Identification and optimization of relevant factors for chronic kidney disease in abdominal obesity patients by machine learning methods: insights from NHANES 2005–2018

Abstract

Background The intake of dietary antioxidants and glycolipid metabolism are closely related to chronic kidney disease (CKD), particularly among individuals with abdominal obesity. Nevertheless, the cumulative effect of multiple comorbid risk factors on the progression and complications of CKD remains inadequately characterized. MethodsThis study analyzed data from the National Health and Nutrition Examination Survey (NHANES) dat abase (2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017)(2018), to examine potential factors related to CKD, including glycolipid metabolism, dietary antioxidant intake, and pertinent medical history. To explore the associations between these variables and CKD, the present study used a multivariable-adjusted least absolute shrinkage and selection operator (LASSO) regression model, along with a restricted cubic spline (RCS) model. Furthermore, an optimal predictive model was developed for CKD using ten machine learning algorithms and enhanced model interpretability with the Shapley Additive Explanations (SHAP) method.Results A cohort comprising 8,764 eligible individuals (52% male, including 1,839 CKD patients) with abdominal obesity aged 20-85 years were included. The findings revealed significant positive correlations in patients with abdominal obesity between the presence of CKD and age, a history of heart failure, hypertension, diabetes, elevated lipid accumulation product (LAP) and triglyceride glucose-waist circumference (TyG-WC) levels. Conversely, negative correlations were identified between CKD and variables such as sex, high-density lipoprotein cholesterol (HDL-C) levels, and the composite dietary antioxidant index (CDAI). In parallel, RCS regression analysis revealed significant nonlinear associations between the CDAI, HDL-C, TyG-WC, and CKD among patients with abdominal obesity aged 60-80 years. The development of predictive models demonstrated that the CatBoost model surpassed other models, achieving an accuracy of 86.74% on the validation set. The model's area under the receiver operator curve (AUC) and F1 score were 0.938 and 0.889, respectively. The SHAP values revealed that age was the most significant predictor, followed by diabetes history, hypertension, HDL-C levels, CDAI index, TyG-WC, and LAP.