Development and validation of a machine learning-based risk prediction model for hyperkalemia in patients with chronic kidney disease

Abstract

To develop an optimal predictive model for hyperkalemia in patients with chronic kidney disease (CKD). Clinical data of CKD patients were collected from Shandong Provincial Hospital Affiliated to Shandong First Medical University between January 2017 and December 2023, including 343 hyperkalemia cases and 713 cases with normal potassium levels. The data were divided into training and test sets at a 7:3 ratio. Important features were screened via univariate analysis, collinearity diagnosis, and LASSO regression, identifying 20 feature variables for hyperkalemia. Five machine learning models were established: logistic regression (LR), decision tree (DT), gradient boosting machine (GBM), support vector machine (SVM), and K-nearest neighbors (KNN). Models were compared using the area under the curve (AUC), Brier score, calibration curve, decision curve analysis (DCA), and overfitting control. The LR model was identified as the optimal model, showing excellent performance in predicting hyperkalemia. The AUCs were 0.899 (training set) and 0.868 (test set), with corresponding F1 scores of 0.819 and 0.725. Calibration and DCA curves demonstrated high predictive accuracy and clinical benefit. Additionally, the nomogram based on the LR model could assist in clinical decision-making. Our findings suggest that the LR model is the optimal predictive model for the risk of hyperkalemia in CKD patients.