【摘要】 目的:基于机器学习算法构建老年肌少症临床预测模型,识别早期运动干预的重点指标,为老年肌少症高危人群的早期筛查及精准运动处方制定提供预测工具与理论依据。方法:利用中国健康与养老追踪调查（CHARLS^）数据库2011年全国基线调查数据训练模型,并通过研究团队2021-2024年收集的老年人社区横断面数据进行外部验证。通过结合三种特征工程方法与10折交叉验证的特征选择技术筛选预测变量;运用四种机器学习算法构建预测模型,通过受试者工作特征曲线(_ROC)等指标评估模型性能,并使用SHAP方法对预测指标进行排序;结合限制性立方样条(_RCS)分析可被干预改善的预测指标与老年肌少症患病风险的潜在关联。结果:（₁）经特征筛选确定最大握力平均值(_MGSA)、BMI、性别、四肢骨骼肌量(_ASM)、年龄、身高、平均步行用时(_AWT)等7项预测指标;（₂）_XGBoost算法在外部验证集的综合性能表现较好(AUC=0.974,)且具有较高的敏感性(0.978;（）3⁾SHAP与RCS分析表明,在MGSA、AWT、ASM、BMI四个可被干预改善的预测指标中,MGSA与老年肌少症患病风险呈负向线性关联,BMI与患病风险呈U型关联,两者风险阈值与保护区间分别为24 kg与23～25_kg/_m²。结论:（₁）基于XGBoost算法构建的预测模型,在外部验证中表现出较好的预测性能与稳健性,作为辅助工具可为老年肌少症的社区筛查提供参考;（₂）_MGSA和BMI可作为老年肌少症高危人群运动干预的优先监测与干预重点;二者的风险阈值和保护区间可为精准运动处方的制定提供量化依据。更多还原

【Abstract】 The purpose of the study was to construct a clinical prediction model for sarcopenia in older adults using machine learning algorithms and identify key indicators for early exercise intervention, so as to provide a predictive tool and theoretical evidence for early screening of high-risk groups and the formulation of precise exercise prescriptions for geriatric sarcopenia. The model was trained using the 2011 national baseline data from the China Health and Retirement Longitudinal Study(CHARLS) database and externally validated with community-based cross-sectional data of older adults collected by the research team from 2021 to 2024. Predictive variables were selected via feature selection integrating three feature engineering methods and 10-fold cross-validation. Four machine learning algorithms were adopted to establish prediction models, and model performance was assessed using indicators such as the receiver operating characteristic(ROC) curve. The Shapley Additive Explanations(SHAP) method was used to rank predictive indicators, and restricted cubic spline(RCS) analysis was performed to examine potential associations between modifiable predictive indicators and the risk of geriatric sarcopenia. It could be seen that seven predictive indicators were identified after feature selection: mean maximum grip strength(MGSA), BMI, gender, appendicular skeletal muscle mass(ASM), age, height, and average walking time(AWT). The XGBoost algorithm achieved favorable overall performance in the external validation set(AUC=0.974), with high sensitivity(0.978). SHAP and RCS analyses showed that among the four modifiable indicators(MGSA, AWT, ASM, BMI), MGSA was negatively and linearly correlated with sarcopenia risk, while BMI exhibited a U-shaped association. The risk threshold for MGSA was 24 kg, and the protective range for BMI was 23-25 kg/m~2. The prediction model based on the XGBoost algorithm showed satisfactory predictive performance and robustness in external validation, and can serve as an auxiliary tool for community screening of geriatric sarcopenia. MGSA and BMI could be used as priority monitoring and intervention targets for exercise intervention in high-risk groups; their risk thresholds and protective ranges provided a quantitative basis for developing precise exercise prescriptions.更多还原