Enhanced swarm optimization for feature selection in electroencephalogram classification: investigating visibility graph and persistent homology-based features

Abstract

The analysis of high-dimensional, nonlinear electroencephalogram (EEG) remains challenging, particularly for non-medical EEG, which shows only subtle distinctions between data classes, compared to medical EEG. This study proposed a novel persistent homology (PH) pipeline by incorporating visibility graphs and an enhanced binary particle swarm optimization (BPSO) with four improvement strategies into a range of PH representations and filtrations, to classify non-medical EEG recordings in a visual recognition task under varying auditory conditions. By integrating multi-domain features and robust feature selection, the proposed pipeline fills a crucial gap left by earlier PH-based EEG studies that mainly focus on narrow, single-domain feature sets. The highest increases of 23.71% in accuracy and 17.77% in F1-score were achieved when classifying the alpha EEG from the O2 channel using k-nearest neighbors classifier. The comparative analysis demonstrated the superiority of the enhanced BPSO over standard BPSO, while persistence landscape, silhouette, Vietoris-Rips filtration, and weighted visibility graph consistently surpassed the others in performance. Alpha EEG exhibited better classification performance than beta EEG, indicating a stronger link between alpha activity and attentional modulation. The statistical significance test, hyperparameter sensitivity analysis, and benchmarking results using a public epilepsy EEG dataset validated the applicability of the proposed pipeline in different EEG analysis tasks. These findings corroborated the capability and impact of the proposed pipeline in complex EEG analysis, promoting the development of the brain-computer interfaces.