Scholarlink

Background <p>Cardiometabolic multimorbidity (CMM) is an increasing public health concern. Physical activity (PA) and prolonged sitting time are recognized as key modifiable lifestyle factors. This study aimed to develop an explainable machine learning (ML) model incorporating PA, sitting time, and other sociodemographic and clinical predictors to classify CMM status among U.S. adults. SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) were applied to improve model interpretability.</p> Methods <p>This cross-sectional study analyzed data from 23,635 U.S. adults in the National Health and Nutrition Examination Survey (NHANES) 2007–2018. CMM was defined as the coexistence of at least two of the following conditions: hypertension, diabetes, stroke, or myocardial infarction. Self-reported total physical activity (MET-hours/week) and daily sitting time were considered the primary exposures. Associations were evaluated using multivariable logistic regression, restricted cubic splines (RCS), and subgroup analyses. After feature selection using LASSO regression and the Boruta algorithm, twelve machine learning models were trained on 70% of the dataset. Models were evaluated using stratified tenfold cross-validation and assessed by AUC, sensitivity, and calibration. Survey sampling weights were applied to descriptive statistics and logistic regression analyses but were not used during machine learning training or evaluation.</p> Results <p>Higher physical activity was associated with lower odds of CMM (adjusted OR for highest vs. lowest quartile = 0.59, 95% CI 0.51–0.68), whereas longer sitting time was associated with higher odds of CMM (adjusted OR = 1.27, 95% CI 1.07–1.49). Restricted cubic spline analysis revealed a nonlinear relationship, with a stronger inverse association at lower physical activity levels (inflection point ≈36 MET-hours/week). The gradient boosting classifier showed the best performance (AUC = 0.84, 95% CI 0.83–0.86) and good calibration (Brier score = 0.116; calibration slope = 0.961). SHAP analysis identified age as the most influential predictor, followed by body mass index, hyperlipidemia, income-to-poverty ratio, and physical activity. LIME provided local explanations illustrating the contribution of individual features.</p> Conclusion <p>This study supports associations between PA, sitting time, and CMM among U.S. adults. SHAP and LIME were used to explain model predictions and identify key contributing features. The explainable ML framework provides a transparent approach for cross-sectional classification and hypothesis-generating, individual-level interpretation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Physical activity and sitting time as correlates of cardiometabolic multimorbidity risk in U.S. adults: an explainable machine learning classification model using SHAP and LIME

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月9日

查看全文
  • Yi Yang,
  • Zhenxiang Guo,
  • Dong Li,
  • Bin Wu,
  • Changnan Xu,
  • Mengbiao Cai,
  • Zhiming Wang

摘要

Background

Cardiometabolic multimorbidity (CMM) is an increasing public health concern. Physical activity (PA) and prolonged sitting time are recognized as key modifiable lifestyle factors. This study aimed to develop an explainable machine learning (ML) model incorporating PA, sitting time, and other sociodemographic and clinical predictors to classify CMM status among U.S. adults. SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) were applied to improve model interpretability.

Methods

This cross-sectional study analyzed data from 23,635 U.S. adults in the National Health and Nutrition Examination Survey (NHANES) 2007–2018. CMM was defined as the coexistence of at least two of the following conditions: hypertension, diabetes, stroke, or myocardial infarction. Self-reported total physical activity (MET-hours/week) and daily sitting time were considered the primary exposures. Associations were evaluated using multivariable logistic regression, restricted cubic splines (RCS), and subgroup analyses. After feature selection using LASSO regression and the Boruta algorithm, twelve machine learning models were trained on 70% of the dataset. Models were evaluated using stratified tenfold cross-validation and assessed by AUC, sensitivity, and calibration. Survey sampling weights were applied to descriptive statistics and logistic regression analyses but were not used during machine learning training or evaluation.

Results

Higher physical activity was associated with lower odds of CMM (adjusted OR for highest vs. lowest quartile = 0.59, 95% CI 0.51–0.68), whereas longer sitting time was associated with higher odds of CMM (adjusted OR = 1.27, 95% CI 1.07–1.49). Restricted cubic spline analysis revealed a nonlinear relationship, with a stronger inverse association at lower physical activity levels (inflection point ≈36 MET-hours/week). The gradient boosting classifier showed the best performance (AUC = 0.84, 95% CI 0.83–0.86) and good calibration (Brier score = 0.116; calibration slope = 0.961). SHAP analysis identified age as the most influential predictor, followed by body mass index, hyperlipidemia, income-to-poverty ratio, and physical activity. LIME provided local explanations illustrating the contribution of individual features.

Conclusion

This study supports associations between PA, sitting time, and CMM among U.S. adults. SHAP and LIME were used to explain model predictions and identify key contributing features. The explainable ML framework provides a transparent approach for cross-sectional classification and hypothesis-generating, individual-level interpretation.