CUA-Moe: A Clinical Subspace and Uncertainty Aware Patient-Adaptive Mixture-of-Experts Framework for Chronic Kidney Disease Classification

Abstract

Chronic kidney disease (CKD) classification from heterogeneous clinical measurements presents a modelling challenge because laboratory, urinary, demographic and comorbidity variables may provide complementary and partially overlapping information. This study proposes CUA-MoE (Clinical Subspace and Uncertainty-Aware Mixture-of-Experts), a patient-adaptive mixture-of-experts framework that assigns observation-specific contributions to specialised predictive experts rather than applying fixed global model weights. The framework comprises laboratory, clinical-risk, urinary/inflammatory and global experts, with each expert trained on a predefined feature subspace. A patient-specific routing layer integrates out-of-fold expert probabilities with entropy-based uncertainty, inter-expert disagreement, missingness patterns and reliability-diversity priors to determine the relative contribution of each expert. To reduce information leakage, pre-processing, expert training and routing components were fitted within the cross-validation workflow, while calibration and threshold selection were separated from the outer test folds. Experiments were performed on the supplied 400 observation CKD dataset using stratified five-fold outer cross-validation with three-fold inner out-of-fold estimation. The dataset comprised 24 predictive variables, with 250 CKD and 150 non-CKD observations in the supplied cleaned representation. CUA-MoE achieved an accuracy of 97.75%, balanced accuracy of 97.42%, precision of 97.61%, recall of 98.79%, specificity of 96.05%, F1-score of 98.20%, ROC-AUC of 0.9899, PR AUC of 0.9872, and Matthews’s correlation coefficient of 0.9522. Aggregated across the outer test folds, the resulting confusion matrix contained 246 true positives, 145 true negatives, 5 false positives and 4 false negatives, corresponding to 9 misclassified observations out of 400 at the 0.50 reporting threshold. These findings demonstrate strong predictive performance for CUA-MoE on the evaluated benchmark and support the value of patient-specific expert routing for integrating heterogeneous feature subspaces. Further evaluation on independent external datasets is required to determine the generalisability and robustness of the framework.