Q-learning enhanced differential evolution for feature selection in high-dimensional medical data analysis

Abstract Feature selection (FS) plays a crucial role in improving predictive performance and computational efficiency in high-dimensional medical data analysis. However, the diversity and complexity of medical datasets pose significant challenges, such as the risk of local optima and high computational costs, especially in large-scale feature spaces. Many feature selection methods fail to reconcile the need for both global and local search, often resulting in non-optimal feature subsets. In this study, we propose a hybrid optimization algorithm named QDEHHO, where Differential Evolution (DE) serves as the backbone search framework, Q-learning adaptively selects mutation strategies and parameter combinations, and Harris Hawks Optimization (HHO) provides directional masks to guide the crossover process. This design enables QDEHHO to dynamically balance exploration and exploitation, achieving robust global search in the early phase and precise local refinement in the later phase. The effectiveness of QDEHHO was validated on multiple benchmark suites, where it was compared against the baseline algorithms and several state-of-the-art competitors. Furthermore, the binary version, bQDEHHO, was applied to 13 publicly available high-dimensional medical datasets with feature dimensions ranging from 2,000 to 15,000. The results show that bQDEHHO has the potential to achieve dimensionality reduction while maintaining or moderately improving classification accuracy compared to traditional methods.