摘要
目的 基于中西医双维特征,借助于机器学习技术构建老年高血压轻度认知障碍(MCI)预测模型.方法 收集分析2020 年 1 月至 2023 年 3 月院内就诊的 502 例 60 岁以上原发性高血压患者数据,按照 7∶3 的比例随机划分为训练集和验证集,并将其分为认知障碍组(n=104)和认知正常组(n=398).运用LASSO回归分析对临床指标数据降维分析,筛选出核心预测因子.采用logistic回归、XGBoost、AdaBoost、SVM、GNB及MLP 6 种机器学习算法构建模型,绘制ROC曲线比较6 种模型的AUC、准确度、敏感度、特异度及F1 分数.SHAP模型揭示预测因子的特征重要性.结果 腰臀比、气郁质、年龄、总胆固醇、痰湿质、湿热质、气虚质和空腹血糖是老年高血压患者早期MCI的核心预测因子.XGBoost模型AUC、准确度、灵敏度、特异度、F1 分数分别为 0.938、0.885、0.846、0.896、0.755,均优于其他算法模型.结论 基于腰臀比、年龄、总胆固醇、气郁质、痰湿质、湿热质、气虚质和空腹血糖构建的XGBoost模型预测性能最优,可为临床老年高血压群体中MCI风险的早期辨识和诊治决策提供参考依据.
Abstract
OBJECTIVE To construct a prediction model of mild cognitive impairment(MCI)in elderly patients with hyperten-sion based on the bi-dimensional features of Chinese and western medicine with the help of machine learning(ML).METHODS The clinical data of 502 patients over 60 years old with essential hypertension treated in hospital from January 2020 to March 2023 were collected and analyzed,randomly divided into training set and verification set according to a ratio of 7∶3,and divided into cognitive impairment group(n=104)and cognitive normal group(n=398).LASSO regression analysis was used to reduce the dimension of clin-ical indicator data and screen out the core predictors.Six ML algorithms,logistic regression,XGBoost,AdaBoost,SVM,GNB,and MLP were used to construct the models,and ROC curves were plotted to compare the AUC,accuracy,sensitivity,specificity,and F1 scores of the 6 models.SHAP models were adopted to reveal the characteristic importance of predictors.RESULTS Waist-hip ratio,qi depression,age,total cholesterol,phlegm-dampness,damp-heat,qi deficiency and fasting blood glucose were the core predictors of early MCI in elderly hypertensive patients.The AUC,accuracy,sensitivity,specificity,and F1 scores of the XGBoost model were 0.938,0.885,0.846,0.896,and 0.755 respectively,which were superior to those of other algorithmic models.CONCLUSION The XGBoost model constructed on the basis of waist-to-hip ratio,qi depression,age,total cholesterol,phlegm-dampness,damp-heat,qi deficiency and fasting blood glucose has the best prediction performance,which can provide a reference basis for early identifi-cation of MCI risk and diagnostic and therapeutic decision-making in the clinical elderly hypertensive population.
NETL
NSTL
万方数据