Risk Classification and Risk-Adjusted Model in Capitation Payment

[[abstract]]根據健保統計資料，實際醫療支出分配具有偏斜性的特點，加上變數間呈現之相關性，為達估計或預測的準確性，必須區分個人醫療支出之異值性，及選取正確之預測模型。為能區分個人醫療支出之異質性，及建構具可行性之風險調整模式，統計模型均以人為設定方式來區分觀察樣本，以及判別變數間的函數型態。而類神經網路模型中之自我組織映射圖網路具有聚集類似特徵向量的特性，可根據樣本資料之風險特性，將風險程度相似之觀察樣本聚類在一起，對整體樣本資料予以適當分群。此外，倒傳遞網路具有自動學習自變數影響應變數之規則的特色，可不事先假設自變數與應變數間之關係，只需將應變數與自變數之數值資料輸入預測模型，可免除預測模型選取不當的影響。 具體而言，本文目的在於，利用SOM模型物以類聚之特性，將具備偏斜性之健保資料予以適當分類，使風險程度相近的被保人併為同一集合，以避免因醫療支出分配的高度偏斜特性，造成分析結果的偏誤。其次應用BPN模型與統計模型建構個人的風險調整模式，來探討兩種不同分析工具的適當性及可行性。同時，在進行迴歸分析時，所得到的迴歸模式必須符合常態性的假設，因此透過迴歸模式的檢定結果，來探討本文所採用之分群資料是否可再細分，並運用風險值觀念來界定模式的設限門檻，作為區分高醫療風險者的依據，將醫療費用較高的被保險人與其餘被保險人區分開處理，以期能將醫療費用控制於保險人可承擔之範圍內。[[abstract]]According to the annual report of the Bureau of National Health Insurance (NHI), NHI program has faced the financial deficit problem if there is no further premium rate increase. In order to maintain financial balance, individual medical expenditure predictions play an important role. According to previous research, difference in expected medical care cost across individuals may be reflected in the per capita prior utilization. Prior utilization made by individuals depends on the cause-of-ill risk, which can be quantified into different types of risk factors. Thus, individual’s health expenditure can be estimated by adopting different types of risk factors with a selected statistics model. However, data from health insurance are usually highly skewed with a thick-tail distribution. Empirical analysis without data classification or clustering may not obtain the unbiased estimation of individual’s health expenditure. This study intends to develop a risk classification system based on Self-Organizing Map neural network model so that estimation can be made within normal distribution. Together with Back Propagation Neural network model, this study compares the prediction power between the neural network model and the statistic regression model