- 學位論文
調節型 EEMD 實現對心電圖T波之識別
Towards T-wave recognition of ECG signal with modulated ensemble empirical mode decomposition
摘要
1998年黃鄂院士提出一種新型態分解法則，經驗模態分解法(Empirical mode decomposition, EMD)，此法具備非線性及非穩態訊號的拆解能力，因而廣泛應用在生醫訊號分析上。特別是體表電位、血液動力學等訊號特徵之拆解應用。然而，當產生間歇性訊號時，EMD方法易出現混模問題(Mode mixing problem)，為此，總體經驗模態分解法(Ensemble EMD, EEMD)採加入雜訊方式來克服，但卻讓雜訊混合在原訊號內，導致遺留在拆解結果中，徒增生醫訊號特徵判斷之困擾。 在此研究中，提出一種調節型總體經驗模態分解法 (Modulated EEMD, mEEMD)，此法類似於EEMD而加入雜訊，但只雜訊加進原始訊號內作為尋找端點的參考點，而未加入接續的拆解動作。如此作法仍可避免混模問題外，拆解後特徵亦只包含原始訊號。為驗證效能，使用2009年Dr. Wu所發表文章中測試訊號—兩個訊號混疊的模擬訊號。結果顯示：mEEMD可拆解出兩個原始訊號(相關係數分別達到0.999及0.968，方均根差為0.125及0.195)。另EMD與EEMD等結果比較，mEEMD法之結果有較高的相關性係數、較低的方均根差。計算時間與EEMD法相近。另外，值得注意之處，mEEMD法因參考點緣故，可有效地抑制邊界效應(Boundary side effect)，對於接續特徵判定上有所幫助。 考量ECG訊號可呈現心血管系統之機轉，例如，T波代表心室再極化的過程，高尖、倒轉或相位改變皆可呈現相對應的心血管疾病。目前ECG訊號解構多採用小波分析法，或確認R波位置後再接續尋找其他特徵。然而，小波分析法會侷限於母小波的挑選限制，無法適用於不同階段、不同導程ECG訊號；而R波後搜尋T波時易受雜訊影響，常需先去噪處理。本研究嘗試mEEMD法拆解PysioNet QT資料庫的ECG訊號，結果顯示，拆解後的T波與原特徵相符，且不會受雜訊影響、無須事前濾波。接續將以mEEMD法應用在不同階段 ECG訊號的T波特徵拆解與判定。
並列摘要
Empirical mode decomposition (EMD), proposed by Huang et al. (1998), can be used to decompose nonlinear and nonstationary signals. Therefore, EMD is widely used on nonlinear and nonstationary biomedical signals, especially those relating to surface potential and hemodynamics. However, when a signal has intermittent patterns, the use of EMD potentially leads to a mode mixing problem. Hence, other EMD-based algorithms, such as ensemble EMD (EEMD), have been proposed. To solve mode mixing problems, these algorithms add Gaussian white noise to signals. However, such noise may remain in the decomposition result, making the detection of biomedical signal features more difficult. This study proposes modulated EEMD (mEEMD) as a solution. Rather than adding noise to a signal, mEEMD uses Gaussian white noise to assist the search for reference points. Thus, mEEMD can solve mode mixing problems with almost no influence from noise. The boundary side effect is another problem in EMD. This study’s mEEMD solution was demonstrated to have a less problematic boundary side effect. To verify the adequacy of mEEMD, this study modified and used the simulated signal formulated in Wu et al. (2009). In the analysis results, the correlation coefficients between the decomposition result and two source signals were as high as 0.999 and 0.968, respectively, and the root-mean-square errors were 0.125 and 0.195, respectively. This study also compared mEEMD against EMD, EEMD, and complementary EEMD. In the results, mEEMD had higher correlation coefficients for this test than the other three methods did. The root-mean-square errors of mEEMD were lower than those of the other three methods. Moreover, mEEMD took almost as much time to complete the test as EEMD did. Electrocardiography (ECG) signals can track the operation of the heart and circulatory system. T-waves, in particular, are a type of waveform in an ECG, and they track ventricular repolarization. The occurrences of hyperkalemia T-waves, T-wave inversion, or T-wave alternans may indicate the presence of a heart disease. Because of ECG’s characteristics, traditionally, the decomposition and detection of T-waves have usually been used for wavelets or tracing after the R-peak. However, wavelet analysis has tended to be limited to the mother wavelet being selected. Because the backward search from the R-peak is easily influenced by noise, the signal must be denoised first. This study also used mEEMD to decompose ECG signals from the PhysioNet QT database. The results indicate favorable performance in T-wave waveform decomposition, with no influence from the power line noise. Filtering is not required prior to processing. The mEEMD method can thus be used to decompose and determine the T-wave characteristics of ECG signals.