Group task-related component analysis (gTRCA): a multivariate method for inter-trial reproducibility and inter-subject similarity maximization for EEG data analysis

Abstract

Group task-related component analysis (gtRcA): a multivariate method for inter-trial reproducibility and inter-subject similarity maximization for eeG data analysis Hirokazu tanaka eeG is known to contain considerable inter-trial and inter-subject variability, which poses a challenge in any group-level EEG analyses. A true experimental effect must be reproducible even with variabilities in trials, sessions, and subjects. extracting components that are reproducible across trials and subjects benefits both understanding common mechanisms in neural processing of cognitive functions and building robust brain-computer interfaces. this study extends our previous method (task-related component analysis, tRcA) by maximizing not only trial-by-trial reproducibility within single subjects but also similarity across a group of subjects, hence referred to as group tRcA (gtRcA). the problem of maximizing reproducibility of time series across trials and subjects is formulated as a generalized eigenvalue problem. We applied gTRCA to EEG data recorded from 35 subjects during a steady-state visual-evoked potential (SSVEP) experiment. The results revealed: (1) The group-representative data computed by gtRcA showed higher and consistent spectral peaks than other conventional methods; (2) Scalp maps obtained by gTRCA showed estimated source locations consistently within the occipital lobe; And (3) the high-dimensional features extracted by gTRCA are consistently mapped to a lowdimensional space. We conclude that gTRCA offers a framework for group-level EEG data analysis and brain-computer interfaces alternative in complement to grand averaging. A major issue in subject-level and group-level EEG analysis is intra-subject and inter-subject variability across trials and sessions that originates from both endogenous factors and exogenous factors 1. Endogenous factors of intra-subject variability include includes artifacts such as blinking, psychophysiological fluctuations, and fatigue, and exogenous factors include a temporal change in electrode impedance and positions. Inter-subject variability includes anatomical differences across subjects such as head shapes, skull conductivity, and patterns of brain gyrification (i.e., folding of the cerebral cortex), in which genetic differences play a primary role. These factors of variability consist of effects of non-interest and conceal an effect of interest related to a task. To sum, EEG is dynamic and non-stationary across sessions and subjects 2 , and the variability within and across subjects obfuscates both subject-and group-level analysis. Trial-mean across subjects (grand mean) is a simple and robust solution to improve signal-to-noise ratio by averaging out effects of non-interest within the group. In addition to univariate averaging, several multivariate methods have been proposed to enhance an experimental effect, including PCA-based ERP analysis 3-11 , partial-least-squares analysis 12 , independent component analysis 13,14 (see more detail in Discussion). Here, we propose a multivar...

Application of Transfer Learning in EEG Decoding Based on Brain-Computer Interfaces: A Review

“…The results showed that this method significantly improves the information transfer rates and classification accuracy. Based on this research, Tanaka [41] improved the TRCA method by maximizing the similarity across group of subjects, and they named this novel method group TRCA. The results showed that the group representation calculated by the group TRCA method achieve high consistency between two domains and offer effective data supplementation during brain-computer interaction.…”

tACS phase-specifically biases brightness perception of flickering light

“…Critically, to get an optimal estimate of the tACS-targeted 8 Hz rhythm evoked by visual flicker, we applied spatio-spectral decomposition (SSD) to data from all seven parieto-occipital electrodes contralateral to the flicker [31]. SSD is a dimensionality reduction method that has been repeatedly applied in the framework of flicker-evoked response analysis to optimally extract rhythmic components in the data [32,33]. This procedure resulted in spatially weighted SSRSSD time series with larger signal-to-noise-ratio compared to single-channel analysis (Fig 3D).…”

“…SSD maximizes signal power at the frequency of interest while simultaneously minimizing power at surrounding frequencies [31]. This procedure has been repeatedly applied in the framework of flicker-evoked response analysis [32,33]. Based on PLV topography, we computed SSD on all seven occipito-parietal electrodes contralateral to the flickering stimulus.…”

Reformulating Task-Related Component Analysis for Reducing its Computational Complexity

“…The xTRCA employs a temporal optimization to compensate for trial-by-trial latency variability, providing a better solution to enhance induced potentials than the conventional TRCA. Another extension is group TRCA (gTRCA) designed for group-level analysis [6]. The gTRCA finds linear coefficients that maximize not only trial-by-trial reproducibility within single subjects but also similarity across a group of subjects.…”