Investigation of the stability of features based on functional connectivity of EEG signals originating from motor imagery for application in BCIs

Author(s)

Pedro Felipe Giarusso de Vazquez

• Document type: Master's Dissertation
• Press: Campinas, SP.
• Institution: Universidade Estadual de Campinas (UNICAMP). Instituto de Física Gleb Wataghin
• Defense date: 2024-02-05

Examining board members

Gabriela Castellano; Thiago Bulhões da Silva Costa; André Monteiro Paschoal Paschoal

• Advisor: Gabriela Castellano

Abstract

Brain-computer interfaces based on electroencephalography (EEG) and focused on motor imagery (MI-BCIs) can improve motor rehabilitation processes. Nevertheless, the large variability of intra and inter-subject EEG signals has precluded translation of this technology to the clinical setting. The aim of this work was to evaluate the reproducibility of EEG functional connectivity (FC) features (strength, eigenvector centrality, clustering coefficient, local efficiency) used to discriminate between left- and right-hand MI. Ten subjects underwent 12 EEG-MI-BCI sessions performed in different days, with five runs per session. Two frequency bands (? and ?), three FC methods (imaginary coherence – iCoh –, weighted phase lag index – wPLI – and motif synchronization – MS), four graph parameters (strength – S –, clustering coefficient, eigenvector centrality – EC – and local efficiency) and six electrode sites (C3, C4, F3, F4, Cz, Fz) were evaluated, using a statistical analysis (Wilcoxon test with Bonferroni correction) and a classification analysis (support vector machine with wrappers). The ? band produced the largest number of significantly discriminating features (statistical analysis), and also the best features in the classification analysis and also the most stable features (statistical analysis) for the majority of subjects, suggesting that engagement of needed brain regions may be more important for stability and distinguishability among MI tasks than inhibition of unneeded cortical regions. Electrode sites that stood out in terms of largest number of significantly discriminating features were Cz in the statistical analysis and C4 in the classification analysis. The MS method also produced the largest number of significantly discriminating features, the most stable and most discriminating features for most subjects and also appeared most within the best classification features for the majority of subjects. Since the MS method ignores the amplitude of the changes (it only looks at change patterns), this seems to indicate that signal variation patterns are more important for feature stability and class separability. Finally, EC was the most stable feature for most subjects while both EC and S were the most discriminating, considering both statistical and classification analysis. In summary, using features from the MS method, ? band, S and/or EC parameters, and central electrodes (Cz or C4), results in the best combination for the task of distinguishing left from right-hand MI (AU)

• FAPESP's process: 21/06397-6 - Investigation of the stability of functional brain networks obtained from electroencephalography data for application in brain-computer interfaces
• Grantee: Pedro Felipe Giarusso de Vazquez
• Support Opportunities: Scholarships in Brazil - Master