Machine Learning-Based Metabolic Syndrome Identification

Objective: To explore the efficacy of machine learning techniques in identifying Metabolic Syndrome (MetS) and examine the performance of models when applied to target populations with different distributions. Methods: This study utilizes data from the National Health and Nutrition Examination Survey (NHANES) and local physical examinations, where MetS is diagnosed based on the International Diabetes Federation (IDF) standards. We first employ demographic and blood test data from NHANES and predicted MetS using machine learning models (including MLP, Logistic Regression, Random Forest, XGBoost, Catboost, and Multi-layer Perceptron), and then test these predictions on different population data. Results: Models employing 59 features demonstrate commendable performance in the NHANES test set (same population testing), with the MLP model exhibiting the best performance (AUROC= 0.93). Models constructed with 32 features (excluding height, weight, and certain blood test information) still show promising results (MLP AUROC = 0.89). However, when the models are tested on the local physical examination dataset (cross-population testing), there is a substantial decline in performance (MLP AUROC = 0.71). Conclusion: Machine learning techniques can predict MetS on the NHANES dataset with high accuracy. Due to the distribution shift, examined machine learning models perform better in the setting with same population distribution.