Predictive modeling the probability of suffering from metabolic syndrome using machine learning: A population-based...

Area of Science:

• Cardiology
• Endocrinology
• Preventive Medicine

Background:

• Metabolic syndrome (MetS) prevalence is rising, contributing to cardiovascular disease, cancers, and diabetes.
• The long asymptomatic phase of MetS hinders timely diagnosis and intervention.
• Predictive tools are needed for early MetS risk assessment in clinical practice and daily life.

Purpose of the Study:

• To develop highly effective models for predicting individuals' probability of suffering from MetS.
• To enable timely prediction of MetS risk in the general population.

Main Methods:

• Developed and compared LightGBM (LGBM) and logistic regression (LR) models for MetS prediction.
• Enrolled 8964 adults aged 40-75 years in the REACTION study.
• Created three models: Model 1 (lab tests, lifestyle, anthropometrics), Model 2 (Model 1 minus MetS components), Model 3 (Model 2 minus blood biochemicals).

Main Results:

• LGBM models showed high efficacy: AUCs of 0.993 (Model 1), 0.885 (Model 2), and 0.859 (Model 3).
• LR models had lower AUCs: 0.938, 0.839, and 0.820 respectively.
• Key predictors included gender (women higher risk), height (over 58 years), and resting pulse rate (RPR) (40-62 years).

Conclusions:

• Machine learning models demonstrate effective and accurate prediction of MetS probability.
• Women have a significantly higher risk of MetS.
• Age-specific factors like height (over 58) and RPR (40-62) are vital predictors for MetS.