基於心電圖辨識心室頻脈與心室顫動及其用於預測心因性猝死之研究

Identification of Ventricular Tachycardia and Ventricular Fibrillation for Sudden Cardiac Death Prediction Based on Electrocardiogram

蔡柏鋐(BO-HUNG TSAI)

指導教授 ： 余松年

國立中正大學/工學院/電機工程學系/碩士(2016年)

摘要

本論文提出了一個基於心電圖的心室頻脈(Ventricular Tachycardia, VT)及心室顫動(Ventricular Fibrillation, VF)的辨識方法並探討運用於心因性猝死(Sudden Cardiac Death, SCD)預測的研究。VT、VF的辨識與SCD的預測分別使用5秒及60秒的心電圖(ECG)作為訊號來源，其中SCD預測是以VF發病前固定區段的時間進行研究及分析。
本研究主要分為兩部分，第一部分為VT、VF之心律不整辨識，訊號來源為MIT-BIH Malignant Ventricular Arrhythmia Database (VFDB)，第二部分為SCD預測，訊號來源為MIT-BIH Sudden Cardiac Death Holter Database (SDDB)。在系統架構上可分為前處理、特徵擷取、特徵正規化、特徵挑選與分類。首先，先將原始訊號經過前處理去除不必要的雜訊，然後進行特徵擷取及特徵正規化來取得可用以分類的特徵，接著探討傳統基因演算法(GA)、改良型基因演算法(MGA)、多目標基因演算法(NSGA-II)及P值法(P-value)進行特徵挑選，以降低特徵維度時，提升辨識及預測正確率的效能。論文中並探討如何讓挑選出的特徵同時適用於訓練集與測試集的方法，其中分類使用支持向量機(SVM)，並藉由五折(five-fold)的方式來進行交叉驗證。
結果顯示，第一部分嚴重心律不整辨識在訓練集得到的最佳靈敏度、特異度及正確率分別為93.05%、96.84%及95.6%。而在測試集則為92.55%、96.43%及95.52%。第二部分心因性猝死預測，於發病前第10分鐘的訊號在訓練集得到的最佳靈敏度(Sensitivity, Se)、特異度(Specificity, Sp)及正確率(Accuracy, Ac)則分別為76.25%、82.5%及79.38%。而在測試集則為72.5%、82.5%及77.5%。
本系統對於VT、VF的辨識以及SCD預測上皆有不錯的正確率，研究成果可應用在醫療救護上，作為醫護人員掌握病患病情的一項重要輔助。

關鍵字

心因性猝死 ； 心電圖 ； 基因演算法 ； 多目標基因演算法 ； P值法 ； 支持向量機

並列摘要

This thesis proposes a Ventricular Tachycardia (VT) and Ventricular Fibrillation (VF) identification method and its application to Sudden Cardiac Death (SCD) prediction based on electrocardiogram (ECG). These two systems used ECG of 5 seconds and 60 seconds to separately identify VT, VF and predict SCD. The research of SCD prediction were based on signals in specific period of time before the onset of VF.
This study is divided into two parts. The first part is arrhythmia (VT and VF) identification. The ECG signals used in the experiment were obtained from the MIT-BIH Malignant Ventricular Arrhythmia Database (VFDB). The second part is SCD prediction. The ECG signals were obtained from the MIT-BIH Sudden Cardiac Death Holter Database (SDDB). The system architecture was divided into pre-processing, feature extraction, feature normalization, feature selection and classification. Firstly, the original signals were preprocessed to remove unnecessary noise. Secondly, feature extraction and feature normalization were used to obtain the features for classification. Thirdly, traditional genetic algorithm (GA), modified genetic algorithm (MGA), multi-objective genetic algorithm (NSGA-II) and P value were used for feature selection to reduce the feature dimensions. Their performance in improving the identification and prediction accuracy rate was compared. Strategies for selecting features that are suitable for using both in the training set and test set were discussed. Finally, support vector machine (SVM) was used to classify these heart rhythms, and the five-fold scheme was used for cross-validation.
In the first part, a sensitivity (Se) of 93.05%, a specificity (Sp) of 96.84%, and an accuracy (Ac) of 95.6% were obtained on the training dataset. And, an Se of 92.55%, an Sp of 96.43%, and an Ac of 95.52% were obtained on the test dataset.
In the second part, this study can predict the SCD ten minutes before its onset with an Se of 76.25%, an Sp of 82.5%, and an Ac of 79.38% on the training dataset, and an Se of 72.5%, an Sp of 82.5%, and an Ac of 77.5% on the test dataset. According to the results, the systems for arrhythmia identification and SCD prediction proposed in this study imposing accuracy rates, and demonstrate their capability to be used in medical care to assist health care workers handle their patient’s condition.

並列關鍵字

Sudden Cardiac Death ； Electrocardiogram ； Genetic algorithm ； Multi-objective genetic algorithm ； P value ； Support Vector Machine

參考文獻

動控制器設計”，國立中山大學機械工程研究所，90年

[14] S. G. Mallat, “A Theory for Multiresolution Signal Decomposition: The Wavelet

Representation”, IEEE Transactionson on Pattern Analysis and Machine Intelligence , Vol. 11, No. 7, 1989, Page(s): 674-693.

[15] V. Vapnik, “Statistical Learning Theory”, New York: Wiley, 1998.

[17] “LIBSVM-A Library for Support Vector Machines”