Development of risk-prediction tool for peritoneal dialysis related peritonitis and external validation in the PDTAP cohort

To the Editor: Peritoneal dialysis (PD) related peritonitis is a common complication of PD, leading to 25–70% of hospitalization, 10–20% of technical failure, and 6–20% of mortality.[1] Identifying high-risk individuals is urgent. While some predictors of treatment failure are recognized, clinicians still struggle with treatment decisions based on these factors in a specific individual. Several risk-prediction models have been developed, which are limited by small sample size, single-center design, and lack of external validation.[2–4] The current peritonitis guideline in 2022 only recommends an empirical antibiotics regime at the onset of peritonitis, without offering individualized treatment based on patients’ risk stratification.[1] Thus, this study aims to develop and externally validate a risk-prediction model for the prognosis of peritonitis. This study consists of two prospective cohorts. The peritonitis cohort of Peking University First Hospital (PKUFH) is used for model derivation, and comprised patients aged ≥14 years who experienced their first-episode of peritonitis after entering the study cohort between 2008 and 2021. Exclusions are for lost follow-up or missing data. The Peritoneal Dialysis Telemedicine Assisted Platform Cohort (PDTAP) is used for model validation and includes patients with similar criteria from 27 hospitals (2016–2019), excluding patients from PKUFH or those lacking relevant data. This study protocol was approved by the Ethics Committee of PKUFH (No. 2018-100). All enrolled patients have signed informed consent forms. The study outcome was a composite of treatment failure including peritonitis-related death and peritonitis-related transfer to hemodialysis within 30 days of peritonitis onset. Baseline clinical characteristics and laboratory tests were collected, with the latter obtained within six months prior to peritonitis onset. PD effluent was examined for white cell counts (WCC) and microorganisms. WCC on days 1 and 3 were recorded and divided by 100. The effect of empirical antibiotics, based on the initial regimen within 3 days, was also analyzed. Continuous variables were expressed as mean ± standard deviation or median (Q1, Q3), with comparisons made via independent sample t-test or Mann–Whitney U test. Categorical variables were analyzed using the chi-squared test or Fisher’s exact test. All variables were analyzed by univariable logistic regression. For significant continuous variables, the cut-off value was explored using a restricted cubic spline (RCS) curve or a set as a commonly recognized critical value, for converting into categorical variables. Significant variables with P <0.1 were included in the multivariate logistic regression analysis with a forward method. Since effluent WCC and pathogenic bacteria were validated as predictive factors, two core models were constructed. The basic model was constructed based on pre-peritonitis basic information, and the extended model added peritonitis information to the basic model. C statistics measured discrimination ability, while model fit was assessed using Akaike information criterion (AIC) and Bayesian information criterion (BIC). Calibration was evaluated using the Hosmer–Lemeshow (HL) statistic and visualized with calibration curves. A point system was created based on the Framingham Heart Research method,[5] assigning scores based on the relative risk of each predictor. The models were applied to the PDTAP cohort for external validation using C statistics and the HL test. Statistical analyses were performed using SPSS version 20.0 (IBM, Armonk, NY, USA) and R version 3.6.1 (R Foundation for Statistical Computing, Vienna, Austria). Two-tailed P <0.05 findings were considered statistically significant. A total of 905 peritonitis episodes were recorded, leading to the enrollment of 528 patients with first-episode peritonitis in the derivation cohort. For the external validation cohort, 7735 patients in the PDTAP cohort were followed, with 1190 patients meeting the inclusion criteria for first-episode peritonitis [Supplementary Figure 1, https://links.lww.com/CM9/C605]. Baseline clinical characteristics were compared between two cohorts [Supplementary Table 1, https://links.lww.com/CM9/C605]. Patients in the derivation cohort were older, and had higher education levels and income. They exhibited a higher prevalence of diabetes mellitus and peritonitis history. In contrast, hemoglobin and serum albumin levels were higher in the derivation cohort, with better indices of lipids spectrum and mineral bone disease despite a decreased residual renal function. Peritonitis information and the outcome of two cohorts were illustrated in Supplementary Table 2, https://links.lww.com/CM9/C605. Patients in the derivation cohort had higher effluent WCC levels on days 1 and 3, with more Gram-negative bacteria and Gram-positive bacteria other than Staphylococcus aureus. Treatment failure occurred in 104 patients from the derivation cohort and 197 from the validation cohort, with failure rates of 19.70% and 16.55%, respectively (P = 0.114). Eight candidate predictors with P <0.1 emerged from univariable logistic regression analysis [Supplementary Table 3, https://links.lww.com/CM9/C605]. Among these, five continuous variables—age, PD duration, serum albumin, serum potassium, and WCC on day 3—were analyzed using RCS curves to determine cut-off values and convert them into categorical variables. Subsequently, four models were constructed (basic models A and B, extended models C and D) to include both baseline factors as continuous or categorical variables, along with the presence or absence of peritonitis information [Supplementary Table 4, https://links.lww.com/CM9/C605]. To facilitate clinical practical risk scoring, we further analyzed the basic Model B and extended Model D presented as categorical variables [Supplementary Table 5, https://links.lww.com/CM9/C605]. The basic model B included PD duration (>25 months) and albumin (<35 g/L), with reasonable discrimination (C statistic 0.635 [95% confidence interval (CI): 0.574–0.696]), overall fit (AIC 509.74 and BIC 522.50), and calibration (HL statistic 3.27 [2df; P = 0.195]). The final predictors in extended model D contained age (>60 years old), PD duration (>25 months), effluent WCC on day 3 (>300/mm3), and causative organisms. Compared to model B, the extended model D performs better with higher C statistic (0.740 [95% CI: 0.680–0.800] vs. 0.635 [95% CI: 0.574–0.696], P = 0.013) and HL statistic (5.40 [8df; P = 0.714] vs. 3.27 [2df; P = 0.195]), and lower AIC (397.22 vs. 509.74) and BIC (438.09 vs. 522.50), indicating better discrimination and overall fit. Receiver Operating Characteristic (ROC) curves for both models are shown in Supplementary Figure 2A, B, https://links.lww.com/CM9/C605 and calibration curves further validate the accuracy of predicted risks [Supplementary Figure 3A, B, https://links.lww.com/CM9/C605]. A points system based on Framingham Heart Research[5] was developed for treatment failure risk [Supplementary Table 5, https://links.lww.com/CM9/C605]. Each predictor in the basic model B and extended model D was assigned a score based on regression coefficients. The total scores ranged from 0 to 5 points for model B and 0 to 31 points for model D. Higher total points correlated with increased treatment failure risk: a score of 5 points indicated approximately 30% treatment failure in model B, while scores >27 points indicated >80% risk in model D. C statistics of basic and extended models reached 0.620 (95% CI: 0.579, 0.661) and 0.875 (95% CI: 0.843, 0.907) in the PDTAP cohort, respectively, showing an improved model discrimination compared to the extended model in the derivation cohort. ROC curves for both models were shown in Supplementary Figure 2C, D, https://links.lww.com/CM9/C605, respectively. The goodness of fit for the basic and extended models was represented as an HL statistic of 0.31 (2df; P = 0.858) and an HL statistic of 7.04 (8df; P = 0.532), respectively, with good calibration also visualized in the calibration curves [Supplementary Figure 3C, D, https://links.lww.com/CM9/C605]. In this study, we developed basic and extended models to predict peritonitis-related death and transfer to hemodialysis. Further external, validation of these models using a national-scale PD cohort showed similar or better discrimination and calibration. We thus calculated two risk-scoring systems based on novel models, where higher predicted risk scores were correlated with a greater likelihood of treatment failure. Our models utilized commonly available clinical data, making them practical for risk stratification of peritonitis across hospitals in various regions. Risk factors, including age, albumin, PD duration, effluent WCC, and causative organism, have been recognized as key predictors for treatment failure of peritonitis in previous literature. The basic model aids early risk stratification upon diagnosis, relying solely on PD duration and albumin. In contrast, the extended model aids in predicting outcomes on day 3 and helps adjust treatment after organism culture and effluent WCC results. According to the International Society for Peritoneal Dialysis (ISPD) guideline, catheter removal for refractory peritonitis should be considered by day 5, highlighting the importance of predicting risk by day 3.[1] Our models differ from the current three models,[2–4] the latter of which include laboratory data and/or effluent tests as necessary variables for risk prediction and clinical decision during peritonitis. Some prognostic predictors in other models,[2–4] such as blood pressure, were excluded from our models, due to the variability in the timepoint measurement in our dataset. Diabetes mellitus, included in other models, was not linked to treatment failure in our data. This inconsistency could be explained by the discrepancy in nutritional status, immune function, and contamination risk due to impaired vision in diabetic patients across studies. While high-density lipoprotein was identified as a protective factor due to its antioxidant and anti-inflammatory activities, it was not strongly predictive in our final risk tools. In addition, our study did not consider novel biomarkers, such as bacterial-derived DNA fragments and soluble C5b-9 complex in the effluent, which could be explored for potential improvement in the performance of models. This study presents the development of prediction models for peritonitis treatment failure, which were externally validated using a national-scale PD cohort. By focusing only on first-episode peritonitis, we minimized the observer bias. A long period of our single-center cohort provided a chance to develop prediction models with good performance. The characteristics of the PDTAP cohort ensured the applicability and generalizability of prediction models. As shown, our validation cohorts covered the broad spectrum of low-risk and high-risk individuals, which demonstrated the feasibility of applying these models to diverse populations. This study addressed the limitations of previous models related to multiple episodes, which included smaller-size or single-center cohorts without external validation. However, limitations exist. The models have been established and validated within the Chinese population, necessitating further validation in different countries. There have been missing data such as effluent WCC, bacterial culture, and prognosis of peritonitis in the validation cohort, all of which may to some extent affect the reliability of the models despite only small proportions of missing data. In this study, traditional effluent culture has produced a relatively high number of negative results, potentially leading to false negatives. Recently, metagenomic next-generation sequencing (mNGS) has gained attention and may become a key tool for monitoring pathogenic bacteria. This project developed and externally validated the basic and extended prediction models for peritonitis prognosis using readily available clinical variables. Conflicts of interest None.