NEUROCOMPUTATIONAL AND DATA MINING TECHNIQUES TO DECODE TASK ENGAGEMENTS THROUGH A MULTIMODAL APPROACH

Abstract

Decoding complex brain functioning is a longstanding challenge in cognitive neuroscience. Over the last few decades, several remarkable studies advanced the more profound understanding of brain mechanisms from temporal to spatial scale and local to global cortical communications. To decipher the mechanisms underlying various cognitive engagements needs such a system that can integrate human cognitive mechanisms to the vital knowledge of neuroscience. Thus, developing a multimodal approach capable of extracting the information and utilizing them for exploring the inner neurocognitive processes will contribute to a potential understanding of complex brain behavior. In line with this approach, the simultaneous EEG-fMRI technique is a promising noninvasive data acquisition technique that exhibits highly complementary characteristics for studying a broad range of brain functions with millimeter spatial resolution and millisecond temporal resolution. However, the hemodynamical mechanisms of the task stimulus induce a complex relationship between neural activity and associated response. The researchers [1,2] have observed that neural activity's HRF (Hemodynamic Response Function) response is sluggish and spreads over time depending upon the intensity and duration of neural activity. Hence, there is a need to consider stimuli and voxel-specific hemodynamic delay and develop an algorithm to explore task-related neural mechanisms. Also, behavioral factors such as task familiarity, practice, current brain state, feedback, and intelligence level of individuals lead to significant changes in an individual's brain functioning and performance. A more profound understanding of these factors has been found essential while designing work to decode the complex neurocognitive mechanisms. The central objective is to decipher the neuronal mechanisms underlying the distinct cognitive engagements by building on the strengths of the multimodal approach of neurocomputational and data mining techniques. For this purpose, this thesis develops a comprehensive understanding of cognitive aspects and subsequent modulations in neuronal mechanisms by combining the proposed and conventional vii neurocomputational methods. The cognitive engagement of the target detection task is explored to understand the neurovascular coupling and task-related modulations using the integrated voxel-wise analysis of functional connectivity modulation with the neurovascular underpinning changes. For this purpose, a novel approach of reorganizing the blood oxygen level-dependent (BOLD) information has been proposed utilizing the HRF parameters. The findings are validated through a graphtheoretical approach at the selected regions. The study revealed the strong association of neurovascular underpinnings with the modulation of functional networks and the associated neuronal activity during task engagement. Further, to understand the indepth knowledge of the factors affecting an individual’s performance during task engagements, the role of feedback learning and fluid intelligence in the task are explored. Functional connectivity estimation and neurovascular approaches analyze feedback learning and its effect on subsequent task engagement. Our findings strengthened the associations of explicit frontal-parietal EEG cortical oscillations with the local and global neuronal engagements. Another effort to explore the role of an individual’s fluid intelligence is understanding the neural mechanisms of intelligence associated with feedback learning and subsequent memory retrieval of encoded objects. The study proposed a deep learningbased framework to classify encoded objects from distractors. Subsequently, neural mechanisms of multifrequency elicitations, which correlate with intelligence, are studied through the multifrequency interaction, EEG-informed-fMRI, and hemodynamical functional connectivity approaches. Finally, an attention-based LSTM deep-learning model is successfully employed to decipher these distinct EEG oscillations during encoded objects’ memory retrieval and learning engagement. Thus, the study brings more insights into the multifrequency cortical engagement caused by frontal-parietal interaction of intelligence-associated feedback learning and memory retrieval of encoded objects. Furthermore, decoding cognitive-task engagement from coarse EEG information is still challenging due to the underlying complex local and distant cortical neural elicitation and communication. This study proposes an attention-based deep learning framework that processes novel, optimized, quasi-stable frequency microstates as viii input information. The frequency microstate was optimized using hemodynamic functional connectivity measures estimated using graph theory analysis. The local and distant cortical neural mechanisms associated with each optimized frequency microstates were analyzed using microstate-informed fMRI to use them as neural signatures of cognitive task engagement. The optimized, quasi-stable frequency microstates were used as input information to train and validate the attention-based Long Short-Term Memory (LSTM) architecture to classify the neural engagement of the target from the distractor cognitive tasks. The study demonstrates better classification results of complex cognitive task neural mechanisms, using an attention-based deep learning framework utilizing optimized quasi-stable frequency microstates.