Dynamic functional network connectivity reveals unique and overlapping profiles of insula subdivisions

Abstract

The human insular cortex consists of functionally diverse subdivisions that engage during tasks ranging from interoception to cognitive control. The multiplicity of functions subserved by insular subdivisions calls for a nuanced investigation of their functional connectivity profiles. Four insula subdivisions (dorsal anterior, dAI; ventral, VI; posterior, PI; middle, MI) derived using a data‐driven approach were subjected to static‐ and dynamic functional network connectivity (s‐FNC and d‐FNC) analyses. Static‐FNC analyses replicated previous work demonstrating a cognition‐emotion‐interoception division of the insula, where the dAI is functionally connected to frontal areas, the VI to limbic areas, and the PI and MI to sensorimotor areas. Dynamic‐FNC analyses consisted of k‐means clustering of sliding windows to identify variable insula connectivity states. The d‐FNC analysis revealed that the most frequently occurring dynamic state mirrored the cognition‐emotion‐interoception division observed from the s‐FNC analysis, with less frequently occurring states showing overlapping and unique subdivision connectivity profiles. In two of the states, all subdivisions exhibited largely overlapping profiles, consisting of subcortical, sensory, motor, and frontal connections. Two other states showed the dAI exhibited a unique connectivity profile compared with other insula subdivisions. Additionally, the dAI exhibited the most variable functional connections across the s‐FNC and d‐FNC analyses, and was the only subdivision to exhibit dynamic functional connections with regions of the default mode network. These results highlight how a d‐FNC approach can capture functional dynamics masked by s‐FNC approaches, and reveal dynamic functional connections enabling the functional flexibility of the insula across time. Hum Brain Mapp 37:1770–1787, 2016. © 2016 Wiley Periodicals, Inc.

INTRODUCTION

Situated at the crossroads of cognitive, homeostatic, and affective systems in the brain, the human insular cortex is a particularly diverse structure, participating in pain processing [Wager et al., 2013], interoception [Craig, 2002; Craig, 2009], sensory [Craig et al., 2000], affective [Barrett and Wager, 2006; Zaki et al., 2012], and high‐level cognitive processes [Menon and Uddin, 2010; Sridharan et al., 2008]. The myriad of cognitive processes involving the insula has recently led to its identification as an important hub for coordinating interactions between large‐scale neurocognitive networks [Uddin, 2015]. Parcellation, seed region‐of‐interest based functional connectivity analyses, structural MRI, and task‐based meta‐analyses reveal the insular cortex as a heterogeneous region consisting of multiple functionally and structurally distinct subdivisions.

Previous studies have divided the insula into various numbers of subdivisions that are dictated by the particular parcellation strategies employed [Cauda and Vercelli, 2012]. For example, several studies have divided the insula into two subdivisions using k‐means clustering of insula voxels [Jakab et al., 2012], structural connections [Cloutman et al., 2012], clustering of a priori instantiated ROIs [Cauda et al., 2011], and meta‐analytic clustering of task‐activation areas [Cauda et al., 2012]. Other studies have divided the insula into three subdivisions using clustering of resting state functional connectivity patterns [Chang et al., 2012; Deen et al., 2011] and into four subdivisions using meta‐analytic approaches [Kurth et al., 2010] and approaches automatically estimating the number of clusters [Ryali et al., 2012]. Finally, there have been several studies creating more than four insula subdivisions using clustering methods [Kelly et al., 2012] and differences in functional connections according to whole‐brain parcellations [Gordon et al., 2014; Power et al., 2011; Yeo et al., 2011].

While there have been many studies that have parcellated the insula, only some of these studies have attempted to identify the functional relevance of the resulting insula subdivisions. These studies have identified two, three, or four functionally relevant insula subdivisions related to cognitive processes. The insula can be coarsely subdivided according to a general anterior‐posterior distinction, with the anterior insula having connections with frontal, anterior cingulate, and parietal areas involved in cognitive control and affective processes while the posterior insula (PI) connects with temporal, and posterior cingulate areas involved in sensorimotor processes [Cauda et al., 2012; Cauda et al., 2011]. Studies proposing three insula subdivisions suggest that the dorsal anterior insula (dAI) has connections with frontal, anterior cingulate, and parietal areas and is involved in cognitive control processes, the ventral anterior insula (vAI) has connections with limbic areas and is involved in affective processes, and the mid‐posterior insula has connections with sensorimotor brain areas and is involved in sensorimotor processing [Chang et al., 2012; Deen et al., 2011; Uddin et al., 2014]. Finally, a four‐insula division scheme mirrors that of the three‐insula division, but proposes that the central insula is related to olfacto‐gustatory processes [Kurth et al., 2010].

