應用大數據分析於手術室排程與品質指標監測之創新模式

隨著醫療需求增加與醫療資源有限，手術室(Operating Room, OR)之排程效率與流程品質管理已成為醫院營運的重要議題。傳統手術室管理多仰賴人工經驗與事後統計分析，缺乏即時預測與前瞻性決策支援能力，因此如何運用大數據分析與時間序列預測技術提升手術室管理效能，成為值得探討之研究方向。 本研究以中部某醫學中心2021年至2024年之手術室歷史營運資料為研究對象，經去識別化後共蒐集8016筆資料，並以「科別 × 月份」進行彙整分析。研究指標包含手術排程取消率、常規手術排程逾時率、急診手術中主治醫師親自上刀比率、第一台刀準時下刀比率、手術患者進入準備室延遲率、手術室使用率及手術室佔用率等七項流程與品質指標。 研究方法首先建立移動平均法(Moving Average, MA)、ARIMA與 Naïve等傳統時間序列模型作為基準模型，再建構長短期記憶網路(Long Short-Term Memory, LSTM)與時序融合Transformer(Temporal Fusion Transformer, TFT)等深度學習模型進行預測分析，並進一步導入知識蒸餾(Knowledge Distillation, KD)架構，以輕量化時序融合 Transformer(Tiny Temporal Fusion Transformer, TinyTFT)作為教師模型(teacher model)、LSTM作為學生模型(student model)，評估其於模型壓縮與預測效能提升之可行性。模型評估指標包括均方根誤差(Root Mean Squared Error, RMSE)、平均絕對誤差(Mean Absolute Error, MAE)、對稱平均百分比誤差(Symmetric Mean Absolute Percentage Error, sMAPE)及加權平均百分比誤差(Weighted Mean Absolute Percentage Error, WMAPE)。 研究結果顯示，在部分單步預測情境下，傳統統計模型於特定品質指標仍具良好預測能力；LSTM於第一台刀準時下刀比率等關鍵流程效率指標中展現較佳預測表現，顯示其能有效學習手術室流程之時間依賴特性。Tiny TFT則具備較佳之多特徵整合能力與模型可解釋性，可有效學習時間序列之結構資訊。進一步導入知識蒸餾後，KD-LSTM於七項品質指標中有四項取得較低之預測誤差，顯示知識蒸餾可提升部分流程品質指標之預測能力；而LSTM則於手術室使用率與手術室佔用率等資源利用指標展現更佳表現。 本研究進一步將LSTM與KD-LSTM模型導入商業智慧(Business Intelligence, BI)平台進行實務驗證，以第1至第47期資料預測第48期實際結果。外部驗證結果顯示，KD-LSTM於七項品質指標中有四項指標取得較低之平均絕對誤差(MAE)，而LSTM則於急診手術主治醫師親自上刀比率、手術室使用率及手術室佔用率等指標表現較佳。其中，LSTM在手術室使用率與手術室佔用率兩項資源利用指標展現高度準確之預測能力，預測結果與第48期實際值高度一致，並於所有納入分析科別皆取得較低之預測誤差，顯示其具有良好的泛化能力與實務部署潛力。 本研究證實，結合時間序列分析、深度學習模型、知識蒸餾技術及商業智慧(Business Intelligence, BI)視覺化平台，可建立具實務應用價值之手術室品質監測與預測架構。除完成模型效能比較外，本研究亦成功將人工智慧預測模型整合至BI平台，並透過第48期真實資料完成外部驗證，驗證模型於實際臨床管理情境中的應用可行性。 研究成果不僅可作為醫療機構進行手術室流程預警、資源配置及決策支援之依據，亦展現知識蒸餾與深度學習模型於智慧醫療管理之應用潛力。

With increasing healthcare demand and limited medical resources, operating room (OR) scheduling efficiency and quality management have become critical issues in hospital operations. Conventional operating room management relies primarily on human experience and retrospective statistical analysis, lacking real-time forecasting and proactive decision support. Therefore, developing predictive models based on big data analytics and time-series forecasting has become an important approach for improving operating room management. This study analyzed de-identified historical operating room data collected from a medical center in central Taiwan between 2021 and 2024. A total of 8,016 records were aggregated by department × month. Seven operating room quality indicators were investigated, including surgical cancellation rate, scheduled surgery overtime rate, attending surgeon presence rate in emergency surgery, first-case on-time surgery start rate, delayed entry to the preoperative holding area rate, operating room utilization rate, and operating room occupancy rate. Moving Average (MA), AutoRegressive Integrated Moving Average (ARIMA), and Naïve forecasting were used as baseline models. Deep learning models, including Long Short-Term Memory (LSTM) and Temporal Fusion Transformer (TFT), were then developed for prediction. In addition, a Knowledge Distillation (KD) framework was introduced, using a Tiny Temporal Fusion Transformer (TinyTFT) as the teacher model and LSTM as the student model to improve prediction performance through model compression. Model performance was evaluated using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Symmetric Mean Absolute Percentage Error (sMAPE), and Weighted Mean Absolute Percentage Error (WMAPE). The results showed that conventional statistical models remained competitive for certain one-step forecasting tasks. LSTM achieved superior performance in key workflow indicators, particularly the first-case on-time surgery start rate, demonstrating its ability to capture temporal dependencies in operating room processes. After incorporating Knowledge Distillation, KD-LSTM achieved lower prediction errors in four of the seven quality indicators, whereas LSTM performed better in operating room utilization and occupancy rates. For practical validation, the LSTM and KD-LSTM models were integrated into a Business Intelligence (BI) platform to predict the outcomes of Period 48 using data from Periods 1–47. External validation showed that KD-LSTM achieved lower Mean Absolute Error (MAE) in four quality indicators, while LSTM produced better performance in the attending surgeon presence rate in emergency surgery, operating room utilization rate, and operating room occupancy rate. Notably, LSTM consistently generated lower prediction errors for operating room utilization and occupancy across all analyzed departments, demonstrating strong generalization capability and practical deployment potential. This study demonstrates that integrating time-series forecasting, deep learning, Knowledge Distillation, and a Business Intelligence platform provides a practical framework for operating room quality monitoring and prediction. The successful external validation using real-world data further confirms the feasibility of applying the proposed framework to operating room workflow monitoring, resource allocation, and clinical decision support, highlighting its potential for intelligent healthcare management.