Christoph Mahler ¹ Felix Friedl ¹ Christian Nusshag ¹ Claudius Speer ¹ Louise Benning ¹ Daniel Göth ¹ Matthias Schaier ¹ Claudia Sommerer ¹ Markus Mieth ² Arianeb Mehrabi ² Christoph Michalski ² Lutz Renders ³ Quirin Bachmann ³ Uwe Heemann ³ Markus Krautter ⁴ Vedat Schwenger ⁴ Fabian Echterdiek ^4,5 Martin Gerd Zeier ¹ Christian Morath ¹ Florian Kälble ^1*

• ¹ Department of Nephrology, University Hospital Heidelberg, Heidelberg, Baden-Württemberg, Germany
• ² Department of General, Visceral and Transplantation Surgery, University Hospital Heidelberg, Heidelberg, Baden-Württemberg, Germany
• ³ TUM Uninversitätsklinikum, Klinikum rechts der Isar, Department of Nephrology, Technical University of Munich, München, Germany
• ⁴ Department of Nephrology, Hospital Stuttgart, Stuttgart, Germany
• ⁵ Department of Nephrology, University Hospital rechts der Isar, Technical University of Munich, Munich, Bavaria, Germany

In the face of growing transplant waitlists and aging donors, sound pre-transplant evaluation of organ offers is paramount. However, many transplant centres lack clear criteria on organ acceptance. Often, previous scores for donor characterisation have not been validated for the Eurotransplant population and are not established to support graft acceptance decisions. Here, we investigated 1353 kidney transplantations at three different German centres to develop and validate novel statistical models for the prediction of early adverse graft outcome (EAO), defined as graft loss or CKD ≥4 within three months. The predictive models use generalized estimating equations (GEE) accounting for potential correlations between paired grafts from the same donor. Discriminative accuracy and calibration were determined via internal and external validation in the development (935 recipients, 309 events) and validation cohort (418 recipients, 162 events) respectively. The expert model is based on predictor ratings by senior transplant nephrologists, while for the data-driven model variables were selected via highdimensional lasso generalized estimating equations (LassoGee). Both models show moderate discrimination for EAO (C-statistic expert model: 0,699, data-driven model 0,698) with good calibration. In summary, we developed novel statistical models that represent current clinical consensus and are tailored to the older deceased donor population. Compared to KDRI, our described models are sparse with only four and three predictors respectively and account for paired grafts from the same donor, while maintaining a discriminative accuracy equal or better than the established KDRI-score.