The relationship between insula subdivision specialization and cognitive processes generally adheres to a tripartite cognition‐emotion‐interoception framework [Chang et al., 2012; Deen et al., 2011; Uddin et al., 2014] where the dAI plays a role in high‐level cognitive processes including network switching, attention, and inhibition, the vAI appears to be more involved in affective processes and the PI engages in more basic sensory functions [Cauda et al., 2011; Chang et al., 2012; Deen et al., 2011; Kurth et al., 2010]. A cognition‐emotion‐interoception framework is also in accord with anatomical information from the macaque in which three insula subdivisions can be delineated based on cytoarchitectonic divisions; a dorsal anterior portion composed of dysgranular tissue, a ventral anterior portion composed of agranular tissue, and a posterior portion composed of granular tissue (Mesulam and Mufson, 1982a). In the macaque, anterior regions share connections with the anterior cingulate gyrus and the posterior region shares connections with the middle cingulate near the supplementary motor area [Luppino et al., 1993; Mesulam and Mufson, 1982a; Mesulam and Mufson, 1982b; Mufson and Mesulam, 1982; Vogt and Pandya, 1987]. Additionally, studies utilizing diffusion weighted imaging techniques in humans have identified structural connections between the anterior portion of the insula with frontal, limbic, and temporal areas while the PI has structural connections with parietal and posterior temporal areas [Cerliani et al., 2012; Cloutman et al., 2012; Dennis et al., 2014].

Of these three subdivisions, particular attention has been focused on the dAI due to its role as a key node in the salience network (SN), alongside the anterior cingulate, subcortical, and limbic structures [Seeley et al., 2007]. In this context, the dAI facilitates the detection of salient exogenous or endogenous stimuli and coordinates network switching between the default mode and central executive networks (DMN and CEN) [Menon and Uddin, 2010]. Task‐based investigations have also demonstrated that the dAI is the most flexible insula subdivision that is involved in more diverse and domain general processes than the other subdivisions [Kurth et al., 2010; Uddin et al., 2014; Yeo et al., 2014]. These studies point to a unique role of the dAI in contributing to higher level cognitive processing across a wider range of functional domains compared with other insula subdivisions.

In the previous literature demonstrating a “tripartite” cognition‐emotion‐interoception organization of the insula, functional connections of the insula have been regarded as static. In static functional network connectivity (s‐FNC) approaches, functional connections are assessed via correlations across the duration of a scan or task. However, functional connections between brain areas are increasingly understood to be dynamic, where connections may transiently emerge between brain areas over time (Calhoun et al., 2014; Hutchison et al., 2013a). Time‐varying patterns of functional connectivity are captured using “chronnectome” approaches that take into account dynamic coupling between brain regions [Calhoun et al., 2014].

A recent technique called dynamic functional network connectivity (d‐FNC) assesses dynamic functional connections by dividing resting state fMRI (rsfMRI) scans into series of “sliding windows” [Allen et al., 2014; Chang and Glover, 2010; Damaraju et al., 2014; Sakoğlu et al., 2010]. Analysis of these sliding windows can identify changes in functional connections on the order of seconds rather than minutes, and offers a much more nuanced characterization of dynamic brain function. Although the sliding window approach can help identify dynamic functional connections, it exponentially complicates interpretations by producing extremely large data sets. In response, several approaches have been developed that identify statistically relevant changes in dynamic connections across large numbers of sliding windows. These methods include using time‐frequency information [Chang and Glover, 2010], clustering nodes from graph‐theory [Schaefer et al., 2014], measuring changes in variation using novel test‐statistics [Zalesky et al., 2014], and hierarchical/k‐means clustering methods [Allen et al., 2014; Yang et al., 2014].

We hypothesized that the functional flexibility of the insula, and the dAI in particular, results from its ability to dynamically couple with discrete brain regions. The current study employed k‐means clustering of sliding windows [Allen et al., 2014] to capture functional dynamics of insula subdivisions and compare these results to static functional connections of the insula. Comparing s‐FNC and d‐FNC approaches within the context of the same study may help to better characterize the function of the insula for two reasons. First, the insula's function as a network hub may emerge from dynamic functional connections that cannot be completely captured by s‐FNC approaches. Second, different analytic approaches may capture different aspects of functional connections of the insula that may also lead to discrepancies across the literature. The current study aimed to utilize information from both s‐FNC and d‐FNC analyses in order to clarify the functional role of the insular cortex and its subdivisions in the context of whole‐brain functional connectivity.

