Bioinspired hybrid optimisation and deep belief neural networks for early chronic kidney disease detection: an explainable clinical AI framework.

The early detection of chronic kidney disease (CKD) can lead to timely and appropriate clinical intervention. However, most CKD diagnostic systems rely on redundant features and provide unreliable results in small and unbalanced datasets. Therefore, it is challenging to apply them in clinical settings due to a lack of transparency, as well as the extensive amount of time and effort to fine-tune them using manual labor. This article outlines an algorithm to test CKD automatically with a hybrid spiral search strategy-based binary gravitational search algorithm (SSS-BGSA) with elephant herding optimization (EHO) to optimize a Deep Belief Neural Network (DBNN). The code pipeline was designed to involve automated feature selection and model parameter optimization. The objective of using SSS-BGSA in the study was to enhance the exploration-exploitation tradeoff of the original BGSA by a non-linear spiral search pattern. Training of the DBNN using the selected features and optimization of the DBNN parameters were also done using EHO to enhance convergence rate and learning efficiency. The framework was evaluated on the UCI CKD dataset (25 attributes) using nested stratified 5-fold cross-validation. The suggested SSS-BGSA-EHO-DBNN demonstrated a competitive performance (Accuracy = 0.973 ± 0.022, AUC =0.996 ± 0.006), as well as identifying a minimum of 7 clinically important features. Specific gravity, hypertension, packed cell volume, glucosuria, and blood urea were identified as the most important features. The proposed SSS-BGSA-EHO-DBNN framework demonstrates that dual-stage optimization combined with explainable deep learning can yield a fully automated, interpretable, and computationally efficient CKD screening pipeline.