DEVELOPMENT OF AN OPTIMAL FEATURE SELECTION SCHEME FOR EPILEPSY CLASSIFICATION FROM DISRUPTIVE ELECTROENCEPHALOGRAM SIGNAL USING AN IMPROVED GRASSHOPPER OPTIMIZATION ALGORITHM

Abstract

Epilepsy is a common type of disorder that causes recurrent seizures and affects approximately 70 million people worldwide. One of the common diagnostic tools is the Electroencephalogram (EEG). EEG is an extremely complex signal that holds information about the various activities of the human brain and neurologists inspect the EEG recordings of an epilepsy patient to identify and analyze epileptic seizures. However, most seizures occur unexpectedly, and finding ways to detect a possible seizure before it happens has been a challenging task for many researchers. This is because the detection of epileptic seizures requires visual monitoring of a patient’s EEG recordings for hours or even days, thus making it a laborious and time-consuming process, and whose outcome may be affected by the experience of the neurologists. As each channel or electrode implanted in the brain provides different statistical measures. A critical issue in epilepsy classification is the selection of suitable statistical features. This necessitated the development of a metaheuristic-based effective and improved grasshopper optimization algorithm (IGOA) using elite opposition-based learning and exponential switching parameters between local and random walks for updating the value of the Grasshopper Optimization Algorithm for the optimizations of feature selection for epilepsy classification from disruptive EEG signals. The original Grasshopper Optimization Algorithm (GOA) was developed using linear switching parameters for updating the iteration value of the Grasshopper Optimization Algorithm, which lead to premature convergence in some complex optimization techniques and drawbacks in exploiting the search space. The IGOA was tested on 14 test functions (unimodal and multimodal benchmark functions) and used to optimize a feedforward artificial neural network for epilepsy classification. From the result, the IGOA outperformed the original GOA in terms of best optimal value, worst, mean and standard deviation and effectively balancing the exploitation and exploration search space. Grasshopper Optimization AlgorithmArtificial Neural Network (GOA-ANN), Particle Swamp Optimization-Artificial Neural Network (PSO-ANN), Salp Swarm Optimization Algorithm-Artificial Neural Network (SSOAANN), Bat Algorithm-Artificial Neural Network (BA-ANN) and Grey Wolf Optimization Algorithm-Artificial Neural Network (GWOA-ANN) were evaluated and compared with IGOA-ANN for their classification accuracies, the number of search agents and features extraction. Also, the result was compared with similar results in the literature. The results showed the classification accuracies performance: IGOA-ANN (99.6%), GOA-ANN (99.40%), GWOA-ANN (98.40%), SSOA-ANN (98.40%), BAANN (98.80%) and PSO- ANN (99.0%) respectively. Based on the previous studies presented in the literature using the University of Bonn EEG dataset, the IGOA-ANN method produced better sensitivity (99.60%), precision (99.60%), and accuracy (99.60%). Considering these metrics and the fact that it requires minimum feature extraction, the IGOA-ANN optimized approach makes it an efficient method for epilepsy classification. The method will help neurologists with efficient and accurate epilepsy classification, thereby saving time.