Advanced Metaheuristics for Feature Selection and Classification in Medical Datasets

Feature selection is a critical step in medical classification, where large and complex datasets can obscure clinically relevant patterns. In this work, we employ a wrapper-based feature selection framework built around a k-Nearest Neighbors classifier and compare ten state-of-the-art metaheuristic approaches for identifying minimal yet discriminative feature subsets. These approaches span swarm-intelligence (e.g., Grey Wolf Optimizer, Marine Predators Algorithm) and evolutionary algorithms, including Self-adaptive Differential Evolution and JADE (Just Another Differential Evolution). Experiments on seven medical datasets-spanning cancers (Arcene, Breast, Lung) and cardiac arrhythmia-show that certain methods, notably Moth-Flame Optimizer, Self-adaptive Differential Evolution, and JADE, achieve robust convergence and consistently high classification accuracy, exceeding 98% on multiple benchmarks and reaching 100% on Lung Cancer. Meanwhile, algorithms like Sine Cosine and Marine Predators can surpass 97% accuracy on specific problems but vary depending on dataset characteristics. The results confirm that no single technique excels universally across all clinical datasets; rather, the optimal choice depends on factors such as dataset dimensionality and class distribution. Overall, this study underscores the value of pairing metaheuristics with effective feature selection strategies in enhancing diagnostic accuracy. Future research will integrate multi-objective optimization and domain knowledge to further elevate clinical interpretability and reliability.