利用機器學習理論探討肝癌病患術後2年內再復發風險之預測模式

研究目的: 肝癌病人手術後常擔心是否會再復發之風險，因此本研究將利用類神經網路(Artificial Neural Network, ANN)、最近鄰居演算法(K-Nearest Neighbor, KNN)、支持向量機(Support Vector Machine, SVM)、單純貝氏分類器（Naïve Bayes Classifiers, NBC）與Cox迴歸分析(Cox Regression Analysis, COX)預測模式評估肝癌病患術後二年內再復發風險之準確性，並利用全域敏感度分析(Global Sensitivity Analysis)評估影響肝癌手術病患術後二年內再復發風險重要預測因子之權重。研究方法：本研究為前瞻性世代研究設計，以問卷方式蒐集病人之研究變項，經人體試驗委員會審核通過，本研究樣本以台灣南部某三家醫學中心於2017年11月至2018年12月接受肝手術之病人，評估病人術前焦慮(BAI)與憂鬱(BAI)之分佈狀況，並以病歷審查蒐集病人人口學特性、臨床特性、醫療機構特性及醫療照護品質特性。本研究利用敏感性(Sensitivity)、特異性(Specificity)、陽性預測值(PPV)、陰性預測值(NPV)、準確性(Accuracy)及曲線下面積(AUROC)作為上述四種預測模式其績效評估之指標，進而利用全域敏感度分析探討影響肝癌手術病患術後二年內再復發風險重要預測因子之權重。本研究以SPSS 22.0版與Statistica Academic 13.0版統計套裝軟體進行統計分析。研究結果：本研究利用類神經網路(Artificial Neural Network, ANN)、最近鄰居演算法(K-Nearest Neighbor, KNN)、支持向量機(Support Vector Machine, SVM)、單純貝氏分類器（Naïve Bayes Classifiers, NBC）與Cox迴歸分析(Cox Regression Analysis, COX)作為預測模型之建構，並使用15個變項，結果顯示肝癌病患術後兩年內再復發之預測，在敏感性（Sensitivity）部份類神經網路(Artificial Neural Network, ANN)、最近鄰居演算法(K-Nearest Neighbor, KNN)、支持向量機(Support Vector Machine, SVM)、單純貝氏分類器（Naïve Bayes Classifiers, NBC）與Cox迴歸分析(Cox Regression Analysis, COX)分別為93％、81％、82％、27％與66％；陰性預測值（NPV）分別為97％、87％、83％、60％與42％；準確性（Accuracy）分別為96％、76％、78％、63％與31％；在AU-ROC曲線分別為94％、77％、78％、60％與40％。整體而言類神經網路預測模式(Artificial Neural Network, ANN)優於最近鄰居演算法(K-Nearest Neighbor, KNN)、支持向量機(Support Vector Machine, SVM)、單純貝氏分類器（Naïve Bayes Classifiers, NBC）與Cox迴歸分析(Cox Regression Analysis, COX)預測模式。最後，經由全域敏感度分析發現腫瘤期別是病患術後兩年內再復發之最重要影響因子，其次為合併症指數(CCI)、30天內再入院及B型肝炎。結論與建議：研究結果顯示，肝癌手術病患其術後兩年內再復發風險之預測，建議應使用類神經網路預測模式進行估計。本研究亦發現，腫瘤期別是影響病患術後兩年內再復發之最重要預測因子，其次為合併症指數(CCI)、30天內再入院及B型肝炎，可用於提供照護癌症病患預計接受手術治療之醫護人員以及病患與家屬，於術後醫療恢復狀況之參考，進而提升醫療照護品質。

關鍵字

肝癌；再復發；機器學習理論；類神經網路；全域敏感度分析

並列摘要

Purposes: Liver cancer patients often worry about the risk of relapse after surgery. In this study, Artificial Neural Network (ANN), K-Nearest Neighbor(KNN), Support Vector Machine (SVM), Naive Bayes Classifiers (NBC) and Cox Regression Analysis (COX) prediction models were used to predict the accuracy of the risk of relapse of liver cancer patients within two years after the surgeries. Global Sensitivity Analysis was also used to predict the weight of the important factors that affects the risk of relapse of liver cancer surgery patients within two years after surgery. Research Methods: This study is designed for a prospective generational study. Questionnaires were used to collect the research variables of patients. After the human trial committee reviewed and approved, the sample of this study was the patients undergoing liver surgery at a medical center in southern Taiwan from November 2017 to December 2018. Evaluate the distribution of patients’ preoperative anxiety (BAI) and depression (BAI), and collect patient demographic characteristics, clinical characteristics, medical institution characteristics and medical care quality characteristics through medical record review. This study used Sensitivity, Specificity, Positive Predictive Value (PPV), Negative Predictive Value (NPV), Accuracy (Accuracy) and Area Under the Curve (AUROC) as the indicator of effectiveness evaluation of above four prediction modes. And then used global sensitivity analysis to discuss the weight of important predictors that affect the risk of relapse within two years after liver cancer surgery. In this study, statistical analysis was performed using SPSS version 22.0 and Statistica Academic version 13.0 statistical package software. Results: This study used Artificial Neural Network (ANN), K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Naive Bayes Classifiers (NBC) and Cox Regression Analysis (COX) as the construction of the prediction model, and used 15 variables. The results showed that the prediction of relapse of liver cancer patients within two years after surgery, in the evaluation of Sensitivity, Artificial Neural Network (ANN), K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Naive Bayes Classifiers (NBC) and COX Regression Analysis (COX) were 93%, 81%, 82%, 27% and 66% respectively; Negative predictive value (NPV) were 97%, 87%, 83%, 60 % And 42% respectively; Accuracy is 96%, 76%, 78%, 63% and 31% respectively; in the AU-ROC curve is 94%, 77%, 78%, 60% and 40%. In general, Artificial Neural Network (ANN) is superior to K-Nearest Neighbor (KNN), Support Vector Machine (SVM), and simple Bayesian classifier (Naive) Bayes Classifiers (NBC) and Cox Regression Analysis (COX) prediction models. Finally, after the global sensitivity analysis, it is found that the tumor stage was the most important factor affecting the relapse of patients within two years after surgery, followed by the comorbidity index (CCI), readmission within 30 days, and hepatitis B. Conclusions and Suggestions: The results showed that the prediction of the risk of relapse within two years after surgery for patients with liver cancer surgery should be estimated by Artificial Neural Network (ANN) prediction model. This study also found that tumor stage is the most important predictor of relapse within two years after surgery, followed by the comorbidity index (CCI), readmission within 30 days and hepatitis B, which can be used to provide care of cancer diseases for medical staff, patients and their families who are expected to undergo surgical treatment. It can also be the reference of the postoperative medical recovery status to further improve the quality of medical care.