建構肝臟移植術後急性排斥之評分系統

Abstract

肝臟移植是嚴重肝病患者的最後一線希望，但是手術的成功與否取決於病患的預後狀況，其中最常遇到的問題就是病患對新移植的肝臟產生急性排斥反應，急性排斥反應不僅可能導致肝衰竭，甚至會直接影響病患的生命安全。目前切片檢查是臨床上判斷急性排斥反應最準確的方式，然而，切片檢查必須要對肝臟進行穿刺取樣，如果病患有凝血功能障礙之問題，將會導致嚴重後果。因此，目前病患術後產生急性排斥之風險判斷還是需要仰賴醫療人員的專業知識與經驗，使得醫療人員的心理負荷大幅增加。

本研究使用隨機森林演算法進行特徵選取，從眾多常態測量的血液檢驗項目中挑選出重要的項目，並透過羅吉斯迴歸演算法進行建模，最後使用模型中的各項係數建立評分系統，將檢驗數值轉換為產生急性排斥反應的風險分數。實驗結果顯示，此預測模型的AUC為0.834，擁有相當準確的預測能力，同時經評分系統轉換後的風險分數，能在產生急性排斥反應的四天前將病患區分開來。未來希望可以實際應用於臨床診斷上，協助醫療人員進行相關決策。

Liver transplant is the last flicker of hope to the patients with severe liver disease. However, the outcome of surgery is dependent on the patient's prognosis. Acute allograft rejection is a common problem in postoperative care of liver transplantation which may cause liver failure or even endanger the patient's life. Currently, most medical staffs assess the risk of acute rejection clinically with their knowledge and experience. This thesis proposes a new scoring system to give a risk score based on patient's physiological measurement values, in which we consider explanatory power and prediction accuracy simultaneously. We use Random Forest to select important features from regularly tested physiological measurement items, and then builds a prediction model with logistic regression. Finally, we use the coefficients generated from the logistic model to build a scoring system. The experimental result indicates that the predictive ability of the prediction model proposed in this study is great (AUC = 0.834). Besides, the scoring system proposed in this thesis can effectively separate the non-acute-rejection patients and the acute-rejection patients four days before the rejection occurred.