Ch-PSO: A Novel Embedded Method based on PSO and Chebyshev Distance for Enhanced Epileptic Seizure Classification Using EEG Brain Signals

This paper presents a novel hybrid feature selection approach, called Ch-PSO, which combines Chebyshev distance and Particle Swarm Optimization (PSO) to enhance seizure epilepsy classification using electroencephalogram (EEG) signal.Traditional feature selection methods, such as filter methods, often suffer from suboptimal feature ranking due to their reliance on statistical measures without considering classifier performance.On the other hand, wrapper methods provide higher accuracy but are computationally expensive due to repeated evaluations of feature subsets.To address these limitations, the proposed Ch-PSO method operates in two stages.The first stage utilizes Chebyshev distance to identify and retain highly similar features (high-harmony features) while eliminating less relevant ones.The process of dimensionality reduction is employed concerning the initial 4097 features as to different fractions from 10% (410 features) to 60% (2458 features).In the second stage, PSO is performed in the space, filtered by the most essential and significant features.This technique not only enhances the accuracy of the classification but also significantly and effectively reduces the computational burden.A new algorithm, based on PSO, is introduced, which is then tested against the Bonn EEG data to quantify the actual improvements in the performance.With Ch-PSO, a Random Forest (RF) classifier, the classification accuracy has risen to 99%, and it even surpasses the performance of the traditional way of doing things, namely PSO-RF, that could only reach 95%.The higher accuracy achieved with the proposed method in the EEG-based seizure event detection is an indication that it actually works.Besides, Ch-PSO, given its high efficiency, will be the most suitable for the medical field, particularly in the cases of epilepsy diagnosis and general neurological disease detection.The capability to remove feature dimension without the loss of classification accuracy makes it an attractive clinical support tool.