MATERIALS AND METHODS

Participants

Data consisted of rsfMRI scans of 81 adult participants between 18‐40 years of age downloaded from the Nathan Kline Institute database (http://fcon_1000.projects.nitrc.org/indi/pro/nki.html). From this initial dataset, participants who were left‐handed, had traits of depression (Beck Depression Inventory > 10) or anxiety (State Trait Anxiety Inventory > 40), and more than 1.5 mm in translational head motion and 1.5 degrees of rotational head motion were excluded from the analysis. These strict exclusion criteria resulted in 31 adult participants analyzed in the current study (14 female; M = 25.61 years of age, SD = 5.28; mean Wechsler Abbreviated Scale of Intelligence (WASI)‐Full: [missing two participants] M= 112.62, SD = 10.21, range = 94‐130; framewise displacement (FD): M = 0.12, SD = 0.06).

Imaging was performed on a Trio Siemens 3.0 T scanner that collected a T2* anatomical image and 10:55 minutes of rsfMRI data (260 volumes; TR = 2.5 secs) while participants were instructed to keep their eyes open and fixate on a central cross that was presented on a screen. Of note, we opted not to use imaging data that was acquired using low TR multi‐band protocols because the current d‐FNC methodology was developed using imaging data with a TR of 2s [Allen et al., 2014] and because of the unknown effect of increased autocorrelation found in low TR imaging data on d‐FNC [Smith et al., 2013]. An interleaved EPI sequence was utilized (38 interleaved slices, TE = 30 ms, flip angle = 80 degrees, slice thickness = 3.0 mm, field of view = 216 mm, no gap, resolution = 3.0 × 3.0 × 3.0 mm³) for fMRI data acquisition.

Data Preprocessing

Preprocessing was conducted using the DPARSF‐A toolbox (http://rfmri.org/DPARSF). Steps included deleting the first five volumes for image stabilization, slice‐time correction, motion correction, realignment to the T2* anatomical image, normalization (2 × 2 × 2 mm) to the Montreal neurological Institute template (MNI), and spatial smoothing (FWHM = 6 mm). Finally, the GIFT toolbox (http://mialab.mrn.org/software/gift/) was used to normalize variance of the images following previous methods of d‐FNC analyses to focus on temporal modulations [Allen et al., 2014] and increase subcortical neural network resolution [Damaraju et al., 2014].

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) (Fig. 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].

Figure 1.

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 x each window for each subject (237 windows x 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. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

ICs 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].

Static Functional Network Connectivity

Following previous work exploring s‐FNC before computing d‐FNC, correlations between all IC pairs were computed using the MANCOVAN toolbox in GIFT [Allen et al., 2014; Allen et al., 2011]. The MANCOVAN toolbox performs both multivariate and univariate significance tests on the correlation between each pair of ICs for each subject individually before calculating the grand average of each IC pair across all subjects. Correlations were fisher‐z transformed before all analyses.

In order to identify significant positive connections between insula subdivisions with other ICs, 51 one‐sample t‐tests were conducted for the connection values of each subdivision using a corrected alpha value of P < 9.8039 × 10 ^{− 4} (Bonferroni corrected at .05/51). This allowed us to determine, for example, if the dAI had a significant positive connection with the frontal pole IC.

In order to assess differences in connection strength between insula subdivisions with other ICs, t‐tests were then conducted between functional connectivity values for each insula subdivision with each IC resulting in 300 t‐tests (six comparisons: anterior vs. middle, anterior vs. posterior, anterior vs. ventral, middle vs. posterior, middle vs. ventral, posterior vs. ventral; each comparison was conducted across the remaining 50 ICs) using a corrected alpha value of P < 1.67 × 10 ^{− 4} (Bonferroni corrected at .05/300). This allowed us to determine, for example, if the dAI had a significantly stronger connection with the frontal pole IC than did the PI, MI, or VI.

Dynamic Functional Network Connectivity and Clustering Analysis of Insula Subdivisions

Dynamic‐FNC was computed across all ICs using a sliding window approach implemented in the d‐FNC toolbox in GIFT (Fig. 1). The entire rsfMRI scan for each subject was divided into sliding windows of 18 TRs (45 secs) in steps of 1 TR creating 237 rectangular windows that are convolved with a Gaussian of sigma 3 TRs to obtain tapering along the edges in order to reduce possible noise due to a small number of time points in each window. A window size of 45 seconds was chosen according to previous d‐FNC analyses using window sizes of 44 [Yang et al., 2014] and 45 seconds [Allen et al., 2014; Damaraju et al., 2014; Rashid et al., 2014] and according to previous research showing that window sizes between 30 and 60 seconds capture additional variations in functional connectivity not found in larger window sizes (Allen et al., 2014; Hutchison et al., 2013b). Additional methodological work exploring ideal sliding window size has also supported the use of sliding windows between lengths of 30–60 seconds [Leonardi and Van De Ville, 2015], a finding that has empirical backing showing that cognitive states can be discerned within such windows [Shirer et al., 2012; Wilson et al., 2015].

