Movie-specific temporal dynamics analysis in functional connectivity using the COBE method
Poster No:
1352
Submission Type:
Abstract Submission
Authors:
Priyanka Mittal1, Anil Sao2, Bharat Biswal3
Institutions:
1Indian Institute of Technology, Mandi, Himachal Pradesh, 2Indian Institute of Technology, Bhilai, Chattisgarh, 3New Jersey Institute of Technology, Newark, NJ
First Author:
Co-Author(s):
Introduction:
Temporal dynamics of brain states, measured using fMRI, could be associated with a sequence of well-defined functional states and can be derived using dynamic functional connectivity (FC). It has been reported that in fMRI scans made during movie watching, the transitions of functional states are temporally aligned to specific features of the movie (Meer et al., 2020). Leveraging the Hidden Markov Model (HMM), such dynamics of whole-brain networks can be modeled as hidden states, where each state can be described by the mean activation vector (Hunyadi et al., 2019; Vidaurre et. al., 2016, 2018a). Most frequently observed states for a window of time may be different across the subjects due to individual-level variations despite they have undergone the same movie stimuli during the scan. In this study, we attempted to suppress the subject level variations present in the dynamic FC observed during movie-watching fMRI data through the Common orthogonal basis extraction (COBE) algorithm. This leads to enhanced consistency in the state's temporal dynamics across the subjects that have undergone the same movie stimuli.
Methods:
We have analyzed movie fMRI data from the publicly available dataset called the "HCP S1200 release." This included data from 184 individuals who underwent fMRI scanning while watching movies on a 7 Tesla scanner at the University of Minnesota (https://db.humanconnectome.org). The stimuli for MOVIE1_CC1 and MOVIE3_CC2 consisted of four short clip compilations consisting of independent Creative Commons videos. MOVIE2_HO1 and MOVIE4_HO2, on the other hand, consisted of scenes from Hollywood movies. In this study, dynamic FC matrices are computed using 160 regions of interest (ROIs) from the Dosenbach atlas using point-by-point multiplication of z-scored ROIs. To estimate the HMM states from the group of dynamic FC matrices related to four different movie stimuli, we have used the HMM-MAR toolbox (Vidaurre et. al., 2016). Further, we applied the Common Orthogonal Basis Extraction (COBE) method (Zhou et al., 2016) to extract common FC patterns across subjects for a movie. Thus movie-specific information present in dynamic FCs is extracted as the common subspace and subject level variations are separated from the FC. For the analysis of movie-specific temporal dynamics present in the dynamic FC, HMM states are extracted from the common subspace of FC across subjects.
Results:
Figure 1 shows the workflow pipeline where HMM states are estimated using group analysis of all four movie stimuli-related dynamic FC matrices. HMM states were also computed in this study after the COBE application on the dynamic FC matrices of each movie stimuli separately. Using the COBE algorithm, the common information present in the dynamic FC which is movie-related (as movie stimuli are common across the subjects for a particular stimuli) is retained after removing the individual components of the COBE decomposition of FC matrices. Using the HMM model, the state path followed by each of the subjects across time can be observed. Figure 2 shows the most frequent states for both cases (with and without using COBE) when a nonoverlapping time window of 10-time points is considered for all the subjects. It is evident from Figure 2 (b) that after COBE application, the state path of most frequent states across the subset of subjects is more consistent than the previous case Figure 2 (a). For all four movies, after the COBE application, fewer states are frequent and comparatively consistent across the subjects.
·Figure 1: Workflow pipeline for extracting HMM states directly from dynamic FC matrices and after performing COBE on dynamic FC matrices while retaining their common information.
·Figure 2: Most frequent States over the time windows across subjects, (no. of states =8) extracted: (a) directly from dynamic FC matrices (b) after applying COBE with no. of components = 30
Conclusions:
In this study, using the COBE algorithm, we have attempted to separate the individual-level variation present in FC matrices and one can observe its effect in the form of consistent state transition across the subjects over the time window. The HMM method is used in this work to model the fMRI data as transitions of some hidden states which will be aligned to the specific features of their movie stimuli.
Modeling and Analysis Methods:
Activation (eg. BOLD task-fMRI) 1
Connectivity (eg. functional, effective, structural) 2
fMRI Connectivity and Network Modeling
Keywords:
Data analysis
FUNCTIONAL MRI
1|2Indicates the priority used for review
Provide references using author date format
Meer, J.N.V.D. et al. (2020), Movie viewing elicits rich and reliable brain state dynamics. Nature communications, 11(1), p. 5004.
Vidaurre, D. et al. (2016), Spectrally resolved fast transient brain states in electrophysiological data. NeuroImage, 126, pp. 81–95.
Vidaurre, D. et al. (2018b), Spontaneous cortical activity transiently organizes into frequency specific phase-coupling networks. Nature Communications, 9(1), p. 2987.
Zhou, G. et al. (2015), Group component analysis for multiblock data: Common and individual feature extraction. IEEE transactions on neural networks and learning systems, 27(11), pp. 2426-2439.