- 學位論文
利用乳癌篩檢前的問卷建構機器學習模型預測乳癌風險
Using machine learning to Predict Correlation of Pre-mammography Questionnaire with Breast Cancer
摘要
機器學習在近數十年來有很大的進展。許多的演算法已被發展出來且應用於各個領域。乳癌(breast cancer)，是全世界女性最常見的癌病，一年約有50萬人死於乳癌。因此，如何早期診斷乳癌就相形重要，而乳房攝影則是目前證實最有效的篩檢工具。美國放射線醫學會已發展出一套系統：乳房攝影報告與資料解讀系統(BI-RADS)，可將乳房攝影的結果，做標準化的判讀，予以分成六類。由於乳房攝影前的問卷內容大多與乳癌發生因子有關，我們可利用病患臨床檢查前的問卷數據和BI-RADS的結果進行分析，設計建構出機器學習的乳癌風險預測模型。 Gail 模型是目前被廣泛使用的乳癌風評估工具。我們使用 3 種機器學習演算法—隨機森林、支援向量機及類神經網路來建構乳癌風險預測模型，並和 Gail 模型作比較。使用混淆矩陣(準確度、靈敏度、特異度、陽性預測值，陰性預測值)以及接收者操作特徵曲線下面積等指標，評估該預測模型效能，並證明機器學習模型預先評估乳癌發生機率的可行性。本研究從 2009年 1 月至 2013 年 12 月，以新北市某醫院之衛生福利部國民健康署免費提供的乳癌篩檢資料為母群體，年齡自45至69歲，共30634位篩檢婦女。 在這受檢母群體中，若其BI-RADS結果為第Ⅰ、Ⅱ類，則歸於低風險，若其結果為BI-RADS第Ⅳ、Ⅴ、Ⅵ 類則歸於高風險。總共收錄了 688 個案例。BI-RADS 分類 Ⅰ、Ⅱ, Ⅳ,Ⅴ,Ⅵ 的個數分別為 299，301， 32 14及42。我們選了問卷中的14 個輸入變項來建構模型。以70% 的數據資料作訓練組，30% 的數據資料作測試組，則得到隨機森林、支援向量機器、類神經網路及 Gail 模型的曲線下面積分別為 0.82, 0.72, 0.72, 0.53。此結果顯示隨機森林模型有較佳的效能。 另外，BI-RADS分類Ⅳ代表著“有異常病灶的可能性”，乳癌的發生機會可低至3%或高至95%，這樣大的落差容易造成婦女極大的恐慌，並且面對未必需要的介入性穿刺或切片檢查。其結果除了極高比例的陰性反應而形成的醫療資源浪費，更造成婦女心理上因乳房外觀缺陷而產生心理陰影或易沉浸在得乳癌的恐懼中。因此，若有一輔助工具加以協助此一部分不足之處，是有其需要性。依據乳房攝影檢查報告結果，依 BI-RADS報告分類系統作分類。BI-RADS分類Ⅳ 並完成切片確診案例共137人，確認得乳癌者46人，未得乳癌者91人。我們選了同樣的14 個輸入變項，並加入乳腺密度來建構模型。70% 之數據資料當作訓練組，30% 之數據資料當作測試組，則得到隨機森林、支援向量機器、類神經網路及 Gail 模型的曲線下面積分別為 0.81, 0.71, 0.85, 0.59。此結果顯示，類神經網路有較佳的效能。 由以上的研究，我們展現了使用機器學習技術發展出臨床決策輔助工具，具有更好的效能。同樣的技術將可用於改造其他的傳統臨床決策輔助工具。未來結合電子病歷，機器學習將有助於臨床人員提供更佳且更有效率的醫療照護。
並列摘要
The progress of machine learning is robust in recent decades. A lot of algorithms for machine learning are developed and applied to distinct fields. Breast cancer is one of the most common female malignancies worldwide, and five hundred thousand female patients die of this disease every year. For early detection of the breast cancer, mammography is a pivotal tool to screen. Recently, American College of Radiology developed a system, Breast Imaging-Reporting and Data System (BI-RADS), was designed to standardize the reading and reporting of mammography, and classified the risk into six categories. Moreover, the examined women will answer the questionnaires before they receive the mammography. Because the contents of most pre-mammography questionnaires are regarding the risk factors of breast cancer, we can correlate the questionnaires with the results of BI-RADS to construct a predicting model for the occurrence of breast cancer, using machine learning, including artificial neural network, random forests and support vector machines. The objective of this study was to construct a breast cancer risk-predicting model, using a more widely used tool, Gail model, as a standard. We applied three algorithms of machine learning, including Random Forests, Support Vector Machines, Artificial Neural Network, to the new breast cancer risk-predicting models, in comparison with Gail model. In our methodology, we input the items of questionnaire as variables into each predicting models. Subsequently, we used area under the receiver operating characteristic curve and the indicators of confusion matrix, including accuracy, sensitivity, specificity, positive predictive value and negative predictive value, to analyze the variables and to evaluate the efficacy of our model. This methodology could validate the feasibility of machine learning model in the prediction of breast cancer risk in advance. Based on the population of breast cancer-screening program provided by Health Promotion Administration, Ministry of Health and Welfare, we selected 30634 women as a basis in a medical center of New Taipei City, aged from 45 to 69 years, joining free-charged breast cancer screening, duration from January, 2009 to December, 2013. In the population, the groups of BI-RADS Category 1-2 and 4-6 were classified as low and high risk, respectively. Besides, we select 14 items of questionnaire as input variables into into the models. After looking into the details, a total of 688 cases were qualified. The numbers of Category 1, 2, 4, 5 and 6 were 299, 301, 32, 14 and 42. Then we used 70% of qualified cases as a training set and 30% as a testing set. The result showed that the AUC of Random Forests, Support Vector Machines, Artificial Neural Network and Gail models were 0.82, 0.72. 0.72 and 0.53, respectively, indicating Random Forests had the best efficiency. Furthermore, we particularly analyzed BI-RADS Category 4. Why did we do that? BI-RADS Category 4 represents for “Suspicious Abnormality” and the probability ranges from 3% to 95%. The range of possibility is so wide that may make a lot of examined women panic and some core biopsies unnecessary. The high possibility of negative findings will cause a medical waste. And the mentality of examined women will be stuck in the fears of defective breast appearance and breast cancer diagnosis. Therefore, it is mandatory to develop an assisting tool to compensate the unmet need. Based on the report of examined population, a total of 137 women had the result of Category 4 and subsequently received a core-biopsy for the further confirmation. The pathological results disclosed the breast cancer in 46 women and non-breast cancer in 91. We selected the aforementioned 14 input variables plus the variable of mammary gland density into the models. Then we used 70% of biopsy-proven cases as a training set and 30% as a testing set. The result showed that the AUC of Random Forests, Support Vector Machines, Artificial Neural Network and Gail models were 0.81, 0.71. 0.85 and 0.59, respectively, indicating Artificial Neural Network had the best efficiency. To be concluded, we revealed an efficient decision-making tool developed by machine learning technique. This technique provides an avenue of methodology for developing other assisting tool in making clinical decision. In future, the model of combining the medical chart with the machine learning technique will support a better and more efficient medical care.