建構即時性肝臟移植病患術後急性排斥反應預警模型

Abstract

肝臟移植病患手術後最容易發生身體的免疫系統對新移植的肝臟產生排斥反應，嚴重的排斥反應除了可能造成肝功能衰竭外，也會對病患的生命安全造成進一步的威脅。理想上，病患排斥反應的嚴重程度應藉由穿刺切片檢查來進行判斷，但因為肝臟移植病患手術後多有凝血功能障礙等相關問題，因此醫護人員轉而透過生化檢驗與血液檢驗來判斷病患排斥反應的嚴重程度，並依此來調整免疫抑制劑的使用量與用藥時間。 本研究期望能建構出一個可於臨床上進行應用的預測模型，因此模型除了必須能即時的提供預測結果外，也必須具有一定的預測準確度與可解釋性的優點。本研究將以資料驅動的角度出發，使用病患發生排斥反應前一天的抽血檢驗資料並透過機器學習演算法來進行建模。在建模的過程中，為了使模型能夠挖掘出發生排斥反應時各個檢測項目數值的變化趨勢，因此將檢驗資料進行變化幅度與類別化等處理；同時為了使模型具有可解釋的特點，因此透過整體式學習之方法來選取發生急性排斥反應時所產生的潛在規則，最終經由醫護人員的專業經驗從中挑選出較符合臨床上進行診斷的規則並將這些規則轉換成資料的新特徵，以建構出最終的預測模型。 本論文考慮了模型解釋力以及預測能力，研究結果顯示，本研究在準確率與其他先進機器學習演算法差異不大的狀況下，還可提供較佳的解釋，同時由於本研究建構出的模型可直接呈現進行預測時的預測規則，因此本研究成功的建構出具有即時性、可有效正確預測發生排斥反應的病患與擁有可解釋性的模型。

Patients who receive the surgery of liver transplant are most likely to have rejection of the new transplanted liver from the body immune system. A serious rejection may causes of not only liver failure, but also a further life threat for patient. Ideally, the severity of the patient's rejection should be judged by puncture biopsy, but the liver transplant patients have more clotting dysfunction after surgery. Therefore, the medical staffs in turn use the biochemical tests and blood tests to determine the severity of the patient's rejection, and adjust the use of immunosuppressive agents and medication time. The goal of this thesis is to construct a predictive model that can be applied clinically. Therefore, the model must provide the instant prediction results, and advantages of predictive accuracy and interpretability. This study proposes to use data-driven method to construct the model, using the day before the rejection of the patient's blood test data and through the machine learning algorithm to construct the model. To construct a model which can dig out the changes in the value of each test item when the rejection occurs, we propose to apply both amplitude of variation and the categorization to the test data. Meanwhile, to construct an interpretable model, we propose to use ensemble learning technique to select the potential rules of the acute rejection. Finally, with the help of the professional experience of health care workers, we select rules more in line with the clinical diagnosis and use these rules to form a new feature of the data to construct the final prediction model. The experimental results indicate that the proposed model could achieve almost identical accuracy as compared with state-of-the-art algorithms. Besides accuracy, the proposed model could offer the practitioners high interpretation. Meanwhile, the constructed model can directly show the forecasting rule once the forecasting is completed, indicating that we successfully construct a real-time, interpretable, and effective model in predicting the rejection of patients.