Following previous d‐FNC analyses [Allen et al., 2014; Damaraju et al., 2014; Yang et al., 2014], the graphical LASSO method [Friedman et al., 2008] was used to apply a L1 regularization to the inverse covariance matrix to further reduce possible noise present due to the limited number of data points in each window [Varoquaux et al., 2010]. The regularization parameter lamba (λ) was optimized for each subject independently [Allen et al., 2014; Damaraju et al., 2014; Yang et al., 2014] producing a regularized correlation matrix consisting of 237 windows × 1,326 functional connections (52 × 51/2 = 1,326) for each subject. For each window, correlations between each insula subdivision and all other ICs were extracted to produce an array of 237 windows × 204 functional connections (four insula subdivisions consisting of 51 pairs of connections for each subdivision) for each subject. Each window is a 1 × 204 vector. All windows for each subject were then concatenated to form an array of 7,347 windows (31 subjects × 237 windows) × 204 functional connections. Correlations were then fisher‐z transformed before all analyses.

An important step in k‐means clustering is choosing the number of cluster centroids. Previous work has demonstrated that the optimal number of clusters can be determined by applying the elbow criterion to the cluster validity index (the ratio between within‐cluster to between‐cluster distance) for values of k = 2–20. This work has also demonstrated that this methodology reliably identifies the optimal number of clusters that can be reproduced using bootstrap and split‐half resampling validation techniques of empirical and simulated data [Allen et al., 2014]. K‐means clustering using values 2–20 was conducted on the concatenated correlation matrix consisting of all subject windows, and produced an optimal value of k = 5 determined by the elbow criterion.

K‐means clustering was then applied to the concatenated correlation matrix using a value of k = 5 to produce five clusters representing five dynamic insula states present across the course of the resting state scan. The k‐means algorithm used the L1 distance function (Manhattan or “cityblock” distance) based off of previous d‐FNC work [Allen et al., 2014]. The final five dynamic insula connectivity states presented were created by calculating individual subject median correlation coefficients for each state, and then averaging all individual subject medians together to create the correlations presented in each dynamic insula state.

Quantification of Within‐State Functional Connections of Insula Subdivisions

For each insula connectivity state, one‐sample t‐tests on the connection strength between each subdivision with another IC were conducted to identify significant positive connections between each subdivision and IC (Bonferroni corrected at P < 9.8039 × 10 ^{− 4}).

Additionally, the k‐means clustering algorithm produces clusters (or brain states) that are statistically separated as a whole, but it does not statistically separate differences across insula subdivisions within each cluster. In order to determine differences in insula subdivision strength with individual ICs (e.g., if the dAI had a stronger connection with the frontal pole than did the MI, PI, or VI), t‐tests were applied to individual subject medians between each insula subdivision for each IC. In order to keep the number of t‐tests and Bonferroni correction values consistent across clusters and to match the s‐FNC analysis, each cluster was subjected to 300 t‐tests (6 comparisons as in the static analysis across 50 brain areas) using a Bonferroni corrected alpha value of P < 1.67 × 10⁻⁴. This resulted in n number of significantly different connections between insula subdivisions within each cluster or state. A chi‐squared test was then applied across all clusters to determine if the amount of within‐cluster differences varied across clusters. Follow up chi‐square tests were then conducted to determine if any individual cluster had significantly greater or fewer differences in insula subdivision connections than another cluster (Bonferroni corrected for 10 chi‐square tests; P < 0.005.)

Comparison of Static and Dynamic Insula Connectivity Profiles

In order to examine how the static profiles of the insula subdivisions compared with their dynamic counterparts, two correlations were calculated. The first examined the relationships of the overall insula profile (all four subdivisions) in the static analysis with the overall insula profile in each dynamic insula state. The second examined the relationship of each individual subdivision's static profile with its counterpart in each dynamic state.

RESULTS

Static Functional Network Connectivity (s‐FNC)

The results of the high model‐order ICA are depicted in Figure 2 (bottom) showing the 52 non‐artifactual ICs. The functional parcellation of cortical and subcortical brain areas is similar to previous high model‐order ICA results [Kiviniemi et al., 2009] and other high model‐order ICA results used in d‐FNC analyses [Allen et al., 2014; Damaraju et al., 2014; Rashid et al., 2014]. The results of the s‐FNC correlation analysis are shown in the correlation matrix presented in Figure 2 (top) in order to demonstrate that the functional parcellation of the current study produced correlations between major brain systems similar to those previously reported. The resulting correlation values were ordered in a correlation matrix according to previous work grouping ICs into brain systems to facilitate functional interpretation. Grouping categories included subcortical, temporal, sensorimotor, visual, frontal, CEN, SN, DMN, and cerebellar systems [Allen et al., 2014].

