Development of Risk-Prediction Tool for Peritoneal Dialysis-Associated Peritonitis and External Validation in the PDTAP Cohort

Background: Although peritoneal dialysis related peritonitis is a common complication among peritoneal dialysis (PD) patients, there is no validated and recognized tool to predict disease prognosis. This limits patient-specific risk stratification and treatment decisions. Methods: A single-center peritonitis cohort was used for model derivation and Peritoneal Dialysis Telemedicine-assisted Platform (PDTAP) cohort was used for external validation. Logistic regression models were used to analyze the risk of treatment failure within 1 month of peritonitis, i.e. a composite of peritonitis-related mortality and transferring to hemodialysis. The discrimination and calibration were evaluated using C statistics, Hosmer-Lemeshow (HL) test, Akaike information criterion (AIC), Bayesian information criterion (BIC) and calibration curves. Predictors were weighted to calculate a risk score. Results: Totally, 528 and 1190 first-episode peritonitis were included in the derivation cohort and validation cohort, respectively. A total point of 5 in basic model was developed including baseline albumin <35g/L and PD duration >25months, and a total point of 29 in extended model was developed including age >60 years old, PD duration >25months, the white cell count >300/mm3 in dialysis effluent on the day 3, and causative organism. Compared with basic model, the extended model performs better with higher C statistic [0.744 (95%CI 0.684-0.804) vs. 0.635 (95%CI 0.574-0.696), P = 0.012], HL statistic [6.73 (8df; P = 0.566) vs. 3.27 (2df; P = 0.195) and lower AIC (389.10 vs. 509.74) and BIC (421.99 vs. 522.50). Both models in the validation cohort performed similar or better discrimination and calibration, as shown in C statistics [0.620 (95% CI 0.579-0.661) in basic model; 0.871 (95%CI 0.837, 0.905) in extended model] and HL statistic [0.31 (2df; P = 0.858) in basic model; 7.86 (8df; P = 0.447) in extended model]. Conclusion: In this study, the basic and extended models for predicting treatment failure of peritonitis were established based on commonly used clinical data, which were applicable to different clinical scenarios and facilitated clinical doctors to identify high-risk individuals and adjust treatment decisions.