Explore the scholarly record.

Abstract

Abstract Feature selection is a crucial preprocessing step in medical machine learning, where datasets frequently contain redundant, irrelevant, or highly correlated features. Although metaheuristic algorithms have shown promising results for wrapper-based feature selection, their effectiveness may be limited by premature convergence and search instability. This paper proposes a sequential hybrid framework for medical feature selection based on a population-transfer strategy. The framework integrates three exploration-oriented metaheuristics, namely Particle Swarm Optimization (PSO), Harris Hawks Optimization (HHO), and Brain Storm Optimization (BSO), with the Binary Al-Biruni Earth Radius algorithm (bABER) used as a refinement stage. The optimization process is divided into two phases: an exploration phase performed by the base metaheuristic, followed by an exploitation phase in which bABER refines the transferred population. The proposed framework is evaluated on six benchmark medical datasets using a Random Forest classifier. Experiments are conducted over 15 independent runs following a unified leakage-aware evaluation protocol. Statistical analysis indicates that hybridization preserves the classification accuracy of the base algorithms: a Friedman test finds no significant difference among the seven methods, and equivalence testing confirms that each hybrid matches its base algorithm to within one percentage point. BSO-bABER outperforms BSO on five of six datasets and reaches nominal significance ( $$p = 0.031$$ ), but this does not survive correction for multiple comparisons and the effect size is negligible. The consistent effect is on subset size: the Friedman test on the number of selected features is significant ( $$p = 0.015$$ ), and HHO-bABER selects fewer features than HHO on five of six datasets at unchanged accuracy, at a measured increase in execution time relative to PSO and BSO. Sequential hybridization with population transfer is therefore best understood as a way of obtaining more compact feature subsets at preserved accuracy.