Dynamic Modeling of Common Brain Neural Activity in Motor Imagery Tasks

Well, motor imagery has been out there since at least January 2013 and we have been guessing about the protocol since then. This does nothing for survivors. 

• motor imagery (52 posts to January 2013)

 

 Dynamic Modeling of Common Brain Neural Activity in Motor Imagery Tasks

Luisa F. Velasquez-Martinez^*, Frank Zapata-Castano and German Castellanos-Dominguez

• Signal Processing and Recognition Group, Universidad Nacional de Colombia, Manizales, Colombia

Evaluation of brain dynamics elicited by motor imagery (MI) tasks can contribute to clinical and learning applications. The multi-subject analysis is to make inferences on the group/population level about the properties of MI brain activity. However, intrinsic neurophysiological variability of neural dynamics poses a challenge for devising efficient MI systems. Here, we develop a time-frequency model for estimating the spatial relevance of common neural activity across subjects employing an introduced statistical thresholding rule. In deriving multi-subject spatial maps, we present a comparative analysis of three feature extraction methods: Common Spatial Patterns, Functional Connectivity, and Event-Related De/Synchronization. In terms of interpretability, we evaluate the effectiveness in gathering MI data from collective populations by introducing two assumptions: (i) Non-linear assessment of the similarity between multi-subject data originating the subject-level dynamics; (ii) Assessment of time-varying brain network responses according to the ranking of individual accuracy performed in distinguishing distinct motor imagery tasks (left-hand vs. right-hand). The obtained validation results indicate that the estimated collective dynamics differently reflect the flow of sensorimotor cortex activation, providing new insights into the evolution of MI responses.

1. Introduction

Motor imagery (MI) is a dynamic mental state in which an individual performs a mental rehearsal of motor action without any overt output. It is believed that real movements and those performed mentally (imaginary movements) are functionally similar (Stolbkov et al., 2019). Therefore, there is sufficient experimental evidence that MI contributes to substantial improvements in motor learning and performance (Aymeric and Ursula, 2019), games and entertainment, sports training, therapy to induce recovery and neuroplasticity in neurophysical regulation and rehabilitation, and activation of brain neural networks as the basis of motor learning (Machado et al., 2019), and education scenarios (Boe and Kraeutner, 2018; MacIntyre et al., 2018; Suica et al., 2018), where the Media and Information Literacy methodology proposed by UNESCO includes many competencies that are vital for people to be effectively engaged in human development (Frau-Meigs, 2007). These applications reinforce the importance of studying the evolving brain organization to model plastic changes accurately, putting strength on dynamic modeling of temporal, spectral, and spatial features extracted from single channels due to most MI systems rely on them to distinguish distinctive neural activation patterns (Hamedi et al., 2016; Allen et al., 2018).

MI systems handle brain data recorded with electroencephalography (EEG), which is a non-invasive measurement of neural activation and interactions, encoding brain dynamics with high temporal granularity, but at a relatively low spatial resolution (Feng et al., 2019). Integrating spatial filtering techniques can reverse the volume conduction effects to some degree, increasing the EEG spatial resolution. Nevertheless, to enhance the analysis of triggering mental activity, feature extraction approaches are performed to derive distinct EEG spatial maps with varying frequency and time characteristics (Tiwari et al., 2018). To begin with, Filter-Bank Common Spatial Patterns are a popular algorithm in MI systems that discriminate multichannel EEG signals by highlighting differences while minimizing similarities, selecting frequency bands appropriately (Baig et al., 2019). Also, Functional Connectivity (FC) networks are extracted because a better understanding of MI mechanisms requires knowledge of the way the co-activated brain regions interact with each other (Stavrinou et al., 2007). Accordingly, the wPLI metric of EEG functional connectivity can account for linear brain interactions but is also expected to be sensitive to non-linear couplings Imperatori et al. (2019). Another approach for characterizing the imaged hand movements is to quantify frequency alterations in time-varying responses to a stimulus (event) through the so-termed Event-Related De/Synchronization (ERD/S), presenting a significant correlate of localized cortical oscillatory activity (Juan et al., 2019). When imagining one hand moving, an increase/decrease in the power of μ and β rhythms becomes more potent in the sensorimotor (electrodes C3 and C4) and pre-motor (Cz) areas located contralaterally to the hand involved in the task (Wierzgała et al., 2018). Due to the non-stationarity of EEG data, however, the effectiveness of feature extraction procedures is reduced in deriving distinct EEG spatio-spectral patterns. Several factors can affect, among others, the following: movement artifacts during recording, temporal stability of mirroring activation over several sessions differs notably between MI time intervals (Friedrich et al., 2013; Pattnaik and Sarraf, 2018), low EEG signal-to-noise ratio, poor performance in small-sample settings (Park and Chung, 2019), and inter-subject variability in EEG dynamics (Saha et al., 2018). Along with variability in the signal acquisition, another circumstance that leads to low accuracy scores is that some subjects may have brain networks, not sufficiently developed for practicing MI tasks (Ahn and Jun, 2015). As a result, the performance of MI systems varies considerably across and within-subjects, severely degrading their reliability.