Figure 2.

Top: Static functional connectivity correlation matrix for IC pairs. Bottom: ICs grouped according to brain systems. IC labels for the correlation matrix denote bilateral activation unless specified by L (left) or R (right). STG, superior temporal gyrus; MTG, medial temporal gyrus; IFG, inferior frontal gyrus; CG, central gyrus; SMA, supplementary motor area; OCC, occipital; CC, calcarine cortex; FG, fusiform gyrus; ACC, anterior cingulate; OBF, orbitofrontal cortex; DLPFC, dorsal lateral pre‐frontal cortex; AG, angular gyrus; MPFC, medial pre‐frontal cortex; DMN, default mode network. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

Overall, the s‐FNC correlation matrix shows strong positive correlations for ICs within sensorimotor, visual, and DMN brain systems, replicating previous work [Allen et al., 2014; Damaraju et al., 2014; Rashid et al., 2014]. The DMN ICs show strong negative correlations with sensorimotor, visual, and SN brain systems. Finally, subcortical ICs show moderate positive correlations with cerebellar ICs and moderate negative correlations with visual ICs.

The main results of interest are the s‐FNC correlations for each insula subdivision. Figure 3 shows the significant positive connections found in the s‐FNC analysis that were identified from one‐sample t‐tests. Importantly, positive connections between insula subdivisions and other ICs showed that the dAI had connections with frontal areas, the middle insula (MI) and PI with sensorimotor areas, and the ventral insula (VI) had connections with affective subcortical areas including the accumbens and hippocampus/amygdala ICs. These connections are consistent with an emotion‐cognition‐interoception framework of insula subdivision function.

Figure 3.

Plots of significant positive connections between insula subdivisions and other ICs. The maximum intensity values from each IC were used to designate the surface plot areas. The connections of the static analysis and State 3 from the dynamic analysis are in accord with the cognition‐emotion‐tripartite framework of insula function; the dorsal anterior insula (dAI) has connections with frontal areas, the middle and posterior (MI and PI) insula subdivisions have connections with sensorimotor areas, and the ventral insula (VI) has connections with the nucleus accumbens and hippocampus/amygdala ICs. State 1 is generally represented by similar connections across subdivisions with subcortical, frontal, sensorimotor, salience, and visual ICs. State 4 is similar to State 1 but with less frontal and sensorimotor connections and virtually no visual connections. States 2 and 5 show that the MI and PI have similar connections while the VI has less sensorimotor connections and the dAI diverges from other subdivisions to have connections with frontal (State 5) and default mode (States 2 and 5) ICs. Percent values refer to the frequency of state occurrence while n refers to the number of subjects that enter into that state. CEN, central executive network; DMN, default mode network; CB cerebellum. Figure created using BrainNet Viewer (https://www.nitrc.org/projects/bnv/). [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

Figure 4 shows a polar plot of s‐FNC correlations for the dAI, VI, MI, and PI. Differences in connection strengths between each subdivision with other ICs are identified by a “*” placed along the radiating axes that represents a significant difference between insula subdivisions for functional connectivity with that particular brain area. The dAI shows stronger connections with frontal brain areas, with significantly stronger connections than all other insula subdivisions with the frontal pole, while the VI shows significantly stronger connections with the accumbens and hippocampus/amygdala components (Supporting Information Figure 1). The PI shows the strongest connections with sensorimotor brain areas, with significantly stronger connections than all other insula subdivisions with the left, right, and medial pre/post central gyrus (Supporting Information Figure 2). Both the PI and MI show significantly stronger connections than the dAI and VI with the precentral gyrus, and pre/post central gyrus ICs. Finally, the PI and MI show the strongest connections with each other and with the mid‐cingulate cortex (mCC; Supporting Information Figure 3) IC. Thus, in addition to each subdivision showing significant positive connections consistent with a cognition‐emotion‐interoception framework, each subdivision's respective connections with specific brain regions was significantly greater than those of other insula subdivisions.

Figure 4.

