A Signal Processing Approach to Post-ACS Patients Risk Stratification Using ECG

急性冠心症肇因於心臟肌肉缺血或缺氧，是長久以來位居我國十大死因前三位的嚴重疾病。是否有一套判斷ACS病患預後程度的機制是非常重要的研究議題。傳統上，TIMI危險分數採用了病患的病史、藥物使用、生物標記、年齡等生理資料作為判斷之依據，是目前較廣被接受的作法。心電圖亦為一種專門用來檢視心臟狀況的方法。近年來，由於心電圖之快速性、方便性、以及非侵入性，漸多以心電圖做ACS病患預後的研究出現。 本論文提出一完整之心電圖分析系統，包含心電圖訊號的前處理、心跳偵測、特徵擷取，以及最後使用人工智慧判斷病患之預後狀況，於系統中每一步驟本論文均有提出改進效能之方法。此系統之設計理念在於能夠實時分析心電圖訊號而非蒐集完訊號後才一次處理，故系統於前處理部份提出一實時運算之集成經驗模態分解法；利用人工智慧協助ACS後病患之危險評估也能達到比人為判斷或用統計模型所得出之預後效果還要好的危險程度判斷效果。傳統上由於使用人為判斷或多變數回歸分析的方式做預後判斷，特徵必須足夠精確才會有顯著之危險分層效果；使用人工智慧的好處在於，即使特徵是由複數個生理因素所影響，在取用足夠多個特徵的狀況下機器仍能自動計算出特徵內的複數個生理因素，並做出危險判斷。換言之，傳統之分析方法需人工去蕪存菁，而人工智慧之方法則是機器自動做計算。本論文之實驗證實本系統所使用之簡單特徵搭配上類神經網路可以得到極佳之危險評估效能。最後，本論文亦提出一決策步驟，提供臨床醫生一簡單判斷方法，即使沒有人工智慧之儀器，仍可得到不錯的預後準確度。Acute coronary syndrome (ACS), caused by rupture of an atherosclerotic plaque and partial or complete thrombosis of the infarct-related artery, has been on the first three places of the top ten causes of death in Taiwan for many years. It is of concern whether these patients have grave prognosis under current medical treatment. Traditionally, TIMI score remained the most popular method for healthcare professionals, which uses information like age, aspirin use, cardiac biomarker as scores to evaluated clinical outcome after ACS. Electrocardiogram (ECG) is also a widely-used instrument for patients with cardiovascular disease. Recently, research focus on whether we could identify high risk patients through ECG reorganization due to it is quick, noninvasive, and easy to use. An ECG analysis system, including preprocessing, beat detection, feature extraction, and using machine learning to make prediction, is proposed. Each part in this system is enhanced by new algorithms or techniques. The design concept of this system is to provide real-time ECG analysis, so a real-time ensemble empirical mode decomposition (EEMD) is proposed. Furthermore, using machine learning to make prediction achieves better stratification outcomes than that made by using inspection or statistics. Machine learning can find the information hidden in features, even if the used features are affected by multiple medical factors, while users have to use much more sophisticated features when they make prognosis by inspection or multivariate statistical analysis. On the other words, traditionally users extract specific information by themselves, while machine learning does it automatically. A set of new features are proposed and proved work via artificial neural network. Finally, we also give simple decision steps so that experts can easily adopt even if they do not use machine learning