Development of metabolic syndrome identification model by integrating Korea national health and nutrition examination survey data with regional particulate matter, air pollutants, and green environmental factors

This study seeks to develop a classification model for metabolic syndrome by integrating environmental characteristics from multiple urban regions. The methodology employs resident health data from the Korea National Health and Nutrition Examination Survey (KNHANES), conducted by the Korea Centers for Disease Control and Prevention, alongside regional data on average air pollution and green spaces sourced from the National Statistical Office. The analysis utilized data from 20,035 adults aged 30 years or older, collected between 2018 and 2022, after excluding cases with missing values. To address class imbalance in metabolic syndrome cases, four correction methods were compared: original data, SMOTE alone, class weighting alone, and a combination of SMOTE and class weighting. Class weighting alone was selected, as it provided the best balance between discriminant power and correction among the four methods compared. The HE_obe variable was excluded from the independent variables, and eight models were constructed for performance comparison: SVM, Logistic Regression, Random Forest, XGBoost, LightGBM, Basic MLP, DeepMLP, and ResNetMLP. The AUC-ROC values for all models ranged from 0.6438 to 0.6664, with DeepMLP achieving the highest discriminative power at 0.6664. However, DeLong and bootstrap tests indicated that DeepMLP did not exhibit a statistically significant performance difference compared to major models such as Random Forest, XGBoost, and LightGBM. Analysis of the contribution of 21 independent variables to the optimal model, using SHAP and permutation importance, identified demographic and socioeconomic factors such as gender, age, and education level as strong contributors to metabolic syndrome identification. Among the environmental factors, green area per capita was associated with a lower likelihood of prevalent metabolic syndrome, contributing to the model to a smaller degree than the demographic and socioeconomic factors. Multicollinearity, indicated by variance inflation factor (VIF) values greater than 5 for NO2 and PM2.5, was identified during construction of the air pollution variable group. As this is a cross-sectional study, the relationship between environmental variables and metabolic syndrome should be interpreted as correlational rather than causal.