運用資料探勘方法建立乳癌患者存活之預測模式

研究背景:資料探勘的技術方法已成功的運用在許多不同的領域之上，且在醫療領域的治療診斷和預後評估上也可以提供廣大效益，而透過癌症資料庫所建立的預測模型對於臨床醫學及研究上也具有實質之幫助。罹患癌症與相關危險因子有密切關係，而癌症預後因子研究對臨床上皆能有所幫助，不僅能夠協助了解疾病惡化程度，也可提供病人有益相關治療的信息，進而提升癌症病人存活率。研究方法:本研究利用資料探勘方法在於進行乳癌資料分析上，使用美國國家研究院The Surveillance Epidemiology and End Results (SEER)乳癌病患資料集，在特徵上我們以醫療文獻來提取特徵因子及機器學習挑選的特徵因子，分別以Decision Tree決策樹、Random Forest隨機森林、Logistic Regression邏輯斯迴歸三種演算法對進行1~5年存活預測模型之建立。研究結果: 在預測乳癌病患五年存活情形比較，研究的結果顯示決在醫療特徵組上，以決策樹Decision Tree預測模型Accuracy (0.9901)、AUC(0.9892 %)，預測效能最好；在機器挑選特徵上，以Random Forest預測模型Accuracy (0.9912)、AUC(0.9913 %)，測效能最好；然而Logistic Regression在預測上的效果較不好，對於此結果可以證實出Decision Tree決策樹及Random Forest隨機森林兩種方法，對於乳癌存活預測上具有不錯的鑑別力，此外，從兩種特徵來看，以機器學習所挑選特徵與醫療所挑選特徵更具有存活影響力；以預後因子重要排序上來看，在存活預測中以腫瘤主要部位(Primary Site)、診斷年齡 (Diagnosis Age)最為影響預後

關鍵字

乳癌；存活率；資料探勘；決策樹；隨機森林；邏輯斯迴歸

並列摘要

Background: The Data mining have been successfully applied in many different fields, and they can provide benefit and prognosis evaluation in the medical field. Cancer prognostic factors can to understand the severity of the disease and providing information about the patient beneficial treatment, increasing the survival rate of cancer patients. Methods: This study use the Surveillance Epidemiology and End Results Program (SEER) of the National Cancer Institute from U.S for breast cancer patients. Also we use two different characteristic factors (medical literature feature and machine learning selection feature), and take the Decision Tree, Random Forest and Logistic Regression three kinds data mining methods of one to five year survival rate. Results: The study results showed that Machine selection feature performed better than medical characteristic selection feature, and random Forest method and Decision Tree method is better than Logistic Regression method. In the prognostic factors important sequencing, in survival prediction with the Primary site and diagnosis age is most important effect on prognosis.