International validation of a pre-transplant risk assessment tool for graft survival in pediatric kidney transplant recipients

TL;DR

This study validates a kidney transplant risk model for children in different countries and finds that local adaptations improve predictions.

Contribution

The study demonstrates the need for country-specific adaptations of a pediatric kidney transplant risk model to improve international applicability.

Findings

The Dutch prediction model showed good internal validation (AUC 0.77 at 10 years) but poor external validation in Germany and France.

Adapted national models outperformed the international model with higher AUCs in Dutch, French, and German cohorts.

Cohorts differed significantly in baseline characteristics and outcomes, highlighting the need for localized models.

Abstract

A pre-transplant prediction model using commonly available factors is valuable for optimizing donor selection, communication, and counseling for pediatric kidney transplant (PKT) recipients. This study aims to externally validate a Dutch PKT prediction model and assess its international applicability.

Data from the Dutch-, CERTAIN-, and CRISTAL registries, covering PKT from 2005 to 2021, were used. The Dutch prediction model was externally validated in a German and French cohort and then adapted to these specific countries. An international prediction model was also developed using all available data. Models were based on 80% derivation cohorts and internally validated using areas under the receiver operating characteristic curve (ROC-AUC) and calibration plots.

Of 3266 transplantations, 2475 (273 Dutch, 356 German, 1622 French, and 224 other) were used for analysis. Cohorts differed significantly in baseline characteristics and outcomes. Internal validation of the Dutch model showed ROC-AUC of 0.77 and 0.75 at 10 and 15 years. External validation in German and French cohorts yielded 10-year ROC-AUC of 0.63 and 0.60, respectively.

Internal validation of the international prediction model showed AUC of 0.61 and 0.60 at 10 and 15 years with poor calibration, indicating inferior performance. The adapted national models showed better internal validation performance, with 10-year ROC-AUC of 0.77, 0.76, and 0.73 in Dutch, French, and German cohorts, respectively.

The Dutch PKT prediction tool requires country-specific adaptations for use in other countries, given the diversity of clinical practice across Europe. A country-specific model is preferable to an international model in the current landscape.

Graphical Abstract

Figures