運用監督式學習技術建構手術病患住院天數評估模式之研究

Study of Using Supervised Learning Techniques to Construct a Model for Assessing Surgical Length of Hospital Stays

蔡楚凡(Chu Fan Tsai)

指導教授 ： 胡雅涵

國立中正大學/管理學院/資訊管理系醫療資訊管理研究所/碩士(2014年)

摘要

根據國民健康保險在台灣的統計資料指出，外科手術很容易造成醫院的標準住院天數延長或違反包裹式的醫療付費方式，主要的原因是手術後所產生的併發症。現今，延長住院天數也容易使醫生和病患之間的議論與糾紛，這將導致越來越多的舞弊行為的發生，尤其是在風險高的手術行為。因此，無論手術病患的住院天數的延長與否，住院天數對於手術風險管理及改善醫師與病人間的關係皆有著強烈的影響。住院天數長短的預估可做為醫師提供給病患家屬溝通的重要階段，亦可做為加強提升醫療照護服務的依據。
本研究資料來源於南台灣某區域級醫院合作，資料收集的對象為2006年1月至2012年12月期間的病患，共彙整出896筆經一般外科手術的病患住院紀錄。使用內部驗證與外部驗證評估，並應用資料探勘軟體weka其內建的分類技術決策樹J48、隨機林、邏輯斯迴歸、支援向量機及搭配加強分類器Bagging，比較預測模型的正確率。評估模型的性能的指標包括準確度、敏感度、特異度和AUC曲線，此外我們還分析了模型中的自變項和依變項之間的關係，包括病人資料、醫生特徵，護理紀錄及手術紀錄等。
本研究主旨在於探討一般外科的手術病患在不同的診斷代碼下的有效性，在單一分類技術及十折交叉驗證之下，羅吉斯迴歸AUC指標高達77.6%。在本研究中有無輸血、有無使用腹腔鏡、進入ICU的次數、血紅素、血比容和病患年齡是最有影響力的因素。
本研究使用不同術式的資料，建構的住院天數評估模型，而不是只能應用在單一術式的評估工具。本研究所建構的模型也能減少醫生可能發生的弊端。此外，醫療機構可廣泛採用以人工智慧為基礎的預測技術，應用在不同的外科手術，進而評估術前診斷階段的手術風險。

關鍵字

手術風險評估 ； 監督式學習技術 ； 住院天數 ； 分類器

並列摘要

According to the statistics data from National Health Insurance in Taiwan, surgical operation is easy to extend hospital standard stays or make the package payment system of medical claims break out and the major reason is the postoperative complications. Extending hospital stays is also easy to make the poor communication between physicians and patients which would result in more and more malpractices happen nowadays, especially in serious surgery cases. Therefore, It has strong implications for surgery risk management and physician-patient relationship improvement whether the length of hospital stays of surgical patient extents standard stays or not. The length of hospital stays estimation also provide an important stage for strengthen density of medical services and more communication with the families of those patients.
The data consists of complete historical records of 896 Taiwan clinical cases who received surgery therapy from January 2006 to December 2012 in general surgery department of the medical center of Taiwan. The 10-fold cross validation technique is adopted for constructing the training and testing datasets in all experiments. We evaluate the prediction performance of four classification techniques, including logistic regression, decision tree J48, random forest, and support vector machine. We further consider homogeneous ensembles as Bootstrap Aggregation (Bagging) algorithm. Measures including accuracy, sensitivity, specificity and receiver operating characteristic curve are used to evaluate model performance. Besides, we also analysis our models in which the relationship among the dependent variable and independent variables included the patient information, surgeon character, nursing character, and operative character.
This study evaluates the effectiveness of different patients with different diagnosis code. Logistic regression based classifiers outperform other investigated classifiers in all evaluation measures and acute peritonitis patients are significantly more accurate than other operation cases (77.6%). Blood transfusion, laparoscopic surgery, length of stay in intensive care unit, the level of Heme and HCT and age are the most influential factors in this study.
The empirical evaluation suggests construct the surgical risk prediction model by different operation instead of single evaluation tool. The investigated models also help physicians to reduce the possible malpractice happened. Moreover, Healthcare institution could widely adopt the AI-based prediction techniques to evaluation the risk of surgery for another surgery department in the pre-surgical diagnosis stage.

並列關鍵字

risk of surgical evaluation ； supervised learning techniques ； length of hospital stay ； classifiers

參考文獻

Badriah, F., Abe, T., Nabeshima, Y., Ikeda, K., Kuroda, K., & Hagihara, A. (2013). Predicting the length of hospital stay of psychiatry patients using signal detection analysis. Psychiatry research, 210(3), 1211-1218.

Banka, P., Sleeper, L. A., Atz, A. M., Cowley, C. G., Gallagher, D., Gillespie, M. J., et al. (2011). Practice variability and outcomes of coil embolization of aortopulmonary collaterals before Fontan completion: a report from the Pediatric Heart Network Fontan Cross-Sectional Study. American heart journal, 162(1), 125-130.

Breiman, L. (1996). Bagging predictors. Machine learning, 24(2), 123-140.

Chang, J. K., Calligaro, K. D., Lombardi, J. P., & Dougherty, M. J. (2003). Factors that predict prolonged length of stay after aortic surgery. Journal of vascular surgery, 38(2), 335-339.

Chang, K.-C., Tseng, M.-C., Weng, H.-H., Lin, Y.-H., Liou, C.-W., & Tan, T.-Y. (2002). Prediction of length of stay of first-ever ischemic stroke. Stroke, 33(11), 2670-2674.