机器学习预测中老年高血压风险的研究

Prediction of Hypertension Risk in Middle-Aged and Elderly Populations Using Machine Learning

摘要: 针对当前中老年高血压风险预测研究中模型泛化性不足、缺乏多方法对比与外部验证支撑等局限，本文基于CHARLS中老年队列数据与NHANES独立外部验证数据，构建“双重特征筛选–多模型对比–跨人群验证”的机器学习预测框架，为中老年高血压风险分层筛查提供科学依据。该方法整合人口学、体格测量、血压指标、生活方式与慢性病变量，采用单因素Logistic回归与Lasso回归联合筛选，剔除多重共线性指标后确定13项核心变量；构建Logistic回归、随机森林与XGBoost三种模型，以ROC曲线、AUC等指标评估性能，结合5折交叉验证与SHAP值解析验证模型稳健性与可解释性。结果表明，XGBoost模型综合性能最优，内部AUC达0.942，外部AUC为0.830，内外AUC差值仅0.112，无明显过拟合；年龄、BMI、收缩压、舒张压为核心危险因素。本研究构建的XGBoost模型兼具最优区分效能与跨人群泛化稳定性，可为中老年高血压早期筛查与个体化防控提供科学参考。此外，基于非血压指标的补充分析模型同样展现出良好的早期预警效能，与主模型形成分层互补，为中老年高血压的社区初筛提供了无创化技术方案。

Abstract: This paper presents a machine learning prediction framework of “dual feature selection-multi-model comparison-cross-population validation” to address insufficient generalizability and lack of external validation in current hypertension risk prediction models for middle-aged and elderly populations. Based on CHARLS cohort data and independent NHANES validation data, we integrated multi-dimensional variables, such as demographic, anthropometric, blood pressure, lifestyle, and chronic disease variables, and performed joint feature selection using univariate Logistic regression and Lasso regression, identifying 13 core variables. Three models—Logistic Regression, Random Forest, and XGBoost—were constructed. Performance was evaluated using ROC curves and AUC, and robustness and interpretability were verified using 5-fold cross-validation and SHAP value analysis. Results showed that the XGBoost model exhibited the best overall performance, with an internal AUC of 0.942 and an external AUC of 0.830, a difference of only 0.112, indicating no significant overfitting. Age, BMI, systolic blood pressure, and diastolic blood pressure were identified as the core risk factors. The XGBoost model constructed in this study combines optimal discriminative power with cross-population generalization stability, providing a scientific reference for early screening and individualized prevention of hypertension in middle-aged and elderly populations. Furthermore, the supplementary analysis model based on non-blood pressure indicators also demonstrated good early warning efficacy, forming a hierarchical complement to the main model and providing a non-invasive technical solution for community-based initial screening of hypertension in middle-aged and elderly populations.