Explainable Adaptive Metaheuristic Optimization for Feature Selection in Breast Cancer Diagnosis: A Review of Transparent and Effective AI Approaches

Abstract

Breast cancer remains the most frequently diagnosed malignancy among women worldwide, and computer-aided diagnostic systems built on machine learning have become central to early detection efforts. A persistent obstacle to clinical adoption is the tension between predictive accuracy and interpretability: high-dimensional diagnostic datasets, drawn from mammography, ultrasound, histopathology and genomic assays, contain many redundant or irrelevant variables, while the models capable of exploiting them are frequently opaque. Metaheuristic optimisation algorithms, including grey wolf, whale, particle swarm, genetic and hybrid variants, have been widely applied to reduce dimensionality by selecting compact, diagnostically relevant feature subsets, and these algorithms are increasingly combined with explainable artificial intelligence methods such as SHapley Additive exPlanations and Local Interpretable Model-agnostic Explanations to render the resulting models auditable by clinicians. This narrative review synthesises the recent literature on explainable, adaptive metaheuristic feature selection for breast cancer diagnosis. It examines the conceptual foundations of explainable artificial intelligence, surveys the principal metaheuristic algorithms used for feature selection, and critically appraises studies that integrate optimisation with post hoc or intrinsic explanation methods across imaging, clinical and multi-omics data modalities. The review finds that hybrid and adaptive metaheuristic strategies consistently outperform single-algorithm baselines in reducing feature dimensionality while preserving or improving classification accuracy, and that explanation layers built atop these optimised models can identify clinically plausible biomarkers, though evaluation of explanation fidelity remains inconsistent across studies. Persisting gaps include limited external validation, scarce prospective clinical testing, uneven reporting of computational cost, and the absence of standardised metrics for explanation quality. Future work should prioritise adaptive parameter control, multi-objective formulations that explicitly balance accuracy against interpretability, and clinician-centred evaluation of explanations. The review concludes that explainable adaptive metaheuristic feature selection is a promising but still maturing pathway toward trustworthy, deployable artificial intelligence for breast cancer diagnosis.