Polar plot showing functional connections of insula subdivisions in the static functional connectivity analysis (s‐FNC). Labels on the outer edges of the polar plots are grouped in the same way as Figure 1, with different brain systems represented by different color fonts. The “*” placed along the radiating axes represents a significant difference among insula subdivisions for that specific brain area. Significant positive correlations can be seen between the dorsal anterior insula (dAI) and frontal ICs, the VI with the hippocampus/amygdala and the nucleus accumbens, and the posterior/middle insula with sensorimotor ICs (Figure 3). Polar plot labels use the same abbreviations as in Figure 2. The scale of the polar plot (−0.7 to 0.7) represents the size of the fisher‐z transformed correlations. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

Insula Dynamic Functional Connectivity (d‐FNC)

Polar plots representing the five dynamic states for each insula subdivision are depicted in Figure 5. It is important to note that not all subjects necessarily enter into all five states. This is because clustering was done on the concatenated subject matrix that assigned each sliding window to one of the five dynamic states regardless of subject assignment. When the dynamic states for each subject were back‐reconstructed by calculating individual subject medians according to window state‐assignment, this allowed for the possibility that not all subjects entered into all states.

Figure 5.

Polar plots representing functional connections of insula subdivisions for the whole insula d‐FNC analysis. Labels on the outer edges of the polar plots are grouped in the same way as Figure 1, with different brain systems represented by different color fonts. The “*” placed along the radiating axes represents a significant difference among insula subdivisions for that specific brain area. Notably, State 1 shows the most convergence among insula subdivision connections (less significant differences among subdivision connections) while State 5 shows the most divergence among insula subdivision connections (more significant differences among insula subdivision connections). Dorsal anterior insula = dAI. Polar plot labels use the same abbreviations as Figure 2. The scale of the polar plot (−0.5 to 0.5) represents the size of the fisher z transformed correlations. Percent values refer to the frequency of state occurrence while n refers to the number of subjects that enter into that state. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

State 3 represented the insula state occurring the most frequently (38%; n = 31 subjects entered into that particular state), and was similar to the s‐FNC findings in that unique functional profiles for each insula subdivision could be observed that mirrored those from the s‐FNC analysis, but were much smaller in magnitude. States 1 (24%; n = 31), 4 (13%; n = 26), and 5 (20%; n =30), represented moderately occurring insula states. Finally, State 2 (5%; n = 9) represented the insula state that occurred the least frequently, with nine subjects presenting with windows contributing to this brain state.

Functional Connections of Insula Subdivisions Within Each State

In State 1, all four insula subdivisions had connections with sensorimotor, temporal, visual, CEN, and SN ICs. Only the PI did not exhibit connections with frontal areas, while only the MI had connections with the cerebellum. This suggests that in State 1, insula subdivisions work together to process and integrate sensorimotor and visual information with the frontal cortex.

In State 2, only the dAI had connections with subcortical ICs and a node of the DMN while the MI, PI, and VI had connections with temporal, sensorimotor, and SN ICs. This suggests that in this state, the MI, PI, and VI work together to coordinate sensorimotor and temporal information while the dAI works independently coordinating subcortical and DMN communication.

State 3 was similar to the static analysis, with the dAI showing significant connections to frontal ICs, the MI and PI showing significant connections to sensorimotor ICs, and the VI showing significant connections to the accumbens and hippocampus/amygdala ICs. The functional connectivity profile observed in State 3 thus mirrored the cognition‐emotion‐interoception divisions found in the s‐FNC analysis.

In State 4, all insula subdivisions had connections to sensorimotor and SN ICs. Only the PI did not exhibit a subcortical connection, only the dAI did not exhibit a temporal connection, and only the PI did not exhibit frontal and CEN connections. Only the MI had a visual connection while only the dAI and MI had connections with a cerebellum IC. Thus, State 4 was similar to State 1 in the sense that insula subdivisions may be working together to coordinate sensory information with the frontal cortex, but without the integration of visual information.

In State 5, all insula subdivisions had connections with subcortical ICs while the MI, PI, and VI had connections with temporal, visual, SN, and CEN ICs. Only the dAI had connections with frontal and the DMN ICs and was the only subdivision without connections to SN ICs. Only the dAI and PI had cerebellum connections. Thus, State 5 was similar to State 2 in the sense that the MI, PI, and VI appear to be working together to coordinate information from the temporal and sensory components while the dAI works independently to coordinate information with frontal and DMN components.

Taking into account connections across all dynamic states, only the dAI had at least one connection with all brain systems and had transient connections with the DMN (Supporting Information Fig. 4). This suggests that the dAI is the most functionally flexible insula subdivision, exhibiting transient connections with all other brain systems. Additionally, only the PI did not exhibit connections with frontal ICs, which suggests that the MI and VI play a larger role than the PI in integrating sensory information with frontal components.

