Supervised piecewise network connectivity analysis for enhanced confidence of auditory oddball tasks

Abstract

We present a network connectivity analysis to assess a relevant vector, valuing the contribution of connection node sets in distinguishing between labeled response stimuli. To this end, the piecewise computation of Phase Locking Index is performed, suggesting a combination procedure to reflect the whole recording span with a single relevance value. Further, we use a supervised, statistical thresholding algorithm to reduce the connectivity matrix dimension, holding the links that mostly differentiate the brain responses to each evoked stimulus. Obtained results in an auditory oddball task show that the developed analysis yields a relevant node set for δ and θ waves that becomes more coherent, connected with improved consistency of performed group-level connectivity graphs.

Introduction

To perform physiologically relevant analysis, oddball tasks are very frequently employed to identify perceptual differences, providing a deeper understanding of attention and memory tasks in applications of affective computing and media and information literacy [1]. In this regard, to provide significant evidence about brain function and cognition tasks, relationships between behavior and neuroimaging measures are increasingly investigated, mostly using electroencephalography (EEG) due to its portability, affordability, and high temporal resolution. In particular, event-related potential (ERP) components are employed to reflect canonical neural operations, which are brain activities predictably modulated within spatio-temporal windows [2]. This situation enables neuroimaging measures to benefit from tracking the evoked time-variant responses in different brain structures. In this regard, graph-based methods are developed to characterize EEG functional connectivity, intending to provide a more nuanced view of the neural dynamics during target detection/novelty processing in normative and pathological populations [3].

Despite its evident impact, analysis of functional brain networks faces several restrictions: A growing need for high-resolution connectivity measures to supply a delicate balance between local specialization and global integration of brain processes [4]; EEG non-stationarity that makes the brain networks intrinsically and dramatically change over time, degrading the assessment of pairwise interactions, which are typically operationalized through the full or partial correlation/information between all pairs of regional time series [5]; An extraction of all possible inter-channel interactions that may result in high dimensional connectivity matrices, including redundant or worthless features extracted from specific tasks and making the connectivity analysis confidence be diminished because of noisy links (not mentioning the computational cost issues) [6].

To overcome the non-stationary nature of EEG signals, most of the approaches rely on the quasi-stationary activity of large neuronal populations, extracting synchronization measures from a set of previously segmented time intervals, which are modeled (or even statistically tested [7]) as stationary as discussed in [8]. Yet, by dropping segments in multi-trial tasks, several neural dynamics may be omitted due to the modulated spatio-temporal ERP activity. As regards the dimensionality reduction of connectivity matrices, thresholding methods are employed, generally retaining the strongest edges (pairwise interaction), either by holding the edges surpassing a given absolute weight or by constraining the edge density [9]. Nonetheless, each particular thresholding rule influences the number of weak connections, which, in turn, yields a distinct effect on the structure and global properties of sparsified networks [10].

Nevertheless, one of the major concerns is to determine the latent structure from single-subject data sets that should directly generalize across subjects. To this end, two main approaches of group-level analysis are widely considered [11]: Clustering of components estimated from EEG data [12], and concatenation of single-subject data into a group array from which a latent structure of sources is computed, being representative of the sample as a whole [13]. Regardless of the group-level strategy, however, the activity patterns are far from being time-locked across trials perfectly. Moreover, there is high variability across subject samples. Consequently, standard approaches for group-level performance degrade sharply.

Aiming at reducing the variability of a neural process across the whole sample, here, we develop a graph network analysis that estimates a relevant connectivity vector that takes into account the temporal characteristics of neural responses, assessing the contribution of a link node set in distinguishing between labeled ERP stimuli. To this end, we perform piecewise computation of group-level connectivity graphs to deal with the non-stationarity of EEG data that makes the brain networks change over time. Also, taking advantage of the available labels for EEG responses, we use a supervised, statistical thresholding algorithm to reduce the worthless features, holding the connections that differentiate the most the brain responses to each evoked stimulus.