Wavelet-Based Multi-lead ECG Data Compression and Recognition

• Main Authors: Jein- Ming-Cheng, 簡明成
• Other Authors: Shou-Gang Miaou
• Format: Others
• Language: zh-TW
• Published: 1998
• Online Access: http://ndltd.ncl.edu.tw/handle/57334936949034302405

碩士 === 中原大學 === 電子工程研究所 === 86 === In this thesis, multi-lead ECG data compression and recognition algorithms based on wavelet transforms are proposed. For the compression part, an orthogonal wavelet transform with six-band decomposition is used and the resulting wavelet coefficients in the highest band are dropped. Then the coefficients in each remaining band are vector quantized adaptivelybased on the Gold Washing mechanism. The MIT/BIH arrhythmia database is used in our experiment, where a normal ECG signal from lead II is used for training initial vector codebooks and others ECG signals from lead II and lead V1 in different are used for testing. For the bit rate below 1.59bpp, the resulting percent root-mean-square difference of the reconstructed signal is less than 1.0%. This result is much better than 10.61% (0.909bpp), 2.79% (2.08bpp) and 6.59% (1.12bpp), which are obtained by using basic vector quantization alone, adaptive vector quantization alone, and basic vector quantization in associated with the wavelet transform, respectively. The required computation time is equivalent to that of wavelet transform in associated with basic vector quantization. It can run approximately 2.5 times faster than adaptive vector quantization and 4.5 times faster than basic vector quantization. In addition, by judging the quality of the reconstructed signals, it is found that the proposed compression scheme is nearly lossless and is applicable to variety and different leads of ECG signals at a relatively low bit rate.For the recognition part, a dyadic wavelet transform is used first to extract the features from multi-lead ECG signals that are acquired in three local hospitals. These signals, including leads I, AVL, V1, V2, V5 and V6, reflect the following symptoms: left ventricular hypertrophy (LVH), right ventricular hypertrophy (RVH), left bundle branch block (LBBB), right bundle branch block (RBBB), anterior septal infarction (ASI), anterior lateral infarction (ALI) and posterior infarction (POI). The backpropagation neural network and its fuzzy version are then used for outside recognition testing, resulting in 76.47% and 82.35% average recognition rates respectively.