To compensate for the variability of EEG dynamics, novel approaches are being developed to integrate information across subjects within a collective framework, combining individual feature sets of neural dynamics to improve the brain representation robustness, as explained in Bigdely-Shamlo et al. (2018). Thus, under the assumption that temporal signatures from an evoked neural activity are similar across subjects, group models can be extracted for decoding the multi-subject mental responses to complex stimuli without explicitly representing the elicitation (Fazli et al., 2015). Several strategies for raw data aggregation can be implemented for building group inferences, including serial/parallel combinations of subject-level feature sets to form a more extensive multi-subject array (Lio and Boulinguez, 2016). Instead, data-driven approaches have also been employed to infer collective feature structures, like joint diagonalization (Gong et al., 2018), temporally constrained sparse representation (Zhang et al., 2019), canonical correlation analysis (de Cheveigné et al., 2019), and versions derived from independent Component Analysis (Emge et al., 2018; Huster and Raud, 2018; Bhinge et al., 2019), among others.

For interpretation purposes, the topographic representation is commonly computed to display the spatial distribution of the extracted common neural dynamics. Nonetheless, the building of multi-subject models implies the accurate aggregation of time-frequency patterns extracted from EEG dynamics across the subjects by adequately selecting the domain parameters (i.e., time window length and filter bandwidth setup) (Huster and Raud, 2018). Moreover, the aggregation can face a different dimensionality derived from the feature extraction methods involved. Due to the difference in captured dynamics, each engaged extraction method differently reflects the flow of sensorimotor, being one of the issues that arise in identifying group relationships confidently (Bridwell et al., 2018). Besides, to evaluate computational network models, there is a need to establish the meaning of the aggregation of extracted brain-activity patterns (Kriegeskorte et al., 2008). Hence, another issue to consider is to assess the ability of multi-subject sets to preserve the main properties (i.e., the spatial distribution of brain neural activity throughout time and spectral domains) extracted from single-subject models.

Here, we develop a dynamic model for estimating the common neural activity across subjects to provide new insights into the evolution of collective mental imagery processes. After the preprocessing stage, the t-f EEG signal set is fed into a feature extraction algorithm to improve the efficiency of triggering activity representation. Then, we employ a statistical thresholding algorithm to extract a multi-subject model that provides a set of confident estimates contributing the most to discriminating between MI tasks. We present a comparative analysis of the feasibility of three popular feature extraction methods in deriving multi-subject spatial maps: Common Spatial Patterns, Functional Connectivity, and event-related de/synchronization. The obtained validation results indicate that the estimated collective dynamics reflect the flow in the sensorimotor cortex activation differently. Therefore, the common model addresses inter-subject and inter-trial variability sources individually, depending on the engaged extraction method.

The paper is organized as follows: section 2 describes the validated database and methods that are carried out; section 3 presents the experimental setup as well as the performed outcomes; section 4 introduces a detailed discussion of the attained results, providing the main conclusions of this work.