風險預測管理在臨床研究和醫療照護中意義重大，尤其肝硬化病人與一般重症病人相比 預後較差，但目前尚無針對肝硬化病人長期死亡風險的預測指標，因此本研究試圖使用機器 學習方法：Bagging、Adaboost、Support Vector Machine (SVM)、Random forest 建構預測模 型。研究納入2013年7 月1 日至2018年12 月31 日就醫的528 名個案並追蹤至2020 年6 月，並 與目前臨床肝硬化嚴重度量表：Child-Pugh Score (CTP)、MELD Score (Original)、MELDNa Score、MELD Score以及重症單位常用的The Acute Physiology and Chronic Health Evaluation (APACHE II) 進行比較，結果顯示機器學習模型皆優於現有的臨床評估量表。預測1 個月死 亡模型以Random forest 表現最佳( 準確率：0.836；AUC: 0.844)；預測12 個月內死亡模型以 SVM表現最佳( 準確率：0.826；AUC: 0.869)；預測12 個月後死亡模型以Adaboost表現最佳 ( 準確率：0.759；AUC: 0.760) ，充分證實機器學習模型在預測肝硬化病人的長期死亡率方面 具有應用價值。

Risk prediction management is of great significance in clinical research and medical care. The prognosis of patients with liver cirrhosis is worse than compared with other groups of critically ill patients. However, there are no scoring systems design of long-term mortality risk for patients with cirrhosis.The study included 528 patients who received medical treatment from July 1, 2013 to December 31, 2018 and was followed up to June 2020. This research attempts to use machine learning methods(Bagging, Adaboost, Support Vector Machine (SVM) and Random forest) to predicted mortality of patients with cirrhosis were performed and compared with CTP, MELD Score (Original), MELDNa Score, MELD Score, and The Acute Physiology and Chronic Health Evaluation(APACHE II). The result pointed out that compared with the existing scoring systems, machine learning method model has better prediction effect. Random forest is the best model for predicting death within 1 month (accuracy rate: 0.836; AUC: 0.844); SVM is the best model for predict death within 12 months (accuracy rate: 0.826; AUC: 0.869); Adaboost is the best model for predict death after 12 months(accuracy rate: 0.759; AUC: 0.760). It fully proves that the machine learning model has application value in predicting the long-term mortality rate of patients with cirrhosis.