應用類神經網路模式輔助加護病房重症病人之存活預測

Application of Artificial Neural Network for Predicting Survival of Patients in Intensive Care Unit

陳蕙茹(Hui-Ju Chen)

指導教授 ： 邱泓文

臺北醫學大學/醫學科技學院/醫學資訊研究所/碩士(2013年)

https://doi.org/10.6831/TMU.2013.00069

摘要

研究背景與目的：重症加護病人之存活預測有助於有效分派醫療資源。目前建立與驗證預測模式之相關研究，多來自於國外，不見得適用於國人。本研究擬進行24-48小時存活預測模式分析，以提供醫療人員向病人及家屬說明安寧緩和計畫時間點之輔助參考。
方法：由某醫學中心外科加護病房臨床資訊系統( IntelliVue Clinical Information Portfolio, ICIP)資料庫進行回溯性資料分析，擷取2011年-2012年某醫學中心外科加護病房轉入病人(n=2593)之資料。以類神經網路（Artificial Neural Network, ANN）資料探勘（Data mining）演算法建構存活預測模型，變項包括基本人口學資料(年齡、性別)、入院來源、就醫科別、住加護病房日數、轉入加護病房原因及住院24小時及48小時之評估指標-急性生理及慢性健康評估評分(Acute Physiology and Chronic Health Evaluation, APACHE Π)、TISS ( Therapeutic intervention scoring system)及Glasgow昏迷指數(Glasgow coma scale, GCS)。模型的評估指標為準確率（accuracy, ACC）、ROC曲線下的面積（area under the ROC curve, AUC）。輸出變數：存活為1，死亡(含病危自動出院)為0。
結果： 2593位住加護病房病人中，共有224人(8.6%)死亡。24小時及48小時之APACHE II、TISS、GCS三指標分別加入其他臨床因子之預測力佳，其測試組AUC分別為：APACHE II：0.914/0.901；TISS：0.808/0.858；GCS：0.903/0.928。納入APACHE II和GCS之模式的預測又優於納入TISS的模式。進一步以APACHE II加入臨床因子的模式為基礎，分別比較合併加入TISS或(及) GCS之三組模式表現，結果以第48小時之APACHE II+TISS
+GCS模式的預測力優於其他指標組合模式。
結論：研究證實類神經網路於建構存活預測模式準確度佳且臨床可用性高，建議未來可持續開發並運用於臨床。

關鍵字

加護病房 ； 預測 ； 病危自動出院 ； 類神經網路 ； 資料探勘

並列摘要

Background and Purpose: The survival prediction models are effective tools for allocating the limited healthcare resources. Prediction models established based on western populations may not be valid in Asia populations. The aim of this study was to construct survival prediction models among patients admitted to an Intensive Care Unit (ICU).
Methods: Clinical data of 2593 patients admitted to neuro-surgery/surgery ICU of a medical center during 2011-2012 were retrieved from the clinical information system (IntelliVue Clinical Information Portfolio, ICIP) of the ICU. Various regression models were constructed using data mining of artificial neural network, and the area under the ROC curve (AUC), accuracy and positive predictive value (PPV) were estimated to assess the model performance. The models included APACHEⅡ score, TISS and/or GCS at 24hrs stay and 48hrs stay in ICU. Other prognostic factors, including age, gender, days of ICU stay, causes of transfer to ICU, length of stay, were then added to the models to assess whether the model performance improves. The best predicting model was established using MLP; the output variable was survival (0, survive; 1, death, against advise discharge, and critical discharge).
Results: Of 2593 patients, 224 (8.6%) died. The survival prediction models constructed using APACHE II, TISS or GCS alone were improved after adding other prognostic factors to the models. After adding the prognostic factors, the AUC, estimated by using the test dataset, for the model of APACHE II recorded at 24-hour and 48-hour were 0.914 and 0.901, respectively. The corresponding figures for TISS models were 0.808 and 0.858, and for GCS models were 0.903and 0.928. APACHE II or GCS showed better prediction ability than TISS. We further assessed whether the prediction ability of the models differs after adding TISS and (or) GCS to models included APACHE II and other prognostic factors. The 48-hour APACHE II +TISS+GCS model showed the best prediction ability。
Conclusion: This study revealed the accuracy and clinical utility of survival prediction models constructed using artificial neural network, suggesting the importance of further development of the tools and validation for use in clinical practice.

並列關鍵字

intensive care unit ； prediction models ； survival ； artificial neural network ； data mining

參考文獻

Buchman, T. G., Kubos, K. L., Seidler, A. J., & Siegforth, M. J. (1994). A comparison of statistical and connectionist models for the prediction of chronicity in a surgical intensive care unit. Crit Care Med, 22(5), 750-762.

Castillo-Lorente, E., Rivera-Fernandez, R., Rodriguez-Elvira, M., & Vazquez-Mata, G. (2000). Tiss 76 and Tiss 28: correlation of two therapeutic activity indices on a Spanish multicenter ICU database. Intensive Care Medicine, 26(1), 57-61.

Clermont, G., Angus, D. C., DiRusso, S. M., Griffin, M., & Linde-Zwirble, W. T. (2001). Predicting hospital mortality for patients in the intensive care unit: a comparison of artificial neural networks with logistic regression models. Crit Care Med, 29(2), 291-296.

Cullen, D. J., Civetta, J. M., Briggs, B. A., & Ferrara, L. C. (1974). Therapeutic intervention scoring system: a method for quantitative comparison of patient care. Crit Care Med, 2(2), 57-60.

Doi, K., Negishi, K., Ishizu, T., Katagiri, D., Fujita, T., Matsubara, T., . . . Noiri, E. (2011). Evaluation of new acute kidney injury biomarkers in a mixed intensive care unit. Crit Care Med, 39(11), 2464-2469.