Exploring Multimodal Paths in Time-Varying Brain Graphs: Integrating fMRI-Based FNC with sMRI-Based
Poster No:
1831
Submission Type:
Abstract Submission
Authors:
Haleh Falakshahi1, Hooman Rokham2, Vince Calhoun3
Institutions:
1Georgia Institute of Technology, Atlanta, GA, 2Georgia Institue of Technology, Atlanta, GA, 3GSU/GATech/Emory, Decatur, GA
First Author:
Co-Author(s):
Introduction:
The exploration of multimodal brain network analysis shows potential in uncovering the mechanisms at the core of brain disorders. Previous research has predominantly focused on either unimodal brain graphs or local/global graphic metrics, often neglecting the intricate details of disrupted pathways within patient groups. Our study emphasizes the significance of dynamically integrating multimodal brain graphs with a targeted examination of disrupted paths using a Gaussian graphical model (GGM). This approach reveals profound insights that can aid in identifying time-varying, path-based disease biomarkers.
Methods:
In our study, we utilized data from the fBIRN study, incorporating structural MRI (sMRI) and functional MRI (fMRI) data from 160 controls and 151 patients with schizophrenia (SZ). Employing the Neuromark pipeline, we applied a spatially constrained group-independent component analysis to the fMRI data, yielding 53 consistent components representing intrinsic connectivity networks (ICNs). Dynamic functional network connectivity (dFNC) computations were conducted using a sliding window approach, and FNC among segmented ICN time courses was estimated using the GIFT toolbox. Covariance estimation employed a regularized precision matrix obtained through graphical LASSO applied to windowed data, with λ determined via cross-validation. Recurring dFNC patterns were explored using k-means clustering, identifying five states based on the elbow criterion. A univariate test assessed group differences in dFNC, employing t-tests on averaged dFNC values across states and correcting for multiple comparisons using the False Discovery Rate (FDR) method. To create a dynamic multimodal graph, parallel independent component analysis (pICA) was applied to gray matter (GM) and state average features. The optimal number of components, determined using the elbow method, was set at fifteen for GM and fourteen for dFNC states. Loading matrices from pICA were unpacked for each state, and two artifactual GM components were removed. Finally, Gaussian graphical models (GGM) were estimated for control and patient groups for each state, enabling the comparison and analysis of paths between multimodal nodes to identify disconnectors associated with disconnectivity and connectors associated with additional paths (abnormal integration).
Results:
In group difference evaluations, the number of significant FDR corrected p values were 173 for State 1, 55 for State 2, 41 for State 3, 57 for State 4, and 34 for State 5. Figure 1 illustrates the results of the multiple comparison test for the control and SZ group. Figure 2 shows the results of applying path analysis to identify missing and additional links for each state. For example, State 2 shows a multimodal disconnector associated with absent paths between the temporal middle component of gray matter and the FNC component which have a high correlation among sensory-motor, visual, and other domains' ICNs. However, state 2 reveals 23 connectors associated with additional paths both within and between modalities.
·Averaged functional connectivity matrices for control and SZ and group differences. The upper triangle is the differences between the averaged correlation matrix of SZ and control (SZ-C), and the lowe
·Disconnectors and connectors that were identified in each state. Solid red links trigger disconnection (absence of paths), and solid green links trigger abnormal integration (additional paths) in SZ
Conclusions:
In summary, we present an approach to estimate and visualize links within and among multimodal data. Subsequently, we scrutinized pathways capable of discerning connections associated with the absence or presence of links in the patient group. Our approach unveils crucial insights into network disruptions related to the disorder-insights often overlooked when focusing solely on a single modality. The absence of multimodal information hinders our ability to identify significant missing connections in SZ patients, playing a pivotal role in the disconnection or abnormal integration among brain components. This underscores the efficacy of our approach and emphasizes the indispensability of employing multimodal imaging methods in studying intricate mental illnesses.
Modeling and Analysis Methods:
Connectivity (eg. functional, effective, structural) 2
fMRI Connectivity and Network Modeling 1
Keywords:
FUNCTIONAL MRI
Modeling
Schizophrenia
STRUCTURAL MRI
1|2Indicates the priority used for review