Within‐State Differences in Functional Connectivity of Insula Subdivisions

The testing of within‐cluster differences was meant to identify differences in connection strengths between subdivisions for each IC in order to determine if subdivisions had similar or different connections with the rest of the brain. For example, less differences between connection strengths across ICs would indicate greater overlap in functional connections (e.g., convergence of insular subdivision connectivity) while greater differences in connection strengths across ICs would indicate more unique functional connections (eg. divergence of insular subdivision connectivity). State 1 had 41 significant differences across insula subdivisions, State 2 had 33, State 3 had 77, State 4 had 61, and State 5 had 113 significant differences. Thus, states with lower amounts of significant differences were states in which insula subdivisions exhibited more overlapping connections to the various ICs (convergence across subdivision connections), whereas states with greater amounts of differences were states in which insula subdivisions exhibited more unique connectivity profiles with the various ICs (divergence across connections). Notably, although State 2 appears to exhibit large divergence across insula subdivisions, the small number of subjects that entered into State 2 resulted in extremely low power in detecting differences among insula subdivisions.

A chi‐squared test conducted on the amount of significant connections between states was significant (χ²= 62.52, df = 4, P < .0001). Follow up chi‐squared tests between individual states showed that State 5 had significantly more divergence among insula subdivisions than States 1, 2, and 4 (P′s <= 0.0001), State 3 had significantly more divergence among insula subdivisions than States 1 and 2 (P′s < 0.002) (Supporting Information Fig. 5). Finally, State 4 had marginally more divergence than State 2 (P = 0.0053) with no other differences between states (P′s > 0.01). This suggests that functional connectivity of insula subdivisions can demonstrate varying levels of divergence and convergence.

Convergence occurs during states when subdivisions have similar functional connections, and may indicate cooperation across subdivisions (e.g., State 1 had the least amount of differences suggesting insula subdivisions had similar connections to other ICs), while divergence signifies that subdivisions can exhibit unique functional connections that may indicate independence in specific contexts (e.g., State 5 had significantly more differences than State 1 suggesting that insula subdivision connections to other ICs were more varied in state 5 than state 1). These differences are represented visually in the polar plots in Figure 3 where State 1 shows a large amount of overlap for connections between subdivisions and other ICs, while State 5 shows a low amount of overlap for connections between subdivisions and other ICs.

Comparison of Static and Dynamic Insula Connectivity Profiles

The overall correlation comparing all insula subdivisions in the static analysis to each state showed that State 3 was the most similar to the static insula profile (r = 0.932), followed by State 1 (r = 0.90), State 5 (r = 0.84), State 4 (r = 0.829), and State 2 (r = 0.661) (Fig. 6). This is in accord with the s‐FNC results showing that the significant positive connections of State 3 mirrored the cognition‐emotion‐interoception division of the s‐FNC analysis, confirming that State 3 was the dynamic state most similar to the static insula profile.

Figure 6.

Correlations between static and dynamic insula profiles. The top matrix shows the overall correlations calculated using all insula states and show that State 3 has the largest correlation with the static analysis. The bottom matrices show that the dAI has the lowest correlations between the static and dynamic analyses, and also the lowest correlations among dynamic states. This shows that the dAI has more variable connections than the other insula subdivisions. S1, dynamic state 1; s2, dynamic state 2; s3, dynamic state 3, etc. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

The correlations comparing static and dynamic functional profiles showed that the PI generally had the most similarity between the s‐FNC and d‐FNC analysis as well as across d‐FNC states, while the dAI had the least similarity between the s‐FNC and d‐FNC analysis as well as across d‐FNC states. This suggests that the PI stays relatively stable with its connections across the rsfMRI scan, while the dAI shows the most variation across the rsfMRI scan. This lends further support to the role of the dAI as the most flexible insula subdivision with regards to the variation in its possible functional connections with other brain regions over time.

Consideration of Motion Artifacts

Recent investigations have identified the possible influence of motion on functional connectivity results in rsfMRI investigations [Power et al., 2012]. To explore the effects of participant motion on the current findinds, mean FD values were calculated for all TRs included in any given window related to a specific insula state. Additionally, the proportion of high motion windows (FD > 0.2 mm) was also calculated for each state. There were no systematic patterns linking high FD to any dynamic state (Supporting Information Fig. 6). Although state 2 had the highest mean FD, it was by a small amount (0.02 mm). However, state 2 also had the lowest proportion of high FD windows. Finally, state 4 had the lowest mean FD and lowest proportion of windows with high FD.

DISCUSSION

Distinct cortical areas have unique patterns of functional connections, comprising “connectional fingerprints” [Passingham, et al., 2002]. The neural context hypothesis posits that the functional relevance of a brain area depends on the context within which that region is operating, and that a given region can participate in several behaviors through variations in its interactions with other areas [McIntosh, 2004]. These network approaches to brain function have recently dominated the cognitive neuroscience landscape, and are continuing to provide more nuanced views of structure‐function relationships in the human brain [Pessoa, 2014].

