醫療數據挖掘對重症監護室的結局預測

Data Mining in Healthcare for Outcome Predictions in the Intensive Care Unit

梁茂傑(Julian Caesar Blanco)

指導教授 ： 賴飛羆

國立臺灣大學/電機資訊學院/資訊工程學研究所/碩士(2018年)

https://doi.org/10.6342/NTU201704328

摘要

重症監護病房的死亡率可能是許多因素造成的，因此確保其影響力是至關重要的。在過去的研究中，兩個重要特徵是HCT和WBC。隨著ICU部門對高質量醫療保健進步的需求迅速上升，一些研究的目的是確定更多的重要特徵及其臨界值。歐洲重症監護醫學會，世界強化和重症醫學大會以及其他地區和醫療保健機構都在尋求通過先進的方法進行創新。其中一種方法是通過引入數據和算法來幫助解密有助於臨床護理的模式。這些組織的目標是降低成本，同時促進治療指南，從而提高重症監護病房的質量。 2014年在台灣進行的一項研究發現，整個ICU的死亡率降低了20.2％，與之前所說的機構一樣，研究發現，ICU在維護和改善由於技術成本而引起的質量保健方面面臨挑戰。
隨著ICU支出差異的上升，控制和降低成本至關重要。 ICU領域的數據集可以協助這一努力。來自當地醫院ICU的數據集，包含各種實驗室和臨床特徵的組合，用算法進行分析正在研究中。這項研究是一種新穎的方法，已成功進行。研究包括特徵選擇方法和對重量和重要性的調整。在數據集應用中有效的方法。最複雜的數據集由數百到數千個變量組成。在這項研究中，一個較小的數據集進行了研究，但其方法是廣泛可用的。特徵選擇的目的是提高預測指標的預測性能，建立更快，更經濟的預測指標。因此，特徵排序，多種單變量和單變量選擇方法以及特徵驗證評估方法。

關鍵字

機器學習 ； 數據挖掘 ； 功能選擇

並列摘要

ABSTRACT
Intensive care unit mortality can be the result of many factors so determination of the influential features is critical. In previous studies, two important features have been the HCT and WBC. And with the rapid rise in the need for securing advancements in quality healthcare for the ICU department, some research aims are to determine additional important features as well as their cutoffs. The European Society of Intensive Care Medicine, World Congress of Intensive and Critical Care Medicine as well as other regional and international healthcare organizations have sought methods for innovation through progressiveness. One of these methods is through the introduction of data and the algorithms to assist in deciphering patterns that are helpful in clinical care. It is the aim of these organizations to diminish the cost while also promoting treatment guidelines, which in turn improve the quality in the intensive care unit. In 2014, a study in Taiwan found that the overall ICU mortality rate was 20.2% and like the institutions mentioned before, the study found that the ICU has challenges in its maintaining and improving the quality care due to the cost of technology (Cheng, Lu et al. 2014).
With the ICU expenditure difference rising, controlling and reducing the cost is of prime importance. Data mining in the field of ICU can assist with this endeavor. Datasets from the local Hospital ICU containing combinations of various laboratory and clinical features to analyze with algorithms are being studied. In this study, a novel approach was successfully conducted. The research is comprised of methods for feature selection and their adjustments for weight and significance. The feature selection in particular is the focus of much research in some studies. Its research end-goal is to have effective methods in its application with datasets. The most complex datasets are made-up of hundreds to thousands of variables. In this study a smaller dataset is studied, but its methodology is widely applicable. The purpose of feature selection is to improve prediction performance by identifying predictors (features) that are additive or replacing them with better predictors. This establishes predictors that give rise to faster and more cost-effective procedures. As such, feature ranking, multivariate and univariate methods for selection as well as feature validity methods are important (Guyon and Elisseeff 2003).

並列關鍵字

Machine Learning ； Data Mining ； Feature Selection

參考文獻

[1] Adrogué, H. J. and N. E. Madias (1998). "Management of life-threatening acid–base disorders." New England Journal of Medicine 338(1): 26-34.

[2] Arora, P. and S. Varshney (2016). "Analysis of K-means and K-medoids algorithm for big data." Procedia Computer Science 78: 507-512.

[3] Cheng, K.-C., et al. (2014). "ICU service in Taiwan." Journal of intensive care 2(1): 8.

[4] Chung, J., et al. (2012). "Heme metabolism and erythropoiesis." Current opinion in hematology 19(3): 156.

[7] Manley, J. L., et al. (1996). "SR proteins and splicing control 1569." Genes & development.