基於複合式解決方法建立預測模型以降低健保核減

健保核減制度是健保局為避免醫療資源的過度浪費以及醫囑開立的合理性訂定的病歷審查機制。對於各醫療院所來說，若病歷在抽審時被核減，則不僅無法獲得健保的補助點數，更需要花費許多時間成本進行申覆，因此建立一個核減預測模型將有助於醫院在抽審前找出潛在的問題病歷並加以修正從而降低健保核減的成本。本研究預期利用極限梯度提升(eXtreme Gradient Boosting)改善過去預測模型在預測結果表現不佳的情況並提升模型預測之準確率。極限梯度提升算法作為集成式學習模型，其優點在於訓練時會產生多個模型，其中各模型都基於先前的訓練結果所產生的誤差進行優化，同時在面對大量資料時仍能保持其運算效率。最後本研究將使用埔里基督教醫院的健保大數據資料並將訓練出來的模型與基於傳統邏輯回歸、隨機森林所建立之預測模型進行比較。

關鍵字

極限梯度提升；主題模型；潛在狄利克雷分布；全民健康保險

並列摘要

The health insurance reduction system is a medical record review mechanism established by the National Health Insurance Bureau to avoid excessive waste of medical resources and the rationality of medical orders. For medical institutions, if the medical records are deducted during the sampling review, not only the institution can not obtain subsidy points for health insurance, but also the institution have to take a lot of time and cost to process the application. Therefore, establishing a deduction prediction model aims hospitals in the sampling process. Identify potentially problematic medical records pre-trial and correct them to reduce the cost of health insurance deduction. This study expects to use eXtreme Gradient Boosting to improve the poor performance of prediction models in the past and improve the accuracy of model predictions. As an integrated learning model, the extreme gradient boosting algorithm has the advantage that multiple models are generated during training, and each model is optimized based on the error generated by the previous training results, while maintaining its operational efficiency in the face of a large amount of data. Finally, this thesis studies the case of the health insurance big data of Puli Christian Hospital and compares the trained model with the prediction model established based on logistic regression and random forest. The results shows that the model offers a great promise in reducing health insurance deduction.