文献速递创伤骨科

转型骨科创伤诊疗:利用时间序列分析与机器学习预测手术室需求

Transforming Orthopaedic Trauma Care: Forecasting Operating Room Demand by Harnessing Time-Series Analysis and Machine Learning.
PubMed 文献 · J Bone Joint Surg Am · 2026-07-28

本研究基于2012年1月至2023年12月骨科创伤手术患者数据,构建时间序列与机器学习模型预测每日手术室使用时长。共纳入9637例手术,日均负荷8.17±5.38小时。PMD-SARIMA模型测试集MAE为1.80小时,优于滚动平均和直觉模型,多数模型预测每日负荷是否超过8小时手术室区块的准确率≥85%。

要点

  • PMD-SARIMA模型预测每日骨科创伤手术室使用时长MAE为1.80小时,优于滚动平均(4.15-4.37小时)和直觉模型(4.73小时)。
  • 多数时间序列与机器学习模型预测每日负荷是否超过8小时手术室区块的准确率≥85%(95% PI 77%-96%),滚动平均基线仅49%-57%。
  • 整合环境、时间与系统级外生因素可提升每日骨科创伤手术负荷预测准确性,支持手术室排程与资源分配。

核心结论

  • 整合外生环境、时间与系统级因素与历史创伤模式,可准确预测每日骨科创伤手术负荷,显著优于现有模型。

研究设计与样本

  • 回顾性研究,纳入2012年1月至2023年12月骨科创伤手术患者数据,按时间顺序分为训练集70%、验证集15%、测试集15%,训练时间序列与机器学习模型以最小化MAE。

主要发现

  • 共纳入9637例手术,日均负荷8.17±5.38小时。
  • PMD-SARIMA(手动选择特征)测试集MAE最低为1.80小时,优于7天滚动平均(MAE 4.37小时)、14天滚动平均(MAE 4.23小时)、30天滚动平均(MAE 4.15小时)和直觉模型(MAE 4.73小时)。
  • 多数机器学习与时间序列模型预测每日负荷是否超过8小时手术室区块的准确率≥85%(95% PI 77%至96%),滚动平均基线为49%至57%。

关键数值与推荐

  • 模型评估指标为平均绝对误差(MAE),并在7小时和8小时临床负荷阈值下使用95%预测区间(PI)评估预测准确性。

临床怎么用

  • 该预测框架可帮助创伤中心根据预期每日需求主动分配手术室时间和手术人员,优化运营资源,减少手术延迟。
  • 创伤中心可根据本地情况校准该框架,以支持手术室排程、人员配置和资源分配。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

BACKGROUND: Efficient allocation of operating room (OR) time is critical in trauma centers, where unpredictable volumes lead to wasted resources. Forecasting trauma volume is challenging, but integrating exogenous factors with historical patterns may enable proactive scheduling tailored to daily operative demand. This study develops time-series and machine learning models to predict daily orthopaedic trauma operative caseloads. METHODS: Data for patients undergoing operative management for orthopaedic trauma from January 2012 to December 2023 were aggregated. Models predicted daily caseload in hours using features from hospital records plus regional population and environmental data. Data were split in chronological order into the training set (70%), the validation set (15%), and the test set (15%). Time-series and machine learning models were trained to minimize the mean absolute error (MAE). We evaluated the model's accuracy in predicting 7-hour and 8-hour clinical volume thresholds, using 95% predictive intervals (PIs), and compared its performance with rolling averages and an intuitive model. RESULTS: A total of 9,637 procedures were included. The mean daily caseload was 8.17 ± 5.38 hours. Pattern mining and discovery seasonal autoregressive integrated moving average (PMD-SARIMA) with manually selected features achieved the lowest test MAE of 1.80 hours, outperforming rolling averages over 7 days (MAE, 4.37 hours), 14 days (MAE, 4.23 hours), and 30 days (MAE, 4.15 hours) and intuitive modeling (MAE, 4.73 hours). Most machine learning and time-series models correctly predicted whether daily caseload would exceed an 8-hour OR block ≥85% (95% PI, 77% to 96%) of the time, substantially outperforming rolling average baselines (49% to 57%). CONCLUSIONS: Integrating exogenous environmental, temporal, and system-level factors with historical trauma patterns enabled accurate forecasting of daily orthopaedic trauma caseload, substantially outperforming current models. This framework serves as a template that trauma centers can calibrate to their local context to support OR scheduling, staffing, and resource allocation. CLINICAL RELEVANCE: This predictive framework enables trauma centers to proactively allocate OR time and surgical staff in response to anticipated daily demand. By optimizing these operational resources, hospitals can minimize surgical delays and ensure timely, efficient care delivery for patients with orthopaedic trauma.

来源:J Bone Joint Surg Am,查看原文。版权归原作者所有。

引用

  1. Abbas A, Ravindran D, Simone M, et al. Transforming Orthopaedic Trauma Care: Forecasting Operating Room Demand by Harnessing Time-Series Analysis and Machine Learning. J Bone Joint Surg Am. 2026. PMID: 42520102. doi:10.2106/JBJS.26.00014

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。