醫院收入預測模式之建立-以時間序列模型和機器學習分析某區域教學醫院為例

Abstract

醫院收入的準確預測對於財務管理和運營規劃極為重要，它確保醫院能在不斷變化的醫療環境中持續發展。本研究旨在運用機器學習方法，根據醫院的歷史財務數據建立醫療收入預測模型。 研究數據來源於某區域醫院2016年1月至2023年12月的每月醫療收入資料。我們採用了多種預測模型，包括時間序列模型（ARIMA）、迴歸模型（OLS, Lasso, Ridge）以及樹狀模型（Random Forest, Gradient Boost, XG Boost），進行收入預測的建模與分析。 在醫務總收入的預測上，時間序列模型、Lasso 和 XG Boost 模型展現出較好的預測效果。對於2023年的收入預測，這些模型的平均絕對百分比誤差（MAPE）分別為2.38%，2.52%，和2.39%，而判定係數（R²）則分別為72.3%，76.9%，和73.7%。 對於自費收入的預測 Arima，Lasso 和 XG Boost 模型同樣顯示出良好表現，MAPE分別為3.85%，2.30%，和3.84%，判定係數分別為69.7%，78.2%，和55.7%。 綜合研究結果顯示，三種預測模式均有效，其中Lasso模型在多個方面表現最佳。這些發現為醫院管理者提供了一個強有力的工具，以更好地預測和規劃未來的財務走向。

Accurate forecasting of hospital revenue is crucial for financial management and operational planning, ensuring hospitals can sustain development in a constantly changing medical environment. This study aims to utilize machine learning methods to create a medical revenue forecasting model based on historical financial data from a regional hospital spanning from January 2016 to December 2023. We employed various predictive models, including time series models (ARIMA), regression models (OLS, Lasso, Ridge), and tree-based models (Random Forest, Gradient Boost, XGBoost), to model and analyze revenue forecasts. In terms of total medical revenue prediction, the time series model, Lasso, and XGBoost models demonstrated superior forecasting performance. For the revenue forecasts for the year 2023, these models had Mean Absolute Percentage Errors (MAPE) of 2.38%, 2.52%, and 2.39%, respectively, with determination coefficients (R²) of 72.3%, 76.9%, and 73.7%. For self-paid income predictions, the ARIMA, Lasso, and XGBoost models also showed good performance, with MAPEs of 3.85%, 2.30%, and 3.84%, and determination coefficients of 69.7%, 78.2%, and 55.7%. The comprehensive results of the study indicate that all three predictive models are effective, with the Lasso model performing best in several aspects. These findings provide hospital managers with a powerful tool to better predict and plan future financial trajectories.