First experiences with machine learning predictions of accelerated declining eGFR slope of living kidney donors 3 years after donation

• DOI: 10.1007/s40620-024-01967-y
• Publication Date: 2024-06-05T14:01:46Z
• Authors: Leandra Lukomski; Juan Pisula; Tristan Wagner; Andrii Sabov; Nils Große Hokamp; Katarzyna Bozek; Felix Popp; Martin Kann; Christine Kurschat; Jan Ulrich Becker; Christiane Bruns; Michael Thomas; Dirk Stippel
• Keywords: Kidney donation; End-stage kidney disease

Abstract Background Living kidney donors are screened pre-donation to estimate the risk of end-stage disease (ESKD). We evaluate Machine Learning (ML) predict progression function deterioration over time using estimated GFR (eGFR) slope as target variable. Methods included 238 living who underwent donor nephrectomy. divided dataset based on eGFR in third follow-up year, resulting 185 with an average and 53 accelerated declining eGFR-slope . trained three Learning-models (Random Forest [RF], Extreme Gradient Boosting [XG], Support Vector [SVM]) Logistic Regression (LR) for predictions. Predefined data subsets served training explore whether parameters ESKD score alone suffice or additional clinical time-zero biopsy enhance learning-driven feature selection identified best predictive parameters. Results None four models classified AUC greater than 0.6 F 1 surpassing 0.41 despite different subsets. Following machine subsequent retraining these selected features, random forest extreme gradient boosting outperformed other models, achieving 0.66 0.44. After selection, two attributes consistently appeared all models: smoking-related features glomerulitis Banff Lesion Score. Conclusions Training learning-models distinct predefined yielded unsatisfactory results. However, efficacy improved when exclusively suggesting that quality, rather quantity, is crucial learning-model performance. This study offers insights into application emerging learning-techniques screening donors. Graphical abstract