set of essential patterns and the associated weighting vectors. Thanks to the specific feature of cyclic nature and high inter-beat correlations of ECG cycles, SVD especially provides an effective and efficient way to extract the significant information from the original waveforms, and was successfully applied on the applications of ECG data compression and noise reduction. Following the processes of pattern analysis, the significant part of ECG signals can be easily separated from the decomposition processes; the extracted significant patterns with the associated weighting vectors can be retained as the compressed data for representing ECG waveforms, and become the fundamental concept behind the method of ECG data compression used in the present work. Two ECG data compression algorithms are proposed in this paper, one is truncated SVD with error prediction method (TSVD-EP), another one is truncated SVD with residual coding method (TSVD-RC). A set of abnormal ECG signals taken from the MIT-BIH (Massachusetts Institute of Technology - Beth Israel Hospital) arrhythmia database were utilized for evaluating the performances of the two compression algorithms in terms of compression ratio (CR), percent RMS difference (PRD), and visual clinical inspection. The approximate achievement of our proposed methods is presented with an average CR over than 30 with related low reconstruction error. The results base on MIT-BIH database data validate the pattern-based ECG data compression methods using the truncated SVD can provide an efficient way for representing ECG waveforms and effectively remain the critical rhythms for clinical diagnosis. Base on the pattern analysis method, the noise reduction for quasiperiodic signals are also studied in our report. Principle component reconstruction was employed to extract the significant signals from the mutilated waveforms. A set of simulated quasiperiodic signals mixed up with different level of whit background noises was examined to assess the capability of noise canceling using the proposed method. Butterworth FIR filter (N=12) was applied on the same test signals for the comparison of signal-to-noise ratio (SNR) enhanced ability. The proposed method was also applied on the ECG signals that contained high noise interferences for noise reduction. Experimental results show the method significantly upgrade the SNR after principle component reconstruction, suggesting SVD can provide an efficient tool for noise canceling of quasiperiodic signals, and is considered as a robust method on the processing of ECG data compression and noise reduction for clinical applications.