婦科癌症病人焦慮或憂鬱症預測模式之探討-以子宮頸癌、子宮內膜癌及卵巢癌為例

中文摘要研究目的根據統計資料，罹患卵巢癌、子宮頸癌及子宮內膜癌的人數，發現有逐年上升的趨勢，台灣卵巢癌的發生率佔婦女癌症之第三位，僅次於子宮頸癌及子宮內膜癌，但卻是婦癌中致死率最高的癌症。然現今癌症治療不只著重存活率，病人身心理狀況、適時照護，提升患者生活品質已成為醫療照護品質重要指標。因此本研究希望透過類神經網路（ANN）、邏輯斯迴歸（LR）及比例風險迴歸（COX）分析找出之重要因子，以利醫療團隊於醫學的應用領域，故本研究的目的如下：目的一：利用效益指標比較三種預測模式(Forecasting Models)之準確性目的二：利用全域敏感度分析(Global Sensitivity Analysis)評估婦科癌症患者焦慮或憂鬱症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」為研究樣本，採回溯性研究設計，研究期間為1996至2010年，研究對象為年齡≧18歲的婦科癌症病人，並以ICD-9-CM代碼180、182、183為篩選研究對象，樣本數共6,801人。並利用病人特性、醫療機構特性來探討婦科癌症病人焦慮或憂鬱症之重要影響因子，另外利用類神經網路、邏輯斯迴歸與比例風險迴歸，比較婦科癌症病人焦慮或憂鬱症的準確性。研究結果本研究利用類神經網路（ANN）、邏輯斯迴歸（LR）與比例風險迴歸（COX）作為預測模型之建構，並使用13個變項，結果顯示類神經網路預測模式優於邏輯斯迴歸（LR）與比例風險迴歸（COX）預測模式。婦科癌症病人焦慮或憂鬱症的準確性，在敏感性部份類神經網路、邏輯斯迴歸（LR）與比例風險迴歸（COX）分別為67.23％、2.90％與52.02％；陰性預測值（NPV）分別為84.04％、66.91％與0.00％；準確性（Accuracy）分別為81.19％、66.36％與17.32％；在AU-ROC曲線分別為88％、53％與26％。整體而言類神經網路（ANN）表現優於邏輯斯迴歸（LR）與比例風險迴歸（COX）預測模式。結論與建議從本研究結果可以證明類神經網路預測模式是優於邏輯斯迴歸和比例風險迴歸模式。故醫學相關研究可多運用此模式做為臨床的評估及決策參考，找出最重要的預測因子，以提升醫療品質。關鍵詞：子宮頸癌、子宮內膜癌、卵巢癌、焦慮症、憂鬱症、類神經網路

關鍵字

子宮頸癌；子宮內膜癌；卵巢癌；焦慮症；憂鬱症；類神經網路

並列摘要

Abstract Three purposes of this study: 1. To compare the performance indices between artificial neural networks (ANN), logistic regression (LR) and Cox proportional hazards (COX) models; 2. To conduct the global sensitivity analysis in order to weight these significant predictors. Research Methods This nation-wide population-base study retrospectively conducted the claims data from 1996 to 2010. Included criteria were the subjects age larger than 18 years old of gynecologic cancer patients ; the patients with ICD-9-CM diagnosis codes of 180,182,183,and use patient characteristics and hospital characteristics to identify the impact factors of anxiety and depressive disorders.The comparison of performance indices of ANN, LR, and COX models were employed to predict the anxiety and de-pressive disorders . The global sensitivity analysis was also used to weight these sig-nificant predictors. Results The results showed that ANN model is better than LR and COX models in predicting these performance indices: sensitivity (67.23％ vs. 2.90％ vs. 52.02％), NPV (84.04％ vs. 66.91％ vs. 0.00％), accuracy (81.19％ vs. 66.36％ vs. 17.32％), area under the accept operating characteristic curve (AUROC) (88％ vs. 53％ vs. 26％). Overall for the anxiety and depressive disorders, the ANN model also showed the better performance indices than the LR and COX models. Conclusions and Suggestions The ANN model showed the better performance indices than the LR and COX models. Medical research can therefore use this model as a multi-clinical assessment and deci-sion making, and to identify the most important predictors to improve quality of care. Keyword：cervical cancer, endometrial cancer, ovarian cancer, anxiety disorders, de-pressive disorders, artifical neural network (ANN)