Frontiers in Artificial Intelligence · Oct 30, 2026
An Adaptive Knowledge eXploration Optimizer with Explainable Hybrid Feature Selection for Chronic Kidney Disease Prediction
By Sunil Kumar M, RAMYA G
Early detection is key to limit the damage that can be caused by Chronic Kidney Disease (CKD) and to improve the efficiency of treatment. The existing machine learning methods developed for the prediction of CKD employ fixed feature selection methods and optimization is performed without considering the same. To this end, a machine learning approach is developed based on Adaptive Knowledge eXploration Optimizer with Explainable Hybrid Feature Selection framework (AKXO-XHF) by integrating the AKXO algorithm with the XGBoost Classifier and SHapley Additive exPlanation (SHAP). The AKXO-XHF contains adaptive elements, that have the 'memory' of the previous solutions at a low computational cost. The model was tested on a set of 5,575 adults that were reconstructed from raw NHANES (National Health And Nutrition Examination Survey) data from 2021-2023 (972 were labeled as having CKD), and an external set was used for NHANES 2017-March 2020 (8,165 adults), both with repeated nested cross-validation. AKXO-XHF reached an AUC of 0.807 (95\% CI 0.800-0.815) with 12.1 of 22 inputs, compared with 0.812 using all inputs, and an AUC of 0.802 on the external cohort. The Memory operations cost 0.17\% of the runtime, none of its six modules had a measurable individual effect, and it achieved an AUC comparable to six classical optimizers (within 0.0065 AUC over two cohorts) and not significantly different from simple selectors. The SHAP values reflect the dependence of the model, rather than the causal importance, of the variables. A redesigned optimizer AKXO-v2, which learns from all the evaluated subsets thus far, was designed on three non-NHANES datasets, frozen, and then tested with the NHANES folds against GA, PSO, GWO, HHO, SSA, WOA, AKXO and a GA with the same screening step within the same budgets. It achieved the highest mean accuracy, precision, recall, F1-score, AUC (0.810) and MCC but was significantly better than only the screened GA (Holm p = 0.035); on three more datasets the screened GA was first on all three and AKXO-v2 second or third, hence it is considered to be competitive rather than superior.