[[alternative]]The Exploration for Postoperative Complications of Abdominal Aortic Aneurysm by Using Data Mining Classification Models

[[abstract]]論文摘要目前醫生對腹主動脈瘤(abdominal aortic aneurysm)手術併發症，大部分是根據病人的生理檢測數據，加上個人經驗及相關文獻知識來判斷病人可能於術後發生的併發症，但單以醫生個人經驗缺乏客觀判斷，尤其在具有高複雜度及不確定性問題的情況下；醫生的判斷可能無法如預期般的好。因此，本研究希望藉著資訊科技的力量，建置一套腹主動脈瘤術後併發症分類器，讓醫師能再多一項輔助工具，以改善醫師對術後病人可發生的併發症分析。由於記錄主動脈瘤手術資料庫的資料欄位非常多，共有137個資料欄位，導致醫生在判斷病人發生併發症的風險上過於複雜，因此，本研究以資料探勘方式，並以臺北市立榮民總醫院腹主動脈瘤病人的實際個案(包括：病歷資料、生理檢測資料、病人基本資料、手術資料)作為主要資料來源，進行資料前置處理(包括欄位及個案的過濾、資料型態轉換)，並透過裝袋法(Bagging)訓練三個不同的人工類神經網路(Neural Network)分類器，並以十摺交互驗證來穩定分類器，再使用接受器操作曲線(ROC curve)來評估分類器，最後使用C5.0決策樹找出三大類併發症的分類規則及重要因子，及利用貝氏網路，找出重要因子間的因果關係並計算出其聯合條件機率，藉此來提供醫療人員瞭解病人可能發生的併發症，並及早做出預防發生併發症的相關醫療措施，與防範併發症的發生，維護病人健康及提昇整體醫療品質。[[abstract]]AbstractCurrently, the diagnosis of postoperative complications of abdominal aortic aneurysm can only depend upon the patient’s biomedical examine data, physician’s experience and literatures related to abdominal aortic aneurysm at present, but this method is subjective to patient’s diagnosis. Particularly under uncertainty environment or complicated cases, physicians can identify the conditions for patients with personal experience, but the result may not meet final expectation. Therefore, this study intends to develop a data mining processes to assist cardiovascular surgeons diagnose and analyze of patients’ conditions. Due to there are 137 fields in the abdominal aortic aneurysm special chart, this is why the cardiovascular surgeons are difficult to determine what complications would occur after surgical operation. The real dataset for this study was collected from a medical center, Taipei Veterans general hospital, including patients’ historical medical data, biomedical data, fundamental data, and operative data etc. The data mining processes included data preprocessing (filtering and converting of data), and then we used “bagging” learning scheme to train 3 different ANN (artificial neural network) classifiers. The “10-fold cross validation” and “ensemble classifier” was used to enhance the stability of the final classifiers, the ROC curve was employed to estimate the performance of classifiers. Finally, for easy understanding and interpretation of the extracted knowledge, this study employed C5.0 classification rule to express the possible reasons for the arising of three different postoperative complications. In order to find out the causal relationship among features and complications, we used Bayesian Network as a computation tool. We hope the built ensemble classifier and the classification rules could provide more information for cardiovascular surgeons to understand the possibility of arising postoperative complications, and could do some works in advance to prevent the occur of complications after operation, and improve the medical quality for abdominal aortic aneurysm and patients’ life quality