以機器學習法建立Tw-DRGs住院病人之醫療資源利用預測模型

A Model for Predicting Tw-DRGs Medical Resource Utilization

目的：健保署全面（1~5階）實施Tw-DRGs支付制度將直接影響醫院醫療品質與醫療效率。因此本研究導入機器學習方法運算醫院大數據實務資料，以建立住院中資源利用預測模型，提供Tw-DRGs全面上線之疾病管理輔助決策參考。
材料與方法：本研究取自高雄市某區域醫院資料庫，資料區間為2019年月1至2023年7月的特徵數據（n=65,446），隨機分配至訓練集（80%）、驗證集（10%）和測試集（10%），並以33個特徵變數納入機器學習演算法建立「Tw-DRGs醫療資源利用預測模型」。
結果：本研究發現，運用機器學習後，以XGBoost演算法建立預測模型效果最佳，其測試集之混淆矩陣ROC面積為91.09%，預測準確度達83.83%，精確率有82.61%、有80.84%的正收益差個案被成功預測為正收益差個案（召回率）、F1分數達81.59%之結果。年齡、性別、Major Diagnostic Category（MDC）、主診斷、主處置、住院天數為重要預測特徵變數；非青壯年、住院天數愈長、有合併併發症及MDC1（神經系統）相對MDC6（消化系統）虧損機率更高。
結論：經過預測模型的建立和持續優化，機器學習所建立的住院病人醫療資源利用預測模型，有助於在醫療效率方面實現更具預測性、更具精準性的疾病管理模式。

Objective: The implementation of Taiwan’s diagnosis-related group payment system (Tw-DRGs) by the National Health Insurance Administration for hospital inpatient expenses necessitates the use of big data analytics in health-care resource management. This study was conducted to develop a machine learning–based predictive model for inpatient resource utilization, aiming to provide decision support for managing Tw-DRGs.
Methods: Relevant data (January 2019 to July 2023; 65,446 patients) were collected from a regional hospital in Kaohsiung City. The data were randomly allocated to a training set (80%), a validation set (10%), and a testing set (10%). A machine learning algorithm and 33 features were used to construct the Tw-DRGs Medical Resource Utilization Predictive Model.
Results: The use of the deep learning algorithm XGBoost ensured outstanding performance by the predictive model. The model achieved an area under the curve value of 91.09%, an accuracy score of 83.83, a precision score of 82.61%, a recall score of 80.84%, and an F1 score of 81.59%. Key predictive features included age, sex, major diagnostic category, primary diagnosis, main procedure, and hospitalization duration.
Conclusion: The development of a machine learning–based predictive model for inpatient medical resource utilization markedly advances the management of diseases within the Tw-DRGs framework.