A hybrid of an automated multi-filter with a spatial bound particle swarm optimization for gene selection and cancer classification

<h2>Abstract</h2> Cancer is one of the most dangerous diseases and a leading cause of death globally. Therefore, early detection of cancer is critical for effective treatments. However, the main challenge in disease identification and classification, such as cancer microarray dataset, is the large number of genes. Extracting meaningful information from these genes is a challenging task. This study proposes a new hybrid gene selection approach for cancer classification that combines three multi-filters in an automated manner with the proposed spatial bound particle swarm optimization (AMF-SBPSO) to effectively select the most important genes. The proposed AMF-SBPSO selects genes through two phases. In the first phase, the method uses three multi-filters to remove the most irrelevant and noisy genes in an automated manner. This process combines multiple filters based on various metrics and filter size ratios, then votes for the best one. This phase aims to select the best filter type and optimal ratio size for each dataset, taking into account that each dataset has unique gene sizes and characteristics. What may work for one dataset might not be effective for others. This phase makes the filtering process more robust. In the second phase, a new variant of particle swarm optimization (PSO), called spatial bound PSO (SBPSO), based on a spatial bounding strategy is proposed to reduce the remaining number of genes without losing the classifier's performance. The proposed AMF-SBPSO's performance was evaluated through experimental studies on nine microarray gene datasets. The results demonstrate that AMF-SBPSO yields a 100% classification accuracy on all tested datasets, outperforming state-of-the-art methods. These findings indicate the superiority of the proposed method in reducing the number of genes and achieving the highest accuracy for cancer classification.