Machine learning-driven predictions of metabolic syndrome in adults: evidence from a Kurdish cohort in Iran.

Background: The prevalence of metabolic syndrome (MetS) is increasing worldwide. Early detection of MetS by valid and available indicators can help to prevent, control and reduce its complications. This study aimed to identify the most important anthropometric, biochemical and nutritional indices for the prediction of MetS using a machine-learning algorithm.

Methods: This study was conducted with 9602 participants from the baseline data of the Ravansar Non-Communicable Disease Cohort (RaNCD), which is part of the PERSIAN study including adults aged 35-65 years. The reference model for MetS was considered according to the International Diabetes Federation (IDF) criteria. The Boruta algorithm and ROC curve analysis were used to select and assess the most important predictors of MetS.

Results: The importance value (IV) for the components of the models predicting MetS was confirmed before the models were implemented. The identified model with components of age, waist circumference (WC), body mass index (BMI), fasting blood sugar (FBS), systolic-diastolic blood pressure (SBP-DBP), triglyceride, hip circumference and an AUC of 0.89 (95% CI 0.88-0.90) for men and 0.86 (95% CI 0.85-0.88) for women was the strongest model for predicting MetS risk. The AUC for the non-invasive model was 0.75 (95% CI 0.74-0.76) in the total population and has good predictive power for MetS risk with the components age, WC, BMI, SBP, DBP.

Conclusions: This study showed that in addition to aggressive models, non-invasive models (anthropometric indices, blood pressure and energy intake) can be a good and convenient screening tool for predicting MetS. The models can be used in clinical diagnosis as well as in research on large populations.