為達到即時辨識ECG輔助診斷之目的，利用自組織映射圖(Self Organizing Map,SOM)類神經網路模式，藉由自動時域心電圖特徵值擷取 方法，製作檢測心電圖的工具。當SOM網路經心電圖訓練範例訓練完成時 ，網路輸出層結點形成病狀聚類區之拓撲座標，結合拓撲座標上各聚類中 心與特徵值向量連結之權值，構成最小尤拉(Euler)距離方程式，如此可 以標出不正常之心電圖信號。心電圖是診斷心臟疾病最主要的工具之一， 一般病患的心電圖變異常發生於某幾個單一或連續的脈波波形的改變。所 以對病患作全天候心電圖信號之檢測，如霍特(Holter)ECG記錄器、床邊 監視器(Bedside monitor)、運動(exercise)心電圖等儀器，以便能夠在 病變發生時，產生急救的警訊。檢測之方法，利用自動時域心電圖特徵值 擷取方法，擷取 MIT/BIH資料庫導程II心電圖之特徵值，取得訓練範 例1150組，經特徵值正規化後，再輸入自組織映射圖網路進行訓練。當網 路收斂時，輸出層結點被訓練範例特徵值向量映射成為聚類區，由特徵值 向量與各聚類中心所連結之權值，構成病症分類方程式，再將已正規化之 測試範例923組，輸入已完成訓練之網路進行辨識測試。並利用網路輸出 拓撲座標，進行病症轉移之實驗，獲得各病症轉移之主要特徵值參數。本 研究亦探討影響網路辨識能力之因素，如特徵值正規化、網路鄰近半徑參 數、拓撲座標數目等。本研究針對四種病狀：早期心室收縮(PVC)、融合 性早期心室收縮(FUS PVC)、右束枝傳導阻滯(RBBB)、左束枝傳導阻滯( LBBB)及正常（NOR）等五種心電圖波形，擷取十二個特徵值，配合網路架 構採用20×20＝400個輸出處理單元，鄰近半徑等於20，學習循環八十次 ，學習常數等於0.5，供網路學習後進行辨識，結果辨識MIT/BIH資料庫之 訓練範例、測試範例及結合心電圖儀作即時正常心電圖辨識，其正確辨識 率均達百分之九十八以上，而且測試14位正常參與者心電圖可以檢測出 PVC病狀。另外，由病症轉移實驗可化簡構成正常方程式為八個主要特徵 值，如此可以快速經由此正常方程式檢測出不正常之心電圖信號。 ; For the purpose of real time ECG diagnostic, the paper discusses the method to recognize ECG pattern. The ECG is one of the main cardiac diagnosis tools. The abnormal ECG in the sequential Heart beat is the most difficult to identify, e.g. the data of Holter ECG recording, bedside monitor, or exercise ECG recording. When the training of SOM neural network finished, the output layer would classify pattern. The weighting values of the features vector and the center node would be fixed. We combined the weighting values with features vector using Euler distance equation to pick out the abnormal ECG signal. Using MIT/BIH arrhythmia database. The total of 1150 training samples has been used as training groups. The total of 923 samples has been used as to test the method. The feature of lead II ECG has been automatic extracted. A normalized data is fed to SOM neural network for pattern classification. From the classified pattern of SOM, the weighting value of the center node and features vector are forming a pattern distance equation(minimum Euler distance equation). Then, put in all training samples to normal equation that a interval range of NOR value can be obtained. In this thesis, five different ECG patterns has been tested which are normal(NOR), premature ventricular contraction(PVC), fusion premature ventricular contraction (FUS PVC), right bundle branch block(RBBB) and left bundle branch block(LBBB). To test the method, the study extracted 12 features of ECG. The SOM uses 400 processing elements. The neighborhood radius is 20. The learning cycle is 80 iterations. The learning rate coefficient is 0.5. Using ECG pattern in MIT/BIH database, the system demonstrates more than 98 percent correct classification. For the purpose of labeling the abnormal ECG, the equation can be further reduced to only 8 dominant features. Therefore, the method can be implemented into real time process to screen the ECG data. The method has been tested in screen 14 adult subjects to real ; 083CYCU0530001