- 學位論文
乳癌手術病人術後5年死亡預測模式之探討
Prediction Models of 5-year Mortality Analysis after Breast Cancer Surgery
摘要
研究背景與目的 乳癌（Breast cancer）是全世界女性最常見的癌症，也是造成死亡率最高的主因，號稱女性的頭號殺手，更成為亞洲之冠的排名，原本只在西方國家所重視的乳癌防治，已演變成全世界都關心的議題。衛生署100年公布女性乳癌佔癌症死亡排序第4順位，每十萬人口的死亡率由民國1986年的5％逐年上升至2011年的16％。因此，探討影響乳癌手術病人術後5年死亡之影響因子的議題是重要的，藉由比較類神經網路(Artifical Neural Network, ANN)及邏輯斯迴歸(Logistic Regression, LR)的預測模式，找出最佳的重要影響因子，進而依預測的模擬希望助於機構的管理、規劃與國家衛生政策的制定及推動，甚至當成決策支援系統之參考。故研究目的如下： 目的一、探討乳癌手術病人術後5年死亡之長期趨勢分析 目的二、比較乳癌手術病人術後5年死亡不同預測模式之準確度 目的三、利用全域敏感度分析，評估影響乳癌手術病人術後5年死亡重要預測因子之權重 研究方法 本研究架構以「全民健康保險學術資料庫」為研究主體，採回溯性的研究設計，研究時間自1996-2010年，研究對象為年齡＞16歲乳癌手術後的病人，研究樣本共3,632人。利用ICD-9-CM174x（174.0-174.9）之診斷碼，以及處置碼85.20-23、85.33-36、85.4x、85.5x、85.6x、85.7x、85.8x、85.95，找尋有意義的預測因子，分別運用病人的特性、醫院的特性及時間的特性探討乳癌手術病人術後5年死亡分析之重要影響因子；另外將預測的顯著因子利用類神經網路與邏輯斯迴歸建立模型，比較出乳癌手術死亡分析的準確性。 採用SPSS 19.0版本的統計套裝軟體工具進行資料整理與分析，主要的統計方法包括：描述性統計與推論性統計（趨勢分析、單變量分析、多變量分析-含邏輯斯迴歸和類神經網路及敏感度分析）。 研究結果 本研究利用資料探勘技術之類神經網路(ANN)與邏輯斯迴歸(LR)作為預測模型之建構，並使用7個重要變項（年齡、合併症嚴重度指標(CCI)、醫院層級、醫師累計服務量、化療、放療及賀爾蒙用藥），結果顯示以類神經網路預測模式優於邏輯斯迴歸預測模式。乳癌病人術後5年死亡預測，類神經網路與邏輯斯迴歸於敏感性(Sensitivity)分別為5.66%及3.77%；特異性(1-Specificity)、陽性預測值(PPV)、陰性預測值(NPV)及準確性(Accuracy)均為相當表現；在AU-ROC曲線分別為0.70及0.52，整體來說類神經網路的表現優於邏輯斯迴歸模式。 重要預測因子方面，類神經網路的前3項因子為醫師累計服務量、化療、年齡；邏輯斯迴歸前3項因子為醫師累計服務量、年齡、合併症嚴重度指標(CCI)。 結論與建議 研究結果發現乳癌術後5年死亡在病人特性及醫院特性之分佈趨勢上，隨著時間有顯著性改變，表示研究出的相關顯著因子在臨床上仍是值得被應用且成為改善因子，甚至當成治療方針之準則；在不同的預測模式之比較，發現可使用類神經網路預測模式並擴大預測變項，以利各類疾病進行系統性的相關分析探討，有效運用適當的研究方法及發展完善的死亡預測模式，並將此預測模式推廣到其他癌症；另外利用全域敏感度分析出以醫師累計服務量為最首要之因子，代表其影響死亡率為負相關，可讓衛生單位如何協助及培養醫師相關經驗之參考，其次影響因子如年齡增加使死亡率的提升，更應反思與乳癌發生年齡的相關性，應注重如何促進及提升篩檢率等議題，以其達到早期發現早期治療，最終期望能提供給相關醫療人員之臨床參考價值及整合治療與建立醫療決策之預測模式。
並列摘要
Background and Purpose Breast Cancer is the most common cancer in the world women. Also, the breast cancer is causing the highest mortality rate the main reason. The breast cancer is the number one killer for women and become the highest ranking in Asia. This study is therefore comparing artificial neural network (ANN) and logical regression (LR) prediction models to find the best of important effect factors. The purposes of this research are as follows: Ⅰ、To investigate long-term trend analysis of the breast cancer patient after surgery in 5-year mortality; Ⅱ、To compare the accuracy of different predict models for breast cancer patients after surgery in 5-year mortality; Ⅲ、To conduct the global sensitivity analyze and to estimate the significant predictors for breast cancer patients after surgery in 5-year mortality. Research Methods This study subject of "National Health Insurance Research Database" is the research framework. The study design used the retrospective method. The study period is from 1996 to 2010. The study subjects of the breast cancer patients are above sixteen years old after surgery. The samples of study are total 3,632 people. The use of the diagnostic codes ICD-9-CM174x (174.0-174.9) and disposal code 85.20-23,85.33-36,85.4 x, 85.5x, 85.6x, 85.7x, 85.8x, 85.95 find the predictor factors of meaningful. Investigate the important effect factors are by the breast cancer patient analysis after surgery in 5-year mortality and use of the patient's characteristics, hospital characteristics and time characteristics, separately. In addition, the significant predictors are used in the ANN and LR to build model and to compare the accuracy. SPSS 19.0 statistical software was employed for the data collection and analysis. The main statistical methods include: descriptive statistics and inferential statistics (trend analysis, university analysis, and multivariate analysis - including logistic regression and neural networks and sensitivity analysis). Results In this study, the model build are used data mining technology of ANN and LR and seven important variables (age, charlson comorbidity index (CCI), hospital level, hospital volume, surgeon volume, chemotherapy, radiotherapy and hormone therapy). The results show that ANN model is better than LR model. The sensitivity of ANN and LR is 5.66% and 3.77%, respectively. The 1-specificity, positive predictive value, negative predictive value and accuracy are good performance. The AU-ROC curve is 0.70 and 0.52, respectively. Generally, the performance of ANN is better than LR model. The first three important predictors of ANN are surgeon volume, chemotherapy and age. The LR model are the surgeon volume, age, charlson comorbidity index (CCI). Conclusions and Recommendations The research results found that the distribution trend for the breast cancer patient after surgery in 5-year mortality is a significant change with time in patients’ characteristics and hospital characteristics. The means show that the correlation significant factors are worth applied in clinically and become improvement factors even if as standard of treatment guidelines. To compare the different predict models, found that the neural network can be used to expand predict variable items. The various diseases are easy to analyze and investigate, systematically. The death predicts model is use of appropriate research methods with development. In the future, this prediction model can applied to other cancers. In addition, the most important factor is surgeon volume. So that, how to enable health authorities to assist and train surgeons relevant experience reference. Secondary the impact factors such as age are increased with mortality rate. Therefore, the breast cancer should repeated think the relation with age and pay attention in how to promote and improve screening rates issues and to reach early detection and early treatment. Finally, this study expect can provide clinical reference value for medical staff and integrate medical treatment and to establish a predict model of medical decision.