A High Efficient ECG Data Compression System Based On Reversible Round-Off Discrete Wavelet Transform

[[abstract]]心電圖(Electrocardiogram; ECG)提供一種非侵入式檢測方式，廣泛應用於心臟疾病之臨床診斷。心臟病患的照護與即時急救必須仰賴長期的心電圖紀錄與即時資料傳輸，心電圖即時資料壓縮技術即為因應此需求而發展。有別於一般的影音資料壓縮，醫學應用對重建心電圖的品質有極嚴格的要求。為穩定心電圖的重建品質，適應性壓縮機制是心電圖壓縮系統設計的一個必備功能。 由於系統簡單，基於二值化處理(threshold-based)的傳統心電圖壓縮方式仍為目前可滿足即時應用需求的最佳選擇，其中以SPIHT [5] 壓縮法最好。SPIHT是一種基於小波轉換(wavelet-base)的壓縮編碼技術，其具最佳壓縮效能、即時編碼能力、及可硬體實現等優點。然傳統小波轉換採用遞迴式分解演算法存在誤差傳播(error propagation)及位元成長(word-length-growth)效應，此兩效應是導致重建訊號品質在壓縮率增加時呈指數衰減現象的主要因素。為克服此兩效應，本論文提出一可逆式四捨五入非遞迴式小波轉換法，此法結合可逆式四捨五入線性轉換(RROLT)理論和一維非遞迴式離散週期性小波轉換(1-D NRDPWT)。NRDPWT的目的是用於阻斷誤差傳播，而RROLT理論則可提供具最小動態區間(MDR)之小波轉換係數。此外RROLT理論亦提供誤差正規化機制，可有效縮小量化值搜索區間。基於此誤差正規化機制，本論文提出一可用單一量化參數(QF)控制之非線性量化演算法。此演算法的設計規格是要維持失真度與壓縮比率(CR)間呈線性關係。此線性特性不僅可有效提升整體壓縮效能，亦提供建立失真度線性評估模式(linear prediction model)之基礎。本論文使用MIT-BIH之心率不整心電圖資料庫做為測試訊號源並採用三種效能評估方法：均方根誤差百分比(PRD)，最大誤差振幅(MAE)，及重建訊號的視覺觀測法。實驗結果顯示，本系統可獲得比SPIHT壓縮方法更佳的壓縮效能；在 及 兩區間，其PRD改進百分比分別達18.65%及18.39%。 適應性壓縮機制的目的是要使得每一區段心電圖的重建品質皆能満足隨選預設值，此隨選預設值通常採用PRD值。由於PRD值的計算是使用時域訊號，因此傳統系統的適應性壓縮機制通常需要相當複雜的運算過程。以SPIHT壓縮法為例，其所需之計算複雜度為 [4]，其中n為平均重覆預測次數。基於線性的PRD-CR曲線之規劃，本論文提出一線性誤差預測模式及動態量化演算法。此適應性壓縮機制的主要特色是使用頻域訊號預測PRD值，因此其所需之計算複雜度僅為 且m遠小於n。[[abstract]]Electrocardiogram (ECG) signal analysis is a non-invasive modality widely used for heart disease clinical diagnosis. The diagnosis of cardiac patients depends on the quality of the recording signal or effective real-time data transmission. The quality of reconstructed ECG signals is really critical. For keeping a desired quality of reconstructed ECG signal, the adaptive compression mechanism is a required function for ECG compression systems. Because of implementation simplicity, the traditional threshold-based ECG compression methods are good choices for real-time application. Among these traditional methods, the SPIHT [5] has the best compression performance. SPIHT is a wavelet-based coding technique and SPIHT has advantages of optimum compression performance, real-time coding, and hardware implementation. However, traditional wavelet-based ECG data compression methods utilize recursive decomposition algorithm. This recursive decomposition suffers from the error propagation and word-length-growth effects which are the main cause of the quality deterioration. To solve these two problems, a reversible round-off non-recursive wavelet transform is proposed in this dissertation. This reversible round-off non-recursive wavelet transform incorporates a reversible round-off linear transformation (RROLT) theorem and a non-recursive 1-D discrete periodized wavelet transform (1-D NRDPWT). NRDPWT can stop error propagation, and RROLT theorem can obtain the reversible transformed coefficients with minimum dynamic range (MDR). And RROLT theorem with an error normalization mechanism can reduce the search space of the optimum quantization scales efficiently. In this dissertation, we proposed a non-linear quantization algorithm by adjusting the single quantization factor (QF) based on error normalization mechanism of RROLT. The goal of this algorithm expects to maintain a linear relation between distortion and compression ratio (CR). This linear relationship is not only improving the compression performance of the whole system, but also building a linear distortion prediction model. In this dissertation, the MIT-BIH arrhythmia database is used for testing signals, and the performance evaluation is based on the percentage root-mean-square difference (PRD), the maximum amplitude error (MAE), and visual inspection of the reconstructed signals. Experimental results show that our proposed method can obtain a superior compression performance over SPIHT; when , the PRD is improved by 18.65% and when , the PRD is improved by 18.39% . The goal of adaptive compression scheme is to guarantee the reconstructed signal with the desired quality for each segment. The PRD is commonly adopted for quality evaluation. Since the PRD is based on time-domain calculation, the traditional adaptive compression scheme has really high computational complexity. For example, computational complexity of SPIHT is [4], where n denotes the average iteration of PRD prediction. In this dissertation, a linear error predetermination model and a dynamic quantization algorithm are proposed based on the programming of the linear relationship between PRD and CR. The main characteristic of this adaptive compression scheme is based on frequency-domain signal for PRD prediction. The computational complexity is just only , and m is smaller than n