Effect of visceral fat on onset of metabolic syndrome

Abstract

This study analysed the effects of visceral fat on metabolic syndrome (MetS) and developed an algorithm to predict its onset using health examination data from the Iwaki Health Promotion Project in Japan. The dataset included 213 cases of MetS onset within three years and 1320 non-onset cases. The data was split into training and test sets with an 8:2 ratio. In the training set, the MetS onset group had significantly higher visceral fat area than the non-onset group (p < 0.00001). A cut-off value of 82 cm2 for the visceral fat area was determined, with an AUC of 0.86. Additionally, a machine learning algorithm utilizing seven non-invasive factors, including visceral fat, achieved high accuracy with a five-fold cross-validation AUC of 0.90 in the training set and 0.88 in the test set. Visceral fat was identified as the main factor, as supported by the SHAP value. In conclusion, this study found visceral fat to be crucial in predicting the onset of MetS, and a high-accuracy onset prediction algorithm based on non-invasive parameters, including visceral fat, was developed.

Keywords: Early detection; Machine learning; Metabolic syndrome (MetS); Onset prediction models; SHapley additive exPlanations (SHAP); Visceral fat.

Fig. 1

Research steps employed in this study.

Fig. 2

Effect of baseline visceral fat on MetS onset (a) VFA values at baseline in the MetS onset group (n = 169) and non-onset group (n = 1,058), The p value indicates the difference between the MetS onset and non-onset groups using the Mann–Whitney U test (corrected for ties); (b) ROC curve of baseline VFA that determines MetS onset risk. The cut-off value was calculated using the Youden index.

Fig. 3

Comparison of AUC during cross-validation of the two models for MetS onset prediction. Model 3 was constructed using Elastic-Net, and the input feature was VFA. Model 2 was constructed using LightGBM, and the input features were VFA, BMI, number of cigarettes smoked, sex, age, SBP, and DBP. The AUC difference was calculated using the DeLong’s test.

Fig. 4

Feature importance based on the SHAP value.