比較決策樹演算法與邏輯迴歸模式評估事故傷害就醫之相關因子

Comparison of Decision Tree Algorithms and Logistic Regression Models in the Assessment of Factors Associated with Recorded Injuries

10.6200/TCMJ.2012.9.1.04

俞依良(I-Liang Yu)；楊南屏(Nan-Ping Yang)；詹前隆(Chien-Lung Chan)

傷害 ； 邏輯式迴歸 ； 決策樹演算法 ； 全民健康保險資料 ； 台灣國民健康訪問調查 ； Injury ； Logistic regression model ； Decision tree algorithm ； National Health Insurance (NHI) ； National Health Interview Survey (NHIS)

北市醫學雜誌

9卷1期（2012 / 03 / 01）

30 - 44

繁體中文

目的：以傷害就醫族群之特性分析及危險因子探討為實證基礎，比較兩種量化分析方法，應用於臨床資料研究之異同處。方法：以2002年台灣地區12歲以上民眾為抽樣母群體，且能連結當年度全民健康保險資料就醫記錄之個人為研究對象，回溯連結其2001年台灣國民健康訪問調查之內容，以邏輯式迴歸模式及決策樹演算法分析傷害就醫之相關資料。結果：「邏輯式迴歸」顯示：在過去一年中身體質量指數越高、生理健康分數越低、神經肌肉功能良好、循環功能不佳、未使用醫療資源利用者，年齡低者、男性、學歷低者容易造成因傷害而就醫。而「決策樹演算法」於四種模式下所得到綜合規則：身體質量指數越高、生理健康分數越低、未使用醫療資源利用者、年齡低者、學歷低者，在過去一年中容易因傷害而就醫，重要性排名以醫療資源利用與否最重要，其次為年齡與生理健康分數。結論：於臨床流行病學領域，可使用傳統邏輯式迴歸發現較多之危險因子，合併使用決策樹演算法之重要性排序，作為預防醫學資源分配之決策考量。

Objective: The aim of this study was to compare the various different quantitative analytical methods used when studying treated injury epidemiology in order to identify related risk factors, including the geographic characteristics of the injured population, their personal health status and their daily health behavior. Methods: The present study was a retrospective case-control study based on randomly sampled subjects with available medical records who were registered with the Taiwan National Health Insurance (NHI) system in 2002. These enrolled cases were traced to identify their detailed information as provided in household questionnaires and the individual questionnaires during the Taiwan National Health Interview Survey (NHIS) of 2001. Results: The multivariate logistic regression model revealed eight major factors that were associated with medical visits due to injury. However, when decision tree analysis was used, this revealed five major factors that were associated with injury and, furthermore, this approach also ranked the importance of these associated factors based on past use of medical resources, followed by age and physical functioning score. Conclusions: These findings are likely to be useful to healthy policymakers are considering the re-distribution the resources in relation to injury prevention programs in the workplace or in society.