利用可解释机器学习识别关键生物标志物预测代谢功能障碍相关脂肪性肝病患者的慢性肾病风险:基于NHANES数据库的研究OA

Objective To construct a machine learning-based prediction model for chronic kidney disease(CKD)in patients with metabolic dysfunction-associated steatotic liver disease(MASLD)and to develop a diagnostic nomogram for guiding clinical management.Methods This study was based on the National Health and Nutrition Examination Survey(NHANES)database(2007 to 2018).Patients with MASLD,defined as hepatic steatosis with at least one cardiometabolic risk factor,were selected as the analysis sample.After excluding patients with chronic renal failure and other conditions,a total of 2 144 MASLD patients were included,and 34 variables(including demographics,clinical characteristics,and laboratory indicators)were collected.Patients were stratified into a chronic kidney disease(CKD)group(n=347)and a non-CKD group(n=1 797)based on CKD status.The total participants were randomly divided into a training set(n=1 501)and an internal validation set(n=643)using stratified sampling at a 7∶3 ratio.A prospective external validation cohort comprising 110 eligible fatty liver disease patients was recruited from our department during January 2024 and October 2025,adhering strictly to predefined inclusion and exclusion criteria Variables were screened using Least Absolute Shrinkage and Selection Operator(Lasso)regression.Eight machine learning algorithms were then employed to construct the CKD risk prediction models:decision trees(DT),extreme gradient boosting(XGBoost),K-nearest neighbors(KNN),Naive Bayes(NB),support vector machine(SVM),single hidden layer neural networks(NNET),light gradient boosting machine(LightGBM),and random forest(RF).Model performance was comprehensively evaluated using receiver operating characteristic(ROC)curves area under the curve(AUC),sensitivity,specificity,calibration curves,and clinical decision curves.Based on key biomarkers identified through SHAP analysis,a diagnostic nomogram was constructed,and its diagnostic accuracy was assessed using ROC curves.Results Using interpretable machine learning algorithms,we identified 10 key biomarkers:SBP,age,diabetes,BUN,BMI,globulin,HDL-C,neutrophil count,uric acid,and gamma-glutamyl transferase.These biomarkers were utilized to construct predictive models through 8 distinct machine learning algorithms.Comprehensive evaluation revealed that the LightGBM model achieved optimal performance,with an AUC of 0.871,a sensitivity of 0.838,a specificity of 0.756,an accuracy of 0.825,an F1 score of 0.889,and a Brier score of 0.091.Calibration curve analysis demonstrated good consistency of the LightGBM-based predictive model across both training and validation sets.Clinical decision curve analysis indicated significant clinical utility of this model in predicting kidney damage risk in patients with MASLD.Based on SHAP values,we developed a diagnostic nomogram with a risk cutoff score of 81.7.This nomogram enabled categorization of patients into high-risk and low-risk subsets for MASLD-related CKD,and exhibited favorable accuracy and robust predictive reliability with an AUC value of 0.816.Conclusion The LightGBM model significantly outperforms other models and demonstrates potential for application in early clinical prediction of CKD risk among MASLD patients.