An explainable machine learning model for predicting high phosphorus risk in patients on maintenance hemodialysis: a multicenter retrospective study.

Abstract

Background: Hyperphosphatemia is a common complication in hemodialysis and serves as a key marker for evaluating dialysis adequacy. This study aimed to develop an interpretable machine learning (ML) model to predict the risk of hyperphosphatemia in patients undergoing maintenance hemodialysis.

Methods: Data from 718 patients receiving maintenance hemodialysis across five hemodialysis centers in Chongqing between May and September 2025 were included. The dataset was split into training (70%) and testing (30%) sets. Univariate and multivariate analyses were used to identify predictors most strongly associated with hyperphosphatemia. Seven ML algorithms were employed to construct predictive models. Model performance was assessed using the area under the curve (AUC), F1 score, precision, accuracy, and recall. SHapley Additive exPlanations (SHAP) were applied to visualize feature importance and interpret individual predictions.

Results: The support vector machine (SVM) model demonstrated superior performance (AUC = 0.840). The most influential predictors, in descending order, were PreCr, PreUrea, PreK, PreCO2CP, and PTH. Elevated PreCr, PreUrea, and PreK exert a positive effect on hyperphosphatemia, while estimated PreCO₂CP levels demonstrate a negative effect.

Conclusions: The SVM model exhibited relatively high predictive accuracy for hyperphosphatemia risk in hemodialysis patients. This preliminary multicenter retrospective study demonstrated the feasibility of using SVM combined with the SHAP algorithm for risk identification, and the key features identified (PreCr, PreUrea, PreK, PreCO₂CP, PTH) can serve as a set of candidate parameters for future prospective model development.