目的 构建并筛选出评估效能最佳的老年全身麻醉颈椎术后谵妄发生风险模型，为临床干预提供依据。方法 2021年5月至2023年10月，采用便利抽样法选取某院老年全身麻醉颈椎术后患者323例为训练集，将同期另一所医院的同类患者233例为验证集。根据术后3 d是否发生谵妄，将其分为谵妄组和非谵妄组。采用单因素及二元Logistic回归分析确定术后危险因素，采用列线图、决策树CART、随机森林构建风险模型，通过受试者工作曲线（receiver operator characteristic，ROC）分析不同模型的评估效能。结果 单因素分析及二元Logistic回归分析确定，术中输血、术后脑脊液渗漏、糖尿病、手术时长等均为老年全身麻醉颈椎术后谵妄发生的危险因素（均P＜0.05），据此建立列线图模型。决策树CART模型分析显示年龄为根节点，手术时长、糖尿病、术中输血等为子节点。随机森林模型筛选出年龄、手术时长、术中输血、糖尿病等7个重要变量。在训练集和验证集，列线图模型、决策树CART模型、随机森林模型评估的曲线下面积分别为0.888和0.895、0.874 和0.884、0.911和0.911，随机森林模型与决策树CART模型的差异均有统计学意义（均P＜0.05）。结论 三种模型对老年全身麻醉颈椎术后谵妄均有评估价值，随机森林模型效能最优。

Objective To construct and screen a risk model of postoperative delirium in elderly patients undergoing general anesthesia for cervical spine surgery with the best evaluation efficiency，so as to provide a basis for clinical intervention.Methods From May 2021 to October 2023，the convenience sampling method was used to select 323 elderly patients undergoing cervical general anesthesia in a hospital as the training set，and 233 similar patients in another hospital during the same period were selected as the validation set.According to whether delirium occurred 3 days after surgery，the patients were divided into delirium group and non-delirium group.Nomogram，decision tree CART and random forest were used to construct risk models.Receiver operator characteristic(ROC) curve was used to analyze the evaluation efficacy of different models.Results Univariate analysis and binary Logistic regression analysis showed that intraoperative blood transfusion，postoperative cerebrospinal fluid leakage，diabetes，and operation duration were risk factors for delirium after cervical general anesthesia in the elderly (all P＜0.05)，and a nomogram model was established accordingly.Decision tree CART model analysis showed that age was the root node，and operation duration，diabetes，intraoperative blood transfusion were the child nodes.Seven important variables including age，operation duration，intraoperative blood transfusion and diabetes were selected by random forest model.In the training set and validation set，the area under the curve of the nomogram model，the decision tree CART model，and the random forest model were 0.888 and 0.895，0.874 and 0.884，0.911 and 0.911，respectively，and the differences between the random forest model and the decision tree CART model were statistically significant (all P＜0.05).Conclusions All the three models have assessment value for postoperative delirium after general anesthesia for cervical spondylosis in the elderly，and the random forest model has the best performance.

R473；R823

江苏省重点研发计划专项课题（BE2019653）；第九O四医院院管科研课题（MS202111）；江苏省高层次卫生人才“六个一工程” 拔尖人才项目（LGY2019026）