目的 构建老年糖尿病患者认知衰弱风险预测模型，为认知衰弱风险筛查提供工具。方法 便利抽样法选取北京市9所医院收治的501例老年糖尿病患者为研究对象。2022年12月至2023年6月收集的390例为模型开发样本，2023年6-8月收集的111例为模型外部验证样本，采用Logistic回归法确定影响因素，基于影响因素构建逻辑回归、随机森林、支持向量机、XGBoost模型，用5折交叉验证法进行内部验证，确定最佳模型，以外部验证样本对最佳模型进行外部验证。结果 筛选出年龄、文化程度、规律运动、跌倒史、使用胰岛素、抑郁6个影响因素用于构建模型，5折交叉验证结果显示，模型的受试者工作特征曲线下面积（area under curve,AUC）值为0.852，准确率为80.77%，灵敏度为77.62%，特异度为81.13%，布里尔（Brier）评分为0.103，综合表现优于其他模型；外部验证结果显示，模型AUC值为0.833，准确率为72.07%，灵敏度为80.00%，特异度为70.83%，Brier评分为0.098，校准曲线与理想曲线一致性较好。结论 逻辑回归模型预测性能最佳，有较好的区分度和校准度，同时具有较好的临床适用性。

Objective To develop and validate a risk prediction model for cognitive frailty in elderly patients with diabetes mellitus,providing a tool for cognitive frailty risk screening.Methods A convenience sample of 501 elderly patients with diabetes mellitus was recruited from 9 hospitals in Beijing.Data from 390 patients collected between December 2022 and June 2023 were used as the model development sample,while data from 111 patients collected between June and August 2023 served as the external validation sample.Logistic regression analysis was used to identify influencing factors,based on which logistic regression,random forest,support vector machine,and XGBoost models were constructed.5-fold cross-validation was used for internal validation to determine the optimal model,which was subsequently validated using the external validation sample.Results 6 influencing factors were identified for model construction: age,education level,regular exercise,history of falls,insulin use,and depression.5-fold cross-validation results showed that the logistic regression model achieved an area under the receiver operating characteristic curve (AUC) of 0.852,an accuracy of 80.77%,a sensitivity of 77.62%,a specificity of 81.13%,and a Brier score of 0.103,demonstrating superior overall performance compared to the other models.External validation results indicated that the logistic regression model had an AUC of 0.833,an accuracy of 72.07%,a sensitivity of 80.00%,a specificity of 70.83%,and a Brier score of 0.098,with good consistency between the calibration curve and the ideal curve.Conclusions The logistic regression model exhibited the best predictive performance,with good discrimination and calibration,as well as strong generalizability and clinical applicability.

R473.59

国家重点研发计划（2023YFC3603905）