利用類神經網路建構急診部急性胰臟炎嚴重度預測系統

Construction of an Artificial Neural Networks Clinical Outcome Prediction Model for Acute Pancreatitis in Emergency Department

曾俊凱 Chun-Kai Tseng(Chun-Kai Tseng)

指導教授 ： 徐建業 ； 黃衍文

共同指導教授 ： 楊騰芳(Ten-Fang Yang)

臺北醫學大學/醫學院/醫學資訊研究所/碩士(2010年)

https://doi.org/10.6831/TMU.2010.00170

摘要

急性胰臟炎在急診室是一個很常見到的疾病。這些病人大部份預後良好，但仍有少部分病人會併發嚴重併發症，甚至是死亡。如何在早期辨識出這些較嚴重的病人，並給予更積極的醫療照護一直是ㄧ個很重要的課題。傳統有發展出幾種預測急性胰臟炎嚴重度的評分系統，但這些評分系統所要求收集的參數往往都需要很長的時間，甚至超過病人停留在急診的時間，因此這些評分系統並不適合真正利用在急診部評估急性胰臟炎的病人。

人工智慧自1950年代開始蓬勃發展，至今衍生出許多資訊系統來幫助決策的進行。其中利用神經生物學發展而來的類神經網路(Artificial Neural Networks, ANN)，更是因其平行運算及高度容錯的特性，而常被用於大量且複雜資料的分類。而臨床醫學分析本來就具有資料量大且繁雜的特性，因此自1985年起，開始有人將類神經網路應用於醫療決策上。

目的：本研究利用類神經網路架構一個適合急診部使用的急性胰臟炎嚴重度預測系統。此預測系統為適合急診部醫師使用，所使用之參數必須是可以輕易在病人就診早期就可以取得的。我們也將此系統的預測結果和傳統的modified Glasgow scoring system(mGS)以及以logistic regression(LR)所架構出來的預測系統互相比較。

方法：我們收集了台灣中部某醫學中心45個月期間，出院主診斷為急性胰臟炎的病人，以病歷回溯的方式收集了九項候選參數，以此參數架構出新的預測系統。將收集的病人隨機分為訓練模組和測試模組。為明瞭在不同的輸出變數下各種預測系統的預測準確度，本研究規畫三組不同的輸出變數，第一個是以住院是否超過7天為輸出變數。第二個則是以住院是否超過14天為輸出變數。第三個是以住院期間是否死亡做為輸出變數。在三種不同的輸出變數下，分別以訓練模組建構出類神經網路模組，LR預測模組以及計算傳統的mGS評分模組。再分別以測試模組測試預測之準確度並互相比較。

結果：在預測住院天數是否會超過7天的方面，類神經網路模組和LR模組皆比mGS優良(ANN vs. mGS, AUC=0.857:0.787, p<0.001; LR vs. mGS, AUC=0.835:0.787, p=0.003)。在預測住院是否會超過14天的方面，類神經網路的預測結果皆比LR和mGS優良(ANN vs. LR, AUC=0.903:0.803, p<0.001; ANN vs. mGS, AUC=0.903:0.764, p<0.001)。至於在預測病人住院期間是否會死亡方面，類神經網路模組的預測能力亦比LR和mGS優良(ANN vs. LR, AUC=0.978:0.886, p<0.001; ANN vs. mGS, AUC=0.978:0.827, p<0.001)。

結論：本研究顯示，利用類神經網路，以及簡單且早期的變數，是有可能建構出較傳統預測模式優良，可供急診部臨床使用的急性胰臟炎預後評估系統。但因本研究為回溯性研究，病人的組成和真實的情況並不一定相同，實際使用上是否會有相同的準確度則需要再臨床驗證。

關鍵字

急性胰臟炎 ； 類神經網路 ； modified Glasgow score ； logistic regression

並列摘要

Emergency department (ED) is a very busy unit in a hospital. The physicians of emergency department must spend their time on new patients’ diagnosis and old patients’ care. Under the limited resources of medical energy, how to identify the critical patients in ED and give aggressive care is an important entity in emergency medicine. In this study, we tried to construct a new prediction model by artificial neural networks (ANN) for the acute pancreatitis with early phase variables, and compared the result with logistic regression (LR) model and modified Glasgow scoring system (mGS).

Methods: We reviewed the 45-month patients with the diagnosis of acute pancreatitis in a medical center in Taiwan retrospectively and collected ten occult variables. All of the variables must be available in the four hours after admission. This study chose three different output variables, that were length of hospital stay (LOS) (more than seven days and more than fourteen days) and whether patients expired during admission. Under each output variable, we constructed three predicting models (ANN model、LR model and mGS) to predict the output variable, compared the area under receiver operating characteristic (ROC) curve.

Results: In the prediction of whether LOS be more than seven days, the ANN model and LR model all had greater area under the ROC curve (AUC) than mGS (ANN vs. mGS, AUC=0.857:0.787, p<0.001; LR vs. mGS, AUC=0.835:0.787, p=0.003). In the prediction of whether LOS be more than fourteen days, the ANN model was better than LR model and mGS (ANN vs. LR, AUC=0.903:0.803, p<0.001; ANN vs. mGS, AUC=0.903:0.764, p<0.001). In the prediction of whether patients be expired during admission, ANN model was also the best prediction model. (ANN vs. LR, AUC=0.978:0.886, p<0.001; ANN vs. mGS, AUC=0.978:0.827, p<0.001)

Conclusion: In this study, we showed that it is possible to construct a clinical ANN prediction model by early phase variables, and the performance of ANN model can be better than logistic regression model and modified Glasgow scoring system.

並列關鍵字

acute panreatitis ； artificial neural networks ； logistic regression ； modified Glasgow score

參考文獻

1. Beger, H.G., Rau, B., Mayer, J., & Pralle, U., (1997). Natural course of acute pancreatitis. World J Surg, 21(2), 130-135.

2. Corfield, A.P., Cooper, M.J., & Williamson, R.C., (1985). Acute pancreatitis: a lethal disease of increasing incidence. Gut, 26(7), 724-729.

4. Hirano, T. &Manabe, T., (1993). A possible mechanism for gallstone pancreatitis: repeated short-term pancreaticobiliary duct obstruction with exocrine stimulation in rats. Proc Soc Exp Biol Med, 202(2), 246-252.

7. Imrie, C.W., Benjamin, I.S., Ferguson, J.C., McKay, A.J., Mackenzie, I., O'Neill, J., et al., (1978). A single-centre double-blind trial of Trasylol therapy in primary acute pancreatitis. Br J Surg, 65(5), 337-341.

8. Barreto, S.G. &Rodrigues, J., (2007). Comparison of APACHE II and Imrie Scoring Systems in predicting the severity of Acute Pancreatitis. World J Emerg Surg, 2, 33.