Machine learning-based prediction of hernia risk in peritoneal dialysis patients: a comparative study of models and SHAP-driven interpretability analysis.

PURPOSE: This study aimed to predict the hernia risk in peritoneal dialysis patients using machine learning (ML) models and conduct an interpretability analysis.

METHODS

A total of 1,144 eligible PD patients (2010-2024) were divided into training (n = 800) and external validation (n = 344) cohorts. Nine ML models were constructed, and SHAP analysis was used for interpretability. Model performance was evaluated via AUC, accuracy, DCA, etc. An online visualization tool based on the optimal model was developed using R Shiny and deployed for clinical use.

RESULTS

The Random Forest (RF) model performed optimally (training AUC = 97.99%, validation AUC = 93.66%), identifying 9 core risk factors (age, BMI, PDV, albumin, smoking history, history of abdominal surgery, high peritoneal transporter status, COPD, and CAPD modality). SHAP clarified non-linear effects of these factors. The developed R Shiny-based online tool (https://caoyugang.shinyapps.io/appforpub/) enables real-time risk calculation through intuitive input of clinical indicators, providing risk stratification and personalized clinical recommendations.

CONCLUSION

The RF model achieves high-accuracy and interpretable hernia risk prediction in PD patients. The R Shiny-based online tool facilitates clinical risk stratification and early intervention, improving patient prognosis.