Prediction Models for Dialysis Peritonitis Fall Short of Clinical Use, Review Finds

Peritonitis is one of the most feared complications of peritoneal dialysis, a home-based form of kidney replacement therapy. When infection strikes the peritoneal cavity, patients face hospitalization, loss of their dialysis catheter, a forced switch to hemodialysis, and in some cases death. Clinicians have long wanted tools that could tell them, soon after a diagnosis is made, which patients are likely to fail treatment so that catheter removal or other interventions could be timed wisely. A new systematic review now suggests that the mathematical models built for this purpose are far from ready for the bedside.

The review, published in Frontiers in Medicine, systematically searched eight databases, including PubMed, EMBASE, Web of Science, Scopus, the Cochrane Library, and three Chinese databases, from inception to January 30, 2026, with an updated search on March 15, 2026. The team, led by Qishun Wu of the Second Affiliated Hospital of Wannan Medical University in Wuhu, China, included studies that developed or externally validated multivariable models predicting adverse outcomes after a peritonitis diagnosis in adults on maintenance peritoneal dialysis. Two reviewers independently screened records, extracted data, and applied PROBAST, a tool designed specifically to judge whether prediction model studies are methodologically robust enough for clinical use, across four domains: participants, predictors, outcome, and analysis.

The search yielded seven studies reporting seven prediction models. Six of these studies together comprised 3,821 peritonitis episodes; a seventh study did not report its sample size. All seven models were derived in Asia, with five from China, one from Thailand, and one from Taiwan, and none from Europe or Australia. Five models used logistic regression, one used a random forest approach, and one combined logistic regression with a decision tree. Five models predicted risk at baseline, while two reassessed risk early, between 24 and 72 hours after diagnosis. Six studies were retrospective cohorts and one was cross-sectional.

The central finding is stark: no model achieved an overall low risk of bias. Two models, one by Liu and colleagues and one by Qiao and colleagues, were rated low risk in the participants, predictors, and outcome domains, but both received an unclear rating in the analysis domain because of unclear handling of missing data, no explicit assessment of overfitting, and incomplete calibration reporting. That pushed both to an overall unclear rating. The remaining five models were rated high risk overall, driven primarily by reliance on internal or apparent validation alone, incomplete missing-data handling, and small or unreported sample sizes that raise stability concerns.

External validation emerged as a critical weakness. The reviewers classified validation approaches using a four-tier framework, and only one model, from Qiao and colleagues, achieved even limited independent external validation, at Tier 3, meaning geographic separation within the same national registry, the PDTAP cohort. No model achieved Tier 1 or Tier 2 validation, which would require data from a genuinely separate country or a different center without recruitment overlap. The remaining six models relied on internal or apparent validation only. The authors argue this distinction matters because microbiological profiles, antibiotic policies, and catheter removal thresholds differ substantially across countries, creating effect modification that separation within a single network may not capture.

Calibration, the question of whether predicted probabilities actually match observed frequencies, was another pervasive gap. Three studies reported some form of calibration assessment: two published a calibration plot without quantitative metrics and one reported only a Hosmer–Lemeshow test. Not a single study provided a calibration slope and intercept, which the review set as its stricter completeness standard. Without those metrics, the authors note, quantitative recalibration of any included model in a new setting is impossible, because even the basic updating step of adjusting the intercept and slope cannot be performed. A model that ranks patients correctly but systematically overestimates or underestimates risk could drive unnecessary catheter removal, permanently ending a patient’s peritoneal dialysis, or delay essential intervention.

Reported discrimination varied widely. Liu and colleagues reported a C-statistic of 0.80 (95% confidence interval 0.74 to 0.86) from single-center internal bootstrap validation, while Qiao and colleagues reported 0.74 (95% CI 0.68 to 0.80) from the more rigorous geographic limited validation. The models with internal or apparent validation only reported C-statistics ranging from 0.756 to 0.916, though several reported point estimates without confidence intervals. The reviewers caution that internally validated performance typically appears higher than what independent external validation would show, and that even the externally validated figure of 0.74 corresponds to a discrimination level at which misclassification rates remain substantial.

Applicability concerns were identified across all seven models, with high concern in the analysis domain for every one of them, reflecting incomplete calibration reporting, lack of model updating guidance, and absence of implementation tools. No study reported decision curve analysis or net benefit, no model provided a public calculator, code repository, or implementable risk score, and only the Qiao model offered a nomogram, without validation in an independent cohort. Model formulas were incompletely reported in four studies, precluding independent replication. None of the models incorporated patient-reported outcomes or the core outcome domains proposed by the SONG-PD initiative, such as life participation, fatigue, and sleep, and no study engaged patients in defining meaningful prediction thresholds.

The review also acknowledges its own limitations. Twelve potentially relevant non-English records, mostly in Chinese, could not be fully assessed because full texts were unobtainable, and were excluded based on English titles and abstracts alone, a practice the authors classify as potential selection bias of uncertain direction. A quantitative sensitivity simulation suggested that even if one or two high-quality models existed among those excluded records, the conclusion about implementation readiness would remain unchanged. The reviewers also note that they did not search for implementation research specifically, so their conclusions about clinical impact reflect only the absence of such studies within the included model development literature.

The bottom line, according to the authors, is that no currently identified model has sufficient evidence for routine clinical implementation or screening. Under strict supervision, local calibration, and a clear research purpose, the Liu and Qiao models may serve as research tools, most defensibly to conduct prospective calibration studies that report slope and intercept before any clinical use is considered. The reviewers propose that future research follow the IDEAL framework, noting that all currently identified models appear to be at the development or validation stage, with none reaching prospective validation or clinical impact assessment. They call for international collaborative networks for multilingual model validation, transparent reporting of calibration metrics with confidence intervals, decision curve analysis, and greater involvement of patients in defining what prediction models should actually predict.

Subject of Research: Risk of bias and applicability of prediction models for outcomes of peritoneal dialysis-related peritonitis

Article Title: Risk of bias and applicability of prediction models for outcomes of peritoneal dialysis-related peritonitis: a PROBAST systematic review

Article References: Wu, Q., Yu, Z., Wu, Z., Yang, L., & Lin, X. (2026). Risk of bias and applicability of prediction models for outcomes of peritoneal dialysis-related peritonitis: a PROBAST systematic review. Frontiers in Medicine, 13, Article 1882053. https://doi.org/10.3389/fmed.2026.1882053

Image Credits: AI Generated

DOI: 10.3389/fmed.2026.1882053

Keywords: peritoneal dialysis, peritonitis, prediction models, PROBAST, risk of bias, external validation, calibration, systematic review, nephrology, clinical decision support, treatment failure, machine learning