Objective To predict the risk factors of malnutrition in peritoneal dialysis (PD) patients using machine learning (ML) algorithms, and provide the reference for nutritional management decisions.

Methods A retrospective analysis was conducted on 171 PD patients, the nutritional status was assessed using the Subjective Global Assessment (SGA), the patients were divided into the malnutrition group (n = 69) and well-nourished group (n = 102), and the data were preprocessed.The feature variables were screened by collinearity diagnosis, least absolute contraction and selection operator (LASSO). Five machine learning algorithms, namely Random Forest (RF), Extreme Gradient Boosting (XGB), Support Vector Machine (SVM), K-Nearest Neighbor Model (KNN) and Lightweight Gradient Boosting (Light GBM), were selected for predictive modeling. After tenfold cross-validation, the model was comprehensively evaluated using the receiver operating characteristic curve, area under the receiver operating characteristic curve (AUC), precise recall rate (PR) curve, accuracy rate, sensitivity, specificity and F1 index. The Shapley additive interpretation (SHAP) was introduced to process the interpretability of the optimal machine learning model.

Results After LASSO regression analysis, nine characteristic variables were determined for constructing the machine learning model. The comprehensive evaluation showed that the RF model had high AUC (0.994), accuracy rate (0.960), sensitivity (0.905), specificity (0.967), recall rate (0.952) and F1 index (0.952). The results of interpretive analysis of the SHAP model showed that the top 5 characteristics in terms of contribution were hypersensitive C-reactive protein, hemoglobin, albumin, history of hypertension and white blood cells in sequence.

Conclusions In the prediction model of malnutrition in peritoneal dialysis patients, the RF model shows the best performance and can provide extremely valuable basis for formulating nutritional management strategies for peritoneal dialysis patients.