文章摘要
基于联邦学习的老年患者术后谵妄预测模型的建立
Establishment of federated learning-based prediction models for postoperative delirium in elderly patients
  
DOI：10.12089/jca.2025.07.001
中文关键词: 联邦学习  机器学习  数据孤岛  数据隐私  术后谵妄  预测模型  老年
英文关键词: Federated learning  Machine learning  Data silos  Data privacy  Postoperative delirium  Prediction model  Aged
基金项目:北京市自然科学基金-海淀原始创新联合基金资助项目（L222100）
作者单位E-mail
王骞 100853,北京市,解放军总医院第一医学中心麻醉科  
张景伟 中国科学院计算技术研究所  
杨晓东 中国科学院计算技术研究所  
宋玉祥 100853,北京市,解放军总医院第一医学中心麻醉科  
张明宇 100853,北京市,解放军总医院第一医学中心麻醉科  
郑兰圆 100853,北京市,解放军总医院第一医学中心麻醉科  
赵红 北京大学人民医院麻醉科  
彭宇明 首都医科大学附属北京天坛医院麻醉科  
陈益强 中国科学院计算技术研究所  
米卫东 100853,北京市,解放军总医院第一医学中心麻醉科  
曹江北 100853,北京市,解放军总医院第一医学中心麻醉科 caojiangbei@301hospital.com.cn 
摘要点击次数: 2456
全文下载次数: 903
中文摘要:
      
目的：采用联邦学习（FL）方法构建老年患者术后谵妄（POD）的预测模型。
方法：收集2020年4月至2022年4月国内3家医院的非心脏、非神经外科手术老年患者围术期数据，将其按照7∶3随机分为训练集和验证集，构建基于传统Logistic回归的集中式学习模型和基于FL的预测模型（FedAvg、FedLSD、Fedprox）。采用受试者工作特征曲线下面积（AUC）、敏感性、特异性、准确率、校准曲线、决策曲线分析（DCA）评估模型的预测效果。采用Shapley加法解释(SHAP)解释模型特征的重要性。
结果：集中式学习模型AUC在训练集中为0.827（95%CI 0.811～0.842），在验证集中为0.726（95%CI 0.687～0.764）。3种基于FL的预测模型（FedAvg、FedLSD和FedProx）的AUC分别为0.710（95%CI 0.671～0.749）、0.708（95%CI 0.669～0.747）、0.710（95%CI 0.671～0.749）。校准曲线和DCA显示，基于FL的预测模型性能接近集中式学习模型。
结论：FL算法能整合多机构数据，基于FL的预测模型表现与集中式学习模型相似，可有效预测老年患者POD的发生。
英文摘要:
      
Objective: To establish prediction models for postoperative delirium (POD) in elderly patients using federated learning (FL).
Methods: Data of elderly patients who underwent non-cardiac and non-neurosurgical operations in three hospitals in China from April 2020 to April 2022 were included. The dataset was randomly divided into a training set (70%) and a validation set (30%). A centralized learning model based on traditional logistic regression and FL-based prediction models (FedAvg, FedLSD, and Fedprox) were then developed. Area under the receiver operating characteristic curve (AUC), sensitivity, specificity, accuracy, calibration curve, and decision curve analysis (DCA) were used to evaluate the prediction effect of the models. Shapley additive explanations (SHAP) were applied to interpret the feature importance of the models.
Results: The centralized learning models achieved AUCs of 0.827 (95% CI 0.811-0.842) in the training set, and 0.726 (95% CI 0.687-0.764) in the validation set. The three FL-based prediction models (FedAvg, FedLSD, and FedProx) yielded AUCs of 0.710 (95% CI 0.671-0.749), 0.708 (95% CI 0.669-0.747), and 0.710 (95% CI 0.671-0.749), respectively. In the calibration curve and DCA, the performance of the FL-based prediction models are close to that of the centralized learning models.
Conclusion: FL can effectively integrate multi-institutional data. The performance of the FL-based prediction models is comparable to centralized learning model, offering a viable approach for POD prediction in elderly patients.
查看全文   查看/发表评论  下载PDF阅读器
关闭