Donor-Recipient Matching for Kidney Transplantation Using Uncertainty Estimation in Generalized Propensity Score

Kidney donor-recipient matching is a complex process influenced by various clinical and demographic factors, requiring advanced techniques for effective allocation. This study explores the use of Generalized Propensity Score (GPS) modeling with uncertainty estimation and Sufficient Dimensionality Reduction (SDR) methods to enhance matching strategies. We evaluated several dimensionality reduction techniques using Root Mean Squared Error (RMSE) and Pearson Correlation (PC). The regression method (MLP) showed the best performance with the lowest RMSE (1163) and highest PC (0.23) with single dimension. For uncertainty estimation, Quantile Regression outperformed Ensemble and Bayesian MC Dropout methods, with a positive difference of 2695 in mean squared errors between uncertain and certain subsets, indicating its reliability in uncertainty assessment. Our findings establish the benefits of SDR methods for preserving causal relationships and demonstrate that uncertainty-based subset analysis can enhance counterfactual analysis through advanced techniques like propensity score matching and adversarial learning.