探討不同預測模式在肝臟移植手術病患死亡率之比較分析

研究目的肝臟移植(OLT)是罹患肝癌末期患者唯一可根治的術式。從1996年的個位數到2009年的三位數，臺灣進行肝臟移植手術有越來越多的現象，因此，本研究使用類神經網路(Artificial Neural Network, ANN)與對數迴歸(Logistic Regression, LR)預測病患術後住院期間及其術後五年之死亡率，並進行兩種統計方法準確性之探討，期能發展出最佳預測模型，以作為臨床決策之參考。研究方法本研究使用1996年至2010年之全國性資料庫進行探討，研究樣本共2,224位，針對病患人口學特質、疾病特質、醫院特質及術後併發症等四個構面進行死亡率預測，以類神經網路和對數迴歸所建立之預測模型進行準確度比較，進而運用全域敏感度分析(Global Sensitivity Analysis)找出影響術後住院期間及術後五年內死亡率的重要預測因子。研究結果類神經網路與對數迴歸之預測效能方面，術後住院期間死亡率之準確度，ANN=0.906、LR=0.899；術後5年死亡率之準確度，ANN=0.818、LR=0.806。AUROC方面，術後住院期間，ANN=0.779、LR=0.554；術後5年，ANN=0.704、LR=0.555。重要預測因子方面，術後住院期間死亡率之ANN與LR前3項因子相同，為住院天數、敗血症及肝癌，術後5年內死亡率，ANN與LR之第1項預測因子同樣為敗血症。結論與建議依據本研究預測模式類神經網路與對數迴歸皆有良好的準確性，而類神經網路之預測效能比對數迴歸更佳，但不同之預測模型皆有其優劣性及適切性，不宜就單一種統計方法建立預測模型，建議除運用此兩種預測模式之外，可嘗試建立進行複合式預測模型，以提供醫療人員有效資訊；且本研究之研究範圍涵蓋甚廣，所採樣本時間長度橫跨15年，科技日新月異，每一時期病患所處之醫療環境勢必有所不同，建議未來相關研究可將研究時期適當地加以畫分，並可互相比較。而在預測病患住院死亡率及五年死亡率，影響的最重要預測因子皆為醫院服務量，更進一步思考，國內醫療體系資源是否分配不均，政府應先早一步洞悉此現象所代表之意義，並加以預防，以期國內民眾可獲得更公平、更完整之醫療照護。

關鍵字

肝臟移植；類神經網路；對數迴歸；全域敏感度分析

並列摘要

OLT is the only way to completely cure the patient who has advanced liver cancer. Thus, there are getting more liver cancer patients choose to accept the OLT in Taiwan. The purpose of this study was to validate the use of artificial neural network (ANN) models for predicting in-hospital mortality and 5-year mortality in OLT and to compare their predictive capability with that of logistic regression (LR) models. This study compared LR and ANN models based on initial clinical data for 2,224 OLT patients, including age, gender, Charlson co-morbidity index (CCI) score, etiology, liver cirrhosis, hospital volume, surgeon volume, length of stay (LOS) and outcome. A global sensitivity analysis was also performed to assess the relative significance of input parameters in the system model and to rank the importance of variables. The predicted efficiency aspect of ANN and LR, to accuracy of being in hospital is ANN=0.906、LR=0.899; after five years of surgery is ANN=0.818、LR=0.806. In AUROC aspect, be in hospital after surgery is ANN=0.779、LR=0.554; after five years of surgery is ANN=0.704、LR=0.555. In the aspect of important predicted factors, the top three factors of ANN and LR are the same of being in hospital, they are service amount of hospital, age of patient, and complication after surgery. Besides, the first factor of ANN and LR are also the acute renal failure. In comparison with the conventional LR model, the ANN model in this study was more accurate in predicting in-hospital mortality and 5-year mortality and had higher overall performance indices. Further studies of this model may consider the effect of a more detailed database that includes complications and clinical examination findings as well as more detailed outcome data.