本研究提出使用區域化特徵和校正策略來建立辨識持續性心肌缺血心搏和正常心搏的系統,。而所謂的心肌缺血心搏,即是指觀察心電圖波形中的ST段與和正常心搏時相比,下降的電壓值到達一定程度時,則為心肌缺血心搏。而持續性心肌缺血事件,則是指,心肌缺血心搏持續達到一定時間,稱為心肌缺血事件發生。 提出的方法分為兩個部分,第一個部分為使用最直接的方式,以ST段的下降量來直接辨識心肌缺血心搏,此部分也可結合校正策略來做正確率提升。第二個部分為使用分類器辨識,使用到的分類器有兩種,支持向量機(SVM)和倒傳遞類神經網路(BPNN)。而在特徵擷取中,提出區域化特徵的方式來降低一些變化性較大的心搏所造成的影響,最後,再加上區域化校正策略來針對事件的發生去作校正,可以提升整體的辨識率。在分類器辨識中,分為兩個實驗方式,第一個實驗方式是使用排除個體差異性的方式去做系統測試,也就是使用10-fold cross-validation的方式去做驗證。第二個方式則是留一驗證法(leave-one-out cross-validation),將個別測試資料不去做訓練的測試方式。 結果顯示,直接辨識法的平均辨識率可達94.54%。而在分類器辨識方面,在排除個體差異性的情況下,平均辨識率有99.20%。而不排除個體差異性,也就是使用留一驗證法,平均辨識率可達97.51%。以上為對心搏的辨識率,而在對事件的辨識上,直接辨識法和分類器辨識可將本研究所使用的測試資料事件都辨識出來,而在事件上,心搏的覆蓋率分別為93.36%和94.48%。
This study proposes to use segmental features and correction method to build a system to recognize myocardial ischemia heart beats from normal heart beats. Myocardial ischemia heart beats usually show means the voltage drops in the ST segments as compared to the ECG waveforms of normal heart beats. Sustained myocardial ischemic events refer to the continuing of myocardial ischemia heart beats up to a certain time period. The proposed method is divided into two parts. The first part is the direct identification method, which can be further enhance by the correction method. The second part is method using classifiers. Two classifiers, support vector machine (SVM) and back-propagation neural network (BPNN) was used, we proposed to use segmental features to reduce the effect of large heart beat variability shape in the feature extraction. We also used segment correction strategy to improve event outcomes. Classifier validation methods were employed, including a 10-fold cross-validation and a leave-one-out cross-validation. The accuracy of applying the direct method is 94.54%. The accuracy of using the classifier is 99.20% in 10-fold cross-validation and 97.51% in using leave-one-out cross-validation. In recognition of the event, both the direct identification method and the classifier can identify all events. Among them, myocardial ischemia heart beats coverage is 93.36% when using the direct identification method and 94.48% using the classifier method.