以健保資料庫對中風患者再復發之預測

Using NHIRDB to Construct a Predict Model for Recurrence of Stroke

尤哲威(Che-Wei You)

指導教授 ： 許明暉

臺北醫學大學/醫學科技學院/醫學資訊研究所/碩士(2015年)

https://doi.org/10.6831/TMU.2015.00080

摘要

動機：根據我國衛生福利部統計，腦中風連續十年是國人主要死因原因排行榜中的前三名，死亡率仍高於日本、新加坡及大多數西歐國家，僅次於惡性腫瘤及心臟疾病。以台灣地區腦中風的死亡率為十萬分之四十八推估，每年約有將近一萬一千三百人死於中風。台灣自從1995年實施全民健康保險，使國人都可享受健保的醫療服務，目前實質納保率已高達99%，且健保醫院特約率達93%。然而，醫療費用提升及人口結構老化，讓健保費用支出不斷的攀升，有效率、有效益及有效能的使用資源是相當的重要。目的：利用類神經網路建構中風患者再復發中風之預測模型。方法：藉由全民健康保險研究資料庫找出中風患者且再次復發之研究對象，並利用類神經網路建構預測模型，預測模型所使用之因素則利用卡方檢定與T檢定其因素與中風復發之間關聯分析之檢定結果。結果：研究樣本共6,520人，其中共有1,609人中風復發；男性共有3518人，當中929人復發中風；女性3002人，當中680人復發中風。而疾病風險預測模型方面，二元邏輯斯迴歸所建立出之模型其Sensitivity可達0.699、Specificity可達0.562、ROCfm 曲線下之面積可達0.721。結論：所建立出的預測模型準確性均較為低落，其原因可能為腦中風疾病因子與中風復發疾病因子有很高的相關性，腦中風後沒有再次復發的病患也多數同樣擁有這些疾病因子，因此在本研究中我們難以評估出哪些疾病因子是決定性造成腦中風復發的主因。在未來的研究應結合更多不同的資料庫，如個人生活習慣資料庫與健康檢查資料庫，進而再尋找可能導致腦中風患者再次復發之風險因子。

關鍵字

全民健康保險研究資料庫 ； 腦中風 ； 再復發 ； 類神經網路 ； 預測模型

並列摘要

Background: According to statistics by Ministry of Health and Welfare (Taiwan) show that stroke is the third leading cause of death in Taiwan for ten years. The mortality rate is still higher than in Japan, Singapore and most of Western Europe. It's second only to cancer and heart disease. In Taiwan, the death rate for stroke forty-eight hundred thousandths estimates of nearly 10,000 each year about 1,300 died of a stroke. Taiwan implemented universal health insurance since 1995, so that people can use health care medical services. The insurance rate has reached 99% and 93% of Privileged hospital. However, medical costs and improve the aging of population structure let health care expenses continue to rise. It's important that how to efficient, efficiency and efficient use of resources. Objective: Using neural networks to construct a predict model that recurrence of stroke. Methods: By Using National Health Insurance Research Database to identify recurrence of stroke and use of neural network to construct prediction model. Predictive models factors is use of the chi-square test and T test to test the relation between factors and recurrent stroke. Results: In this study we include 3,518 males, and 929 people have recurrent stroke; include 3,002 females, and 680 have recurrent stroke; totally include 6,520 people, and 1,609 people have recurrent. The disease risk prediction models, was created by neural network, Sensitivity is 0.802, Specificity is 0.773, and Area Under ROC is 0.873. Conclusion: The model are lower accuracy, the reason maybe is that stroke and stroke recurrent have relation. Patients don’t recurrent that also have most of the same factors. It’s difficult to assess which is the decisive factor of disease caused by the main cause of stroke recurrence. In future we should be combined with other database, such as personal health database and health record database, which can find more risk factors.

並列關鍵字

National Health Insurance Research Database ； Stroke ； Recurrent ； Neural Networks ； Predictive Models

參考文獻

[46]. 黃琮榆、姚立德、陳文輝，類神經網路於肝癌與卵巢癌質譜資料分類之應用，國立台灣科技大學自動化科技研究所，碩士論文2007。

[4]. 謝邦昌，張?坐憛A資料採礦技術應用於全民健保資料庫分析腦中風病患死亡相關因素之研究，數據分析，2013年12月，8卷6期，165-194。

[1]. 邱弘毅，腦中風之現況與流行病學特徵，台灣腦中風學會會訊，第 15 卷第 3 期，2-4。

[45]. 何子銘、盧瑜芬、許家瑋、白健佑、白璐、周雨青、孫建安、Thomas Wetter、林金定、楊燦、朱基銘，運用三種資料探勘方法預測子宮頸癌存活情形之比較，台灣家醫誌，2006，16 卷3 期。

[47]. 鄭中浩、鄭鴻鈞、邱泓文，應用類神經網路預測鼻咽癌病人之五年存活狀態，放射治療與腫瘤學，2013，20(3):191-197。