Machine learning-based prediction of early-onset peritoneal dialysis-associated peritonitis: the role of the CONUT score

Le résumé fourni par la source

Background Peritoneal dialysis-associated peritonitis (PDAP) remains a major complication of peritoneal dialysis (PD). The controlling nutritional status (CONUT) score, which reflects the immune-nutritional state, may offer predictive value in identifying patients at risk. This study aimed to evaluate the utility of machine learning models in predicting early-onset PDAP and to assess the prognostic importance of baseline CONUT score, 6-month CONUT score, and their dynamic changes. Methods In this multicenter prospective cohort study, 675 patients initiating PD were enrolled. Multivariable logistic regression was performed to identify clinical predictors of early-onset peritonitis, while Kaplan–Meier survival analysis was used to compare peritonitis-free survival among patients with no peritonitis, early-onset peritonitis, and late-onset peritonitis. To enhance predictive performance, machine learning models including XGBoost, LightGBM, and their ensemble were constructed. Feature selection was based on SHapley Additive exPlanations (SHAP) values derived from an initial XGBoost model. The top 10 SHAP-ranked features were used to train all models. Model performance was assessed using area under the receiver operating characteristic curve (AUC), and SHAP summary plots were generated to interpret feature contributions. Results Over a median follow-up period of 41.8 months, 82 patients developed early-onset PDAP. Multivariable logistic regression identified baseline total cholesterol, neutrophil-to-lymphocyte ratio, and 6-month CONUT score as independent predictors of early-onset PDAP (vs. no PDAP; p < 0.05). In comparisons between early- and late-onset PDAP, older age, longer PD duration, and lower 6-month CONUT score were independently associated with a decreased likelihood of early-onset PDAP ( p < 0.05). Using the top 10 SHAP-ranked features, three models (XGBoost, LightGBM, and an ensemble) were trained. For distinguishing early-onset PDAP from no PDAP, LightGBM performed best (AUC = 0.717), followed by the ensemble (0.698) and XGBoost (0.670). In differentiating early- from late-onset PDAP, LightGBM showed the highest AUC (0.781), outperforming the ensemble (0.744) and XGBoost (0.691). SHAP summary plots consistently identified the 6-month CONUT score as the important feature across both classification tasks. Conclusion The 6-month CONUT score is an independent predictor of early-onset PDAP and was among the top contributing features in multiple machine learning models. Integrating SHAP-based feature selection with gradient boosting improved model accuracy and interpretability. Dynamic monitoring of nutritional-immune status may aid in early risk stratification and guide personalized prevention strategies in patients undergoing PD.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.