High-Model Order ICA and Component Postprocessing

Preprocessed data were subjected to a high-model order ICA using the GIFT toolbox infomax algorithm to decompose the group data into 100 independent components (IC) (Figure 1). The number of components for the ICA was based on previous research using a 100 component ICA to investigate d-FNC, that shows this number of components achieves sufficient functional parcellation of major brain systems such as the visual, sensorimotor, DMN, and SN into individual brain areas (Allen, et al., 2014; Damaraju, et al., 2014; Rashid, et al., 2014). Parcellation of the major brain networks is an important initial aspect of a d-FNC analysis, as one of the strengths of such an approach is that it enables investigation of how individual brain areas interact within and across major brain systems, something that is not possible with lower model order ICAs that provide course divisions along brain systems and not individual brain areas. Additionally, previous research has shown that model orders above 100 components show a decrease in ICA repeatability (Abou - Elseoud, et al., 2010). Stability of the IC estimations was ensured by repeating the ICA algorithm 20 times using ICASSO in GIFT. Subject-specific spatial maps and time-courses were acquired using the ‘GICA1’ back-reconstruction approach in GIFT (Erhardt, et al., 2011).

Schematic of analysis steps. (A) high-model group ICA (100 components) creates a functional parcellation of the brain resulting in 52 non-noise components. (B) Subject specific time courses from the group ICA are then used to calculate functional connections. The static analysis entails computing correlations across the entire duration of the rsfMRI scan. The dynamic analysis utilized 45 second tapered-sliding windows slid in 1TR to acquire 237 correlation matrices for each subject (one per window). Connections between each insula subdivision and all other ICs were then extracted for data anaylysis. (C) A concatenated data matrix consisting of all insula subdivision correlations × each window for each subject (237 windows × 31 subjects) was subjected to k-means clustering using values 2-20 that identified the optimal k as 5 using the elbow criterion. K-means clustering using a value of k = 5 then assigned each window to dynamic state k regardless of subject assignment. Subject specific medians were then back-reconstructed for each state k before they were averaged together to produce the final five dynamic insula states.

Independent components were visually inspected following the protocol of previous research to separate noise from non-noise ICs (Damoiseaux, et al., 2006; Uddin, et al., 2013). Those components containing artifacts such as head motion, white matter, cerebrospinal fluid, or containing large amounts of high frequency information were discarded, leaving 52 ICs representing a functional parcellation of cortical and subcortical brain areas. Of the 52 non-noise ICs, the four that fell within the insular cortex (dorsal anterior, ventral, posterior, and middle insula) were the focus of the current investigation. Time courses of the selected 52 ICs were triple detrended (linear, cubic, quadratic), despiked, low-pass filtered (.15 Hz), and subjected to linear regression of the Friston 24 head motion parameters (6 motion parameters of each volume, the preceding volume, and the 12 corresponding squared items) (Friston, et al., 1996) as calculated by the DPARSF-A toolbox. Despiking in GIFT utilizes AFNI's 3dDespike algorithm that replaces signal spikes larger than the absolute median deviation with a third order spline fit to clean portions of neighboring data. This approach to outlier removal is similar to that of the “scrubbing” method (Power, et al., 2012) with the advantage that no volumes are deleted and temporal information is retained for the sliding window approach. This method has been shown to improve the root-mean-square of the temporal derivative across time courses of ICs (“DVARS” in Power et al., 2012) and reduce the impact of outliers on subsequent functional connectivity analyses (Allen, et al., 2014).