Prediction of metabolic syndrome: A machine learning approach to help primary prevention

To describe the performance of machine learning (ML) applied to predict future metabolic syndrome (MS), and to estimate lifestyle changes effects in MS predictions.

We analyzed data from 17,182 adults attending a checkup program sequentially (37,999 visit pairs) over 17 years. Variables on sociodemographic attributes, clinical, laboratory, and lifestyle characteristics were used to develop ML models to predict MS [logistic regression, linear discriminant analysis, k-nearest neighbors, decision trees, Light Gradient Boosting Machine (LGBM), Extreme Gradient Boosting]. We have tested the effects of changes in lifestyle in MS prediction at individual levels.

All models showed adequate calibration and good discrimination, but the LGBM showed better performance (Sensitivity = 87.8 %, Specificity = 70.2 %, AUC-ROC = 0.86). Causal inference analysis showed that increasing physical activity level and reducing BMI by at least 2 % had an effect of reducing the predicted probability of MS by 3.8 % (95 % CI = -4.8 %; -2.7 %).

ML models based on data from a checkup program showed good performance to predict MS and allowed testing for effects of lifestyle changes in this prediction. External validation is recommended to verify models' ability to identify at-risk individuals, and potentially increase their engagement in preventive measures.