The present study presents the times series computational frameworks that classify different task engagement based on temporal modulation of distant brain communications through optimized frequency EEG microstates. The frequency EEG microstates were optimized by associating them with simultaneously eliciting distant hemodynamic functional connectivity measures. Further, the study explored the EEG-informed fMRI approach that has been employed to understand the insights into the neuronal mechanisms associated with frequency microstates that correlate with global task communications. The following sections present the results of each of the above steps in detail.
3.1 Frontal, parietal, and temporal cortical interaction elicited during different task engagements:
The neural correlates associated with each task engagement are assessed through GLM models in fMRI information with the double-sided t-test of p < 0.5, FDR corrected. The results suggest that target engagement enhanced the hemodynamic response in the frontal (frontal orbital cortex (FOC), frontal pole (FP), superior frontal gyrus (SFG), parietal (angular gyrus (AG), Precuneus cortex (PC)), and temporal (Inferior and middle temporal Gyrus (ITG, MTG)) cortices. Further, it involves significant engagement of the cingulate gyrus (CG), lateral occipital cortex, occipital pole (OP), paracingulate gyrus, and insular cortex (IC) regions. The neural correlates of distractor and fixation have also revealed distinct intercortical engagement. The detailed list of neural correlates of each task engagement is tabulated in Supplementary Table (ST.1).
The presence of this distinct inter cortical communication during each task engagements are furthur supported by the graph theoritical functional connectivity metrics such as global (GE) and local efficiency (LE). The target engagement significantly engaged inter and intra-cortical communication at frontal cortex (frontal pole [GE:0.893, LE:0.89], frontal orbital cortex [GE: 0.886, LE:0.89], and superior frontal gyrus [GE:0.886, LE:0.894]), parietal cortex (left angular gyrus [GE:0.872, LE:0.893], right angular gyrus [GE:0.9, LE:0.887], precuneus cortex [GE:0.87, LE:0.894]) and temporal cortical regions (inferior temporal gyrus [GE: 0.88, LE: 0.889], middle temporal gyrus [GE:0.889, LE: 0.891], temporal occipital fusiform cortex [GE:0.88, LE:0.89]) at p < 0.05 with FDR correction.
Functional connectivity elicitation remained distinct during this target detection task and had minimal overlap in neural mechanisms, regions, and intracortical interaction with each task engagement. Notably, frontoparietal and frontotemporal interaction engagement was distinct at the neuronal level for target, distractor and fixation engagement. The detailed information of graph-theoretical measures estimated for each task engagement is given in the supplementary file (section S.2).
3.2 Quasi-stable frequency-microstates and their association with task's fMRI functional connectivity measures
As mentioned in the data analysis section, the preprocessed band passed EEG information of every task engagement (target, distractor, fixation) is subjected to the frequency-microstate estimation. Four dominant frequency microstates are estimated for every EEG frequency band of each task engagement. Figure 4 illustrates spatial topographical patterns of each frequency-microstate topography associated with every task engagement.
Further, the number of occurrences of every task's frequency-microstate prototype is estimated by back-fitting these frequency-microstates to every volunteer's respective frequency information. Figure 5 compares the mean number of occurrences of all volunteers of each frequency-microstate across every task. Most of the researchers in the literature use the alphabetical approach to label each microstate. This nomenclature brings difficulty in following similar observations across the literature. Hence, in this study, the unique directional pattern of each microstate's moderate activation band (green colour band) is used for labelling these microstates. In this study, based on the microstate's green colour band direction, they are labelled as anterior-posterior (AP), left-right (LR), left diagonal (LD), and right diagonal (RD) microstate prototypes. For example, the delta microstate with a green band travelling between anterior-posterior is called the "anterior-posterior delta microstate". As the study primarily focuses on identifying the frequency-microstate that manifests the cortical communication elicits during the task engagement, every frequency microstate's number of occurrences is robustly correlated with the fMRI functional connectivity measures. The significantly (p < 0.01) correlating frequency microstates were colour-coded (+ ve correlation: green, -ve correlation: red) in Fig. 5.
