Predicting metabolic syndrome risk in Chinese adults using integrated features of movement behaviors and dietary...

China is undergoing a rapid nutritional and epidemiological transition. Traditional epidemiological approaches often examine diet and physical activity in isolation, failing to capture the synergistic and non-linear effects of these lifestyle behaviors. This study aimed to develop an explainable machine learning (ML) framework to predict Metabolic Syndrome (MetS) risk using strictly non-clinical lifestyle and environmental features, thereby avoiding the diagnostic circularity prevalent in existing models.

Data were derived from the 2015 wave of the China Health and Nutrition Survey (CHNS), involving 6,854 adults. To eliminate outcome leakage, we excluded all clinical diagnostic components (waist circumference, blood pressure, glucose, and lipids) from the predictor set. We integrated features of movement behaviors (occupational, domestic, transport, and leisure domains) and 3-day dietary recalls. Five ML algorithms were evaluated using a stratified training-testing split followed by nested 10-fold cross-validation within the training set. Model interpretability was achieved using SHapley Additive eXPlanations (SHAP).

The prevalence of MetS was 28.4%. After removing clinical predictors to ensure etiologic transparency, the XGBoost model achieved a robust predictive performance (AUC: 0.792 [95% CI: 0.771-0.813]), significantly outperforming the traditional Logistic Regression baseline (AUC: 0.714). Feature importance analysis identified urbanization index, age, and sedentary time (hours/day) as the top contributors to the model's prediction. SHAP dependence plots revealed non-linear thresholds: predicted risk contribution was significantly higher sharply after exceeding 4 h/day of sedentary time and an urbanization score of 65. Crucially, we observed a conditional association where high carbohydrate intake was associated with increased predicted risk primarily in individuals with low physical activity (< 50 MET-h/week), whereas high physical activity (> 150 MET-h/week) was associated with an attenuated predicted risk.

Machine learning provides a robust and complementary approach for capturing the complex, multi-dimensional nature of MetS risk factors. The findings highlight a conditional association between carbohydrate intake and sedentary behavior associated with metabolic risk in China's urbanizing population.