Development and validation of a machine learning-based prediction model for malnutrition risk in peritoneal dialysis...

To develop and validate a machine learning (ML)‑based model for predicting malnutrition risk in patients undergoing maintenance peritoneal dialysis (PD) using routinely available clinical and laboratory data.

A multicenter retrospective cohort study was conducted. Adults with end-stage renal disease who received PD for at least 90 days were enrolled from two tertiary hospitals. Malnutrition was defined by a 35-point Modified Quantitative Subjective Global Assessment (mSGA) score > 7 at 3-month post-enrollment. Candidate predictors included demographics, laboratory indices, dialysis parameters, comorbidities, and psychosocial scores. After feature selection via LASSO regression, three models-logistic regression (LR), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost)-were trained and optimized. Model performance was evaluated on an internal and an independent external validation set using discrimination (AUC), calibration, and decision curve analysis (DCA). Model interpretability was enhanced with SHAP analysis and a clinical nomogram.

Among 497 included patients, 29.8% were malnourished. Nine predictors were selected by LASSO: age, dialysis vintage, serum albumin, high‑sensitivity C‑reactive protein (hs‑CRP), hemoglobin, total weekly Kt/V, depression score (SDS), anxiety score (SAS), and diabetic nephropathy as the primary cause. XGBoost achieved the highest discrimination, with an AUC of 0.868 (95% CI 0.792-0.944) on internal validation and 0.857 (95% CI 0.805-0.909) on external validation, significantly outperforming LR (p = 0.023). Calibration was excellent (confirmed by calibration intercept, slope, and Brier score), and DCA demonstrated superior net benefit across clinically relevant risk thresholds. Serum albumin, hs‑CRP, and total weekly Kt/V were consistently the most influential predictors across all models.

An XGBoost‑based ML model effectively predicts malnutrition risk in PD patients using readily available clinical data. With robust validation, interpretable outputs (SHAP, nomogram), and confirmed clinical utility, this tool holds promise for early identification and targeted intervention in high‑risk patients.