The significantly associated, optimized frequency microstates with graph-theoretical measures of every task engagement are illustrated in Fig. 6. A total of twelve frequency microstates (target: 4, distractor: 4 and fixation: 4) significantly correlated with the functional connectivity measures of simultaneously acquired fMRI information. During target engagement, the number of occurrences of right diagonal delta-microstate positively correlates with the global efficiency of fMRI functional connectivity measures. On the other hand, the local efficiency of fMRI connectivity measures correlates negatively with both left-right theta and alpha microstate occurrences and positively with anterior-posterior theta microstate.
3.3 Neurovascular analysis of optimized frequency microstates: EEG Informed fMRI analysis
Engagement of task generally elicits local neural clustering (high-frequency quasi-stable oscillations) at distinct brain regions, responsible for efficient local information processing, together with distant cortical intercommunication to facilitate global communication (low-frequency quasi-stable oscillations). The neuro-vascular analysis through EEG-informed fMRI explains these insights through synchronizing neural information (optimized quasi-stable EEG oscillations) with hemodynamic information (vascular) that is elicitated from a specific task engagement. For this purpose, each one of twelve optimized frequency microstates is processed in independent EEG-informed fMRI models, which modelled each microstate as independent regressors (p < 0.01, FDR corrected) to estimate their neuro-vascular information. Supplementary figures (SF.1(a-c)) show neuro-vascular information of each twelve optimized microstates.
Figure 7 shows the neuro-vascular coupling of each significant frequency-microstate at the neural correlates of the respective task engagement. The figure also shows the functional connectivity between each neural correlate of task engagement. It is further evident from Fig. 7 that fMRI functional connectivity optimized frequency microstate associates with almost most of the brain regions involved in each task engagement. Target engagement elucidated the right-diagonal delta-microstate synchronously with the BOLD response in the frontal cortex. Further, anterior-posterior and left-right theta-microstates are found to synchronize with the BOLD response of frontal, temporal, parietal, and occipital regions except for right-lateralized SFG and AG. The role of theta-microstate in BOLD-synchronization of these regions characterizes its significant association with target engagement. The alpha-microstate is observed de-synchronizing with the parietal, PCG's BOLD response, and synchronizing with the frontal, occipital region during target engagement. Our findings also reveal the effects of multiple frequencies on specific brain regions, such that the BOLD response of FP, PCG, SFG, IC, PC, and AG are modulated with delta, theta, and alpha-microstates. Specifically, the study observes the de-synchronization of alpha-microstate with the delta and theta microstates such as PCG, SFG, PC, and synchronization between delta and theta microstates as IC and FP. Hence, the relationships mentioned above reveal that the multi-frequency interactions modulate the BOLD response of task-engaged brain regions.
Performance of Deep Attentional LSTM Model:
The performance of the proposed deep attentional model revealed that hybrid deep learning architecture allows it to apply the attention layer that finds meaningful patterns using LSTM by overcoming the fixed-length input sequences. Figure 8 compares the performance metrics, precision, accuracy, and recall of all the six deep learning models (LSTM and attention-LSTM) in classifying the different task engagement based on the stacked feature vectors that consist of correlation information of each optimized frequency microstate. Three distinct segregation of input feature vectors (200ms/segment, 300ms/segment, and 500ms/segment) with 15, 10, and 6-LSTM cells in input layers were analyzed within the context of each type of network (LSTM and attention-LSTM). As can be observed, LSTM combined with attention appears to perform significantly better in three of the performance metrics. In this case, the choice of nodes (15, 10, and 6) that depends on the EEG time series window (200ms, 300ms, and 500 ms) seems reasonable since it reflects a significant variation in the precision, accuracy, and recall. The results reflected improved performance for using ten nodes from six nodes; a significant decline in performance metrics is observed for using 15 nodes from 10 nodes. In this regard, the proposed attention-based LSTM architecture has proven to enhance the model performance. However, the choice of nodes for the first layer of LSTM architecture leads to variation in deep learning architecture performance.