A Metaheuristic-Driven Diagnostic Optimization Framework for Early Disease Detection Using the Marine Predators Algorithm

Abstract

Early and reliable detection of breast cancer hinges on identifying a compact subset of diagnostically informative biomarkers from high-dimensional clinical data, yet most wrapper-based feature selection studies evaluate only a handful of optimizers and rarely benchmark recent nature-inspired algorithms under a unified statistical protocol. This paper proposes a diagnostic optimization framework in which the Marine Predators Algorithm (MPA), governed by Lévy flight and Brownian motion dynamics, drives binary feature selection over a support vector machine classifier applied to the Wisconsin Diagnostic Breast Cancer (WDBC) dataset. A multi-objective fitness function balancing classification accuracy and subset cardinality, weighted by an accuracy-oriented coefficient of 0.9, directs the search toward clinically actionable feature subsets. To position MPA within the contemporary metaheuristic landscape, ten competing algorithms—Genetic Algorithm, Particle Swarm Optimization, Differential Evolution, Artificial Bee Colony, Grey Wolf Optimizer, Whale Optimization Algorithm, Harris Hawks Optimizer, Aquila Optimizer, Slime Mould Algorithm, and Salp Swarm Algorithm—were executed under identical conditions across thirty independent runs. MPA attained the highest mean accuracy of 98.49 percent with only 11.13 features on average, a 62.9 percent reduction relative to the full thirty-feature set, while improving sensitivity to 98.56 percent. Wilcoxon signed-rank tests confirmed that the gains over every competitor were statistically significant (p < 0.05), with the strongest margins observed against classical methods. The findings indicate that MPA's Lévy-driven exploration delivers compact, biologically coherent feature subsets that retain diagnostic power, supporting its adoption as a principled wrapper optimizer for medical decision support.