Construction and Validation of a Predictive Model for Left Ventricular Hypertrophy in Chronic Kidney Disease Patients Based on Machine Learning

ZENG Xiaoling; ZHOU Wenkao; LIN Lanxiang; HUANG Yaqing; ZHANG Yuhong; GU Tingyu; SUN Huimin; PENG Yongtiao

doi:10.3969/j.issn.2096-305X.2026.01.013

Abstract：Objective To assess the risk factors for left ventricular hypertrophy(LVH)in chronic kidney disease(CKD)patients using machine learning algorithms and to construct a relevant predictive model.Methods A total of 674 patients with mild to moderate CKD,who visited the Nephrology Department of the Fifth Hospital of Xiamen between June 2019 and June 2024,were enrolled in the study.Left ventricular mass index(LVMI)was calculated,and the patients were divided into the LVH group(238)and the normal group(436).Chi-square test and independent sample t-test were used for group comparison.The"randomForest","gbm","treebag","pls","nnet",and"rjags"packages were used for random forest,gradient boosting,treebag model,partial least squares,neural networks,and Bayesian algorithms to explore the best predictive model for CKD patients'LVH risk.A nomogram was created based on the optimal algorithm's significant risk factors.The"p ROC"package was used to plot the ROC curve for model performance.The"rms"and"ResourceSelection"packages were used for calibration curve plotting,calibration C-index analysis,and Hosmer-Lemeshow test.Results The predictive performance of the algorithms for LVH in CKD patients was ranked from highest to lowest as follows:random forest,treebag model,gradient boosting,Bayesian,partial least squares,and neural networks.The most important risk factors for LVH in the random forest model were gender,serum creatinine,glomerular filtration rate,age,BMI,CKD stage,and hemoglobin levels.The nomogram based on these factors demonstrated a high predictive performance.ROC curve analysis of the internal validation showed AUCs of 0.721 for the training set and 0.747 for the validation set.The calibration curves demonstrated that the predicted probabilities closely matched the actual observed results,with both training and validation sets'calibration curves closely aligning with the ideal diagonal,indicating the model's reliability and accurate prediction of risk.Conclusion This study constructed a predictive model for LVH occurrence in CKD patients using machine learning algorithms,with random forest showing the best performance.The nomogram model showed good predictive power,providing an effective tool for the early prediction of LVH.