The diagnostic role of machine learning models related to visceral fat with sex-specific distribution in subtypes of primary aldosteronism

Figures

• 

  Figure 1. Workflow and predictive model setup procedure. (A) The overall study workflow. (B) The predictive model setup procedure. Lasso: least absolute shrinkage and selection operator; CV: cross-validation; XGBoost: extreme gradient boosting; SVM: support vector machine; ROC: receiver operating characteristic curve.

• 

  Figure 2. Visceral fat and epicardial adipose tissue volume measurements from a CT scan. (A) The visceral fat area (VFA) was measured at the third lumbar vertebra. Within this region of interest, pixels between -190 and -30 Hounsfield Units were accounted as fat. (B) After 3-dimensional reconstruction, EAT volume was calculated by summation of all pixels accounted as fat.

• 

  Figure 3. Comparison of body composition of PA patients according to subtype and sex. Comparison of VFA (A), VFA/BMI (B), EAT (C), right RSF (D) between unilateral and bilateral PA patients for each sex. *P < 0.05, **P < 0.01, and ***P < 0.001. P values were calculated using Mann-Whiteny test or Mann-Whitney U test. PA: primary aldosteronism; VFA: visceral fat area; BMI: body mass index; EAT: epicardial adipose tissue; RSF: renal sinus fat.

• 

  Figure 4. Validation set performance for the models across each patient cohort. ROC Curves Comparison of Multiple Machine Learning Models in Whole group (A), male group (B) and female group (C). Performance of XGBoost model in different groups (D). LASSO: least absolute shrinkage and selection operator; XGBoost: extreme gradient boosting; SVM: support vector machine; ROC: receiver operating characteristic curve.

• 

  Figure 5. Waterfall Plots of SHAP Values Exploring the Important Expectation. (A) Features ranked by their mean absolute SHAP values in different primary aldosteronism (PA) groups. The extracted feature after LASSO regression for primary aldosteronism of 1 sample in the male group (B). SBP and VFA played important predictive roles in the male group, the SBP=165mmHg increased the prediction probability by 0.131, while VFA=82cm2 increased it by 0.106. The final model output indicated a probability of 0.986 for the “unilateral group”. (C) The extracted feature after LASSO regression for primary aldosteronism of all samples in the male group. SHAP: shapley additive explanations; LASSO: least absolute shrinkage and selection operator; SBP: systolic blood pressure; VFA: visceral fat area; Lowest K: the lowest potassium during hospitalization; DRC: direct renin concentration; SIT: saline infusion test; CT: computed tomography; PAC: plasma aldosterone concentration; TG: triglycerides; SF: subcutaneous fat.