Feature selection for medical diagnosis using enhanced pelican optimization algorithm

Think of metaheuristic algorithms as intelligent and efficient search strategies used to find a nearly optimal solution to complex real-world problems that are too difficult or computationally expensive to find the optimal solution. Metaheuristic algorithms do not guarantee the absolute optimal solution but provide high-quality, practical solutions for the hard optimization problem. Metaheuristics are being employed to solve Feature Selection (FS) problems in medical diagnostics to identify diseases more effectively than with conventional methods. FS is a crucial step in data mining, as it can enhance classification performance by removing redundant or irrelevant features from the input set. In this work a new metaheuristic method is adopted, called Pelican Optimization Algorithm (POA). The basic POA is easy to use when dealing with simple optimization problems; however, for more complex problems, it may suffer from premature convergence by falling into local optima. Poor exploration and exploitation strategies, along with weak population diversity, result in premature convergence. In this paper, an improved POA algorithm called EPOA was proposed to improve performance. The proposed EPOA adopts the best global selection concept from Particle Swarm Optimization (PSO) to enhance the exploitation ability of the basic POA. To further reinforce the exploration ability, the proposed EPOA applies three evolutionary crossover operators, namely uniform, two-point, and single-point crossovers, which are integrated into the EPOA and controlled by a selection probability. Incorporating these operators into the EPOA enhances the exploration and exploitation capabilities of the algorithm. This helps locate the solution space and accelerates convergence while avoiding the local optima. The POA and EPOA methods are then introduced as wrapper-based FS techniques. Five transfer functions to convert their continuous search spaces into binary ones. This process produced five binary variants for both POA and EPOA. We measured their performance on 24 benchmark datasets. The findings indicate that the new EPOA variants outperformed both the POA variants and other existing wrapper-based algorithms in solving high-dimensional FS problems.