从输入中洞察:利用机器学习预测提升翻修全关节置换术资源配置
本研究比较了行政数据库与机构数据库的机器学习模型在预测翻修全膝关节置换术(rTKA)和翻修全髋关节置换术(rTHA)手术时长、住院时间和30天再入院方面的准确性,并识别重要预测特征。国家数据库纳入22,851例rTKA和14,262例rTHA患者,机构数据库纳入1,025例rTKA和703例rTHA患者。结论认为人工神经网络模型性能因数据集而异,但所有模型均优于使用历史平均值。
要点
- 机构数据库的人工神经网络对rTKA手术时长的30分钟缓冲准确率为76.2%,国家数据库为45.4%。
- 国家数据库的人工神经网络对rTKA住院时间的2天缓冲准确率为81.8%,机构数据库为67.7%。
- 国家数据库预测30天再入院的曲线下面积在rTKA为0.593,在rTHA为0.590。
核心结论
- 人工神经网络模型的性能因数据集不同而存在显著差异,但所有模型均优于使用历史平均值。
- 未来工作应考虑基于已识别的重要特征,使用准确的关节置换术专用数据集。
研究设计与样本
- 查询2014年至2019年的国家质量改进数据库和2012年至2022年的本地机构关节置换术数据库,纳入rTKA和rTHA患者。
- 数据集独立分为训练集、验证集和测试集,并按年份进行归一化处理。
- 为两种手术和每个结局分别创建人工神经网络(ANN),并将其性能与多变量回归模型进行比较。
- 使用缓冲准确率(BA)比较数据集之间的模型,并使用Shapley加性解释值获取ANN的特征重要性。
主要发现
- 国家数据库纳入22,851例rTKA和14,262例rTHA患者,机构数据库纳入1,025例rTKA和703例rTHA患者。
- 对于手术时长,机构ANN优于国家ANN:rTKA的30分钟缓冲准确率为76.2% vs 45.4%,rTHA为55.4% vs 43.1%。
- 对于住院时间,国家ANN的2天缓冲准确率优于机构ANN:rTKA为81.8% vs 67.7%,rTHA为71.4% vs 48.2%。
- 使用国家数据库预测30天再入院,rTKA和rTHA的曲线下面积分别为0.593和0.590。
关键数值与推荐
- rTKA手术时长30分钟缓冲准确率:机构ANN 76.2% vs 国家ANN 45.4%。
- rTHA手术时长30分钟缓冲准确率:机构ANN 55.4% vs 国家ANN 43.1%。
- rTKA住院时间2天缓冲准确率:国家ANN 81.8% vs 机构ANN 67.7%。
- rTHA住院时间2天缓冲准确率:国家ANN 71.4% vs 机构ANN 48.2%。
- 30天再入院预测曲线下面积:rTKA 0.593,rTHA 0.590(国家数据库)。
局限性
- 原文未报告局限性。
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: Revision total knee arthroplasty (rTKA) and revision total hip arthroplasty (rTHA) are among the most resource-intensive orthopaedic procedures. The primary aim of this study was to compare the accuracy of machine learning models between administrative and institutional datasets for predicting duration of surgery, length of stay, and 30-day hospital readmission for rTKA and rTHA based on preoperative factors and identify significant predictive features. METHODS: A national quality improvement database was queried from 2014 to 2019, and a local institutional arthroplasty database was queried from 2012 to 2022 for rTKAs and rTHAs. Datasets were independently split into training, validation, and testing and normalized based on year. Artificial neural networks (ANNs) for both procedures and each outcome were created, and their performance was compared to multivariable regression models. Models were compared between datasets using buffer accuracy (BA). Feature importance of ANNs was retrieved using Shapley Additive exPlanations values. RESULTS: A total of 22,851 and 1,025 rTKA and 14,262 and 703 rTHA patients were included from the national and institutional datasets, respectively. For duration of surgery, the institutional ANNs outperformed the national ANNs with a 76.2 versus 45.4%, and 55.4 versus 43.1% 30-minute BA for rTKA and rTHA, respectively. For length of stay, the national ANNs yielded a superior 2-day BA than the institutional ANNs of 81.8 versus 67.7%, and 71.4 versus 48.2% for rTKA and rTHA, respectively. The 30-day readmission prediction using the national database had area under the curve scores of 0.593 and 0.590 for rTKA and rTHA, respectively. CONCLUSIONS: The performance of ANN models varied significantly by dataset, but all models were superior to using historic averages. Future work should consider using accurate, arthroplasty-specific datasets based on the important features identified. LEVEL OF EVIDENCE: III.
引用
- Lex JR, Entezari B, Abbas A, et al. Insights From Inputs: Enhancing Revision Total Joint Arthroplasty Resource Allocation With Machine Learning Prediction. J Arthroplasty. 2025. PMID: 40339932. doi:10.1016/j.arth.2025.04.072
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。