TY - JOUR TI - Computational methods for probing the spatio-temporo-rhythmic characteristics of the task-associated brain functional networks via electroencephalography DO - https://doi.org/doi:10.7282/t3-0gef-bx97 PY - 2021 AB - The human brain is one of the largest networks known to exist. Recent years have witnessed an increasing interest in understanding the large-scale networks in the brain from the functional point of view, especially, within the context of functional connectivity. In functional connectivity studies, the statistical dependencies among signals recorded from different regions of the brain are evaluated to identify the underlying functional networks. Given the dynamic nature of the brain function, these networks are expected to exhibit dynamic behavior and temporal changes. Most studies in this field, however, use statistical measures that rely on the assumed extended intervals of temporal stationarity, e.g., over the entire scan interval, therefore, ignore the dynamics of the brain functional networks. Moreover, identifying how these networks change spatially and temporally during task execution, e.g., within 500 msec of stimulus onset, has been a challenging task, due to the limitations of imaging techniques as well as the employed computational approaches. These problems create the need for the development of data-driven computational methods capable of detecting the dynamic changes in the functional state of the brain, locating the functional networks forming during each state, and identifying the nature of their association to the studied tasks. In this thesis, we aim to study the dynamic functional networks of the brain in the spectral, temporal, and spatial domains, and to assess the task-association properties of the identified networks. The imaging modality considered here, Electroencephalography (EEG), is a non-invasive modality that measures the neuronal electrical activity traveling from the brain to the surface of the scalp. The excellent temporal resolution that EEG offers makes it a good candidate for studying the dynamics of the brain function at multiple temporal scales. Accordingly, a number of computational frameworks are developed, here, for the purpose of probing the spatio-temporal characteristics of the task-associated, multi-scale, dynamic functional networks, through EEG recordings. The first part of this work addresses the problem of the non-stationary dynamics of the brain function. To detect the temporal structure of the functional state of the brain from EEG recordings, we propose a data-driven segmentation algorithm that is informed by the changes in the underlying cortical activity, without having to solve the inverse problem. The proposed method utilizes singular value decomposition (SVD) to identify the time intervals in EEG recordings during which the spatial distribution of clusters of active cortical neurons remains quasi-stationary. Theoretical analysis shows that the spatial locality features of these clusters can be captured, asymptotically, by the most significant left singular subspace of the EEG data. A reference/sliding window approach is employed to dynamically extract this feature subspace, and the running projection error is monitored for significant changes using Kolmogorov-Smirnov (K-S) test. The proposed source-informed segmentation algorithm is comprehensively benchmarked using simulated data, and the results show that the algorithm is successful in detecting the embedded ground-truth segmental structure of the data. Furthermore, the algorithm is applied to experimental EEG recordings of a modified visual oddball task, and the results are shown to be compatible to the findings in the literature on similar tasks. The second part of this work develops a framework that probes the multi-scale, spatio-temporal characteristics of functional networks. We employ the stationary wavelet transform (SWT) to decompose EEG data into the dyadic scales spanning the delta, theta, alpha, beta, and gamma rhythms. The next step in the framework applies the source-informed segmentation algorithm to each rhythm, independently, in order to detect instants of transition in the brain functional state. The last step of the framework involves localizing, in the cortex, the set of functional networks forming during each segment. Each of these networks is characterized by the common time course binding all its cortical sources together. We implement a SVD-based functional network localization approach that estimates these time courses through the most significant right singular vectors. This approach has the ability to discriminate among and localize multiple simultaneous functional networks. It is supplemented here with a data-driven thresholding technique for assessing the contribution level of each cortical source. We apply the developed multi-scale, dynamic functional connectivity framework, which integrates these three steps, to experimental data of a modified visual oddball task to localize the temporal sequence of functional networks, of each rhythm, from stimulus to response. The third part of this work involves three proposed computational frameworks that allow us to extract the spatio-temporal characteristics of the functionality of the brain specific to a given task or discriminatory across two tasks. The first framework considers the problem of searching the multi-trial-based database of functional networks, generated by the aforementioned dynamic functional connectivity framework, to identify the components that are most recurrent throughout these trials, thus, are deemed "specific" to the probed task. We employ the ASSO algorithm to factor the Boolean functional database into its common network components and their occurrence patterns. Tracing back the occurrences to their original timelines, the recurrence dynamics of each functional network component can be calculated. In contrast, the second framework involves identifying the spatio-temporal characteristics of the functional network components that are "discriminatory" across two different tasks. We formulate this problem as finding the components of the functional networks that are recurrent the most in the trials of one task and least in the other. This is translated to determining a fixed-sized set of Boolean-valued basis vectors with the highest difference in recurrence rate between two Boolean matrices, which we call the discriminative discrete basis problem (DDBP). To address the different classes of DDBP, we develop an ensemble of discriminative-associative (DASSO) algorithms. DASSO algorithms use the cross-matrix difference in pairwise correlation to assess the discriminative associations among the elements of Boolean matrices. The algorithms show promising results in a comprehensive benchmark, based on simulated data. We employ DASSO algorithms to factor the Boolean functional databases into the constituent discriminative network components and to extract their recurrence dynamics within each database. Finally, the third framework generalizes that based on the DASSO algorithms by taking into account the patterns, rather than just the aggregate rates, of recurrence of the sought functional network components, thus, effectively targeting those that are "spatio-temporally discriminatory" across two different tasks. This problem is formulated as finding the functional network components with the occurrence patterns most anti-correlated across the trials of the two contrasted tasks. As it determines a fixed-sized set of Boolean-valued basis vectors with such occurrence patterns, we call it the occurrence-informed discriminative discrete basis problem (OIDDBP). We develop an ensemble of occurrence-informed discriminative-associative (OIDASSO) algorithms, which combine warped anti-correlation and pairwise correlation functions to assess the occurrence-informed discriminative associations among the elements of Boolean matrices, along both dimensions. The OIDASSO algorithms, too, show promising results in a comprehensive benchmark, based on simulated data. We employ OIDASSO algorithms to estimate the spatio-temporally discriminative functional network components and their recurrence dynamics, by accordingly factoring the multi-trial Boolean functional databases. The three aforementioned frameworks were applied to experimental data collected during different visually-cued motor execution/imagery/control tasks to identify the commonality and discrimination among them. Towards demonstrating the capability of the frameworks developed throughout our work to extract the characteristics of the brain functional networks and the applicability of these frameworks in decoding human intention, we employ them in two classification systems. The first system integrates a dynamic feature extractor comprising the source-informed segmentation algorithm and the SVD-based functional network localization approach, with the common spatial patterns (CSP) analysis optionally employed for selecting the most discriminatory features, and a long short-term memory (LSTM) classifier. The resultant dynamic classification system can make a decision as early as the first segment, then progressively update that decision at the end of each following segment. The results of classifying motor execution against motor imagery tasks demonstrate the ability of the proposed dynamic system to provide early detection of human intention, with average accuracy higher than 75% within the first 500 msec of the task planning or task performance phase. In the second system we use the aforementioned multi-scale, dynamic functional connectivity framework, instead, for a more comprehensive feature extraction approach. The DASSO algorithm is then employed for selecting the most discriminatory features which are finally passed to a recurrent neural network (RNN) classifier. This dynamic system is used to classify motor execution against motor imagery tasks. Our results show that an average accuracy above 80% can be achieved within the first 500 msec of the task planning or performance phase. KW - Brain functional networks KW - Artificial intelligence KW - Electroencephalography KW - Electrical and Computer Engineering LA - English ER -