A hybrid parameter-tuning for adaptive variable-length particle swarm optimisation in cancer feature selection

High-dimensional gene expression data pose substantial challenges for cancer classification due to extreme dimensionality, feature redundancy, and limited sample availability. Conventional Variable-Length PSO with Local Search (VLPSO-LS) relies on static acceleration coefficients and predefined inertia-weight schedules, which may limit adaptive search behaviour and increase the risk of premature convergence. To address these limitations, this study proposes a novel Variable-Length Adaptive Particle Swarm Optimisation with Local Search (VL-APSO-LS) framework. The proposed method incorporates a dynamic particle representation, adaptive inertia weight and acceleration control, and a comprehensive velocity update mechanism to enhance diversification and exploitation within a variable-length optimisation structure. The framework was evaluated on five benchmark microarray datasets (Leukaemia, SRBCT, Colon, Lymphoma, and Lung cancer) using four widely adopted classifiers. Experimental results demonstrate that VL-APSO-LS consistently improves predictive performance and feature-selection efficiency relative to VLPSO-LS and other PSO-based variants. The strong-balance performance configuration (Scheme 2) achieves an average classification accuracy of 94.39%, outperforming the baseline of 92.64%. Furthermore, it achieves the highest Accuracy-to-Feature Ratio (0.3775 vs 0.3369), indicating a superior balance between accuracy and feature compactness. Notably, the proposed method achieves perfect classification accuracy (100%) on the Leukaemia dataset and reduces the selected features in the Colon dataset to only 86, compared with over 700 features in competing approaches. Statistical validation confirms consistent non-inferiority and frequent superiority across dataset-classifier combinations, particularly when integrated with Support Vector Machine (SVM). Overall, VL-APSO-LS provides a robust, adaptive, and statistically validated optimisation framework for high-dimensional biomedical classification.