We hypothesized that functional flexibility of the insular cortex may in part arise from the ability of subdivisions of the insula to transiently couple with other brain regions. The current s‐FNC analysis replicates the cognition‐emotion‐interoception functional framework found in previous studies [Chang et al., 2012; Deen et al., 2011; Uddin et al., 2014] showing that the dAI shares functional connections with frontal brain areas, the PI shares functional connections with sensorimotor based areas, and the VI shares functional connections with the limbic system. The d‐FNC analyses demonstrate several novel findings showing that the most frequently occurring dynamic state (State 3) mirrors the profile of the static analysis, while States 1 and 4 represent states where insula subdivisions exhibit similar patterns of functional connections to subcortical, sensorimotor, SN, frontal, and visual ICs. States 2 and 5 showed that the MI and PI have similar functional connections while the VI exhibits fewer connections with sensorimotor ICs, and the dAI diverges from the other subdivisions by having functional connections to frontal and DMN ICs. Accordingly, State 5 had significantly more differences between subdivision connection strengths than States 1 and 4, while State 3 had significantly more differences than State 1. This demonstrates that the insula can occupy states outside of the traditional cognition‐emotion‐interoception framework, and there can be differing amounts of functional overlap of insula subdivisions across dynamic states. Finally, the dAI was the most functionally flexible insula subdivision, showing connections to all brain systems in addition to varying the most between the static and dynamic analyses, as well as between dynamic states. These findings highlight the utility of a d‐FNC analysis for revealing the mutability of functional connections of brain regions across time.

Insula Subdivisions and s‐FNC

The connections of the dAI with the frontal pole IC corroborates research proposing a role for the dAI in cognitive control [Uddin, 2015]. The connections of the VI with limbic and reward circuitry replicates research linking the VI to affective processing via the amygdala [Mutschler et al., 2009] and the nucleus accumbens [Reynolds and Zahm, 2005]. The connections between the PI and sensorimotor ICs replicates research showing the PI plays a role in sensory/interoceptive processing through studies exploring the processing of cutaneous sensations [Craig, 2002; Drzezga et al., 2001] and studies showing functional links between the PI and sensorimotor areas [Deen et al., 2011; Uddin et al., 2014]. The MI showed some overlap with the PI by having similar connections to ICs representing the pre‐ and post‐central gyri. However, it also demonstrated functional independence by showing significantly weaker connections than the PI for ICs representing the left, right, and medial pre/postcentral gyri. Additionally, the PI and MI demonstrated the strongest functional connections with each other compared with the other insula subdivisions. This was not surprising considering that the PI in previous research [Deen et al., 2011] comprised an area that included the PI and MI in the current study. Accordingly, the PI and MI showed the strongest connections with the mid‐cingulate cortex IC, replicating previous research showing that while the dAI has connections with the ACC, the PI has connections with the mCC [Cauda et al., 2011; Nanetti et al., 2009; Taylor et al., 2009]. Overall, the s‐FNC results demonstrate that the functional parcellation in the current study revealed functional connections of insula subdivisions that are consistent with previous research.

The current study also showed a functional separation in the initial ICA of the PI from the MI that was also present in a previous insula parcellation study [Kelly et al., 2012]. The current study shows that there can be both functional overlap and specialization by the PI and MI. In the s‐FNC analysis, although the PI and MI shared the largest functional connections among insula subdivisions and with the mCC IC, the PI had more connections to sensorimotor ICs than the MI. Thus, the current study also suggests that the greater posterior area of the insula can be divided into MI and PI insula subdivisions, with both areas presenting with unique and overlapping functional connections.

Insula Subdivisions and d‐FNC Analysis

The transient functional connections observed for each insula subdivision suggest that the different insula states represented different types of coordination across brain systems. For example, State 3 replicated the cognition‐emotion‐interoception division generally found in static insula analysis approaches. However, States 1 and 4 showed that insula subdivisions exhibited similar functional connections with subcortical, sensorimotor, visual, and SN ICs with only the PI lacking a frontal connection. State 4 showed similar, albeit weaker, connections with sensorimotor and SN ICs but with increased frontal connections, with only the MI showing a single connection to a visual component. Thus, State 1 may represent a strong sensory integration processing state involving subcortical areas with minor frontal integration, while State 4 may represent a sensory integration state with less subcortical and visual connections but more frontal connections.

States 2 and 5 showed similarities in that the MI and PI had similar connections with sensorimotor, visual, and SN ICs while the VI had weaker sensorimotor connections, and the dAI diverged from the other three subdivisions by exhibiting unique connections with the DMN (both states), subcortical (state 2), and frontal areas (state 3). These two states could represent instances where the MI and PI subdivisions maintain their role in sensory processing, but the dAI diverges to communicate with frontal and DMN areas.

Overall, there were generally three types of dynamic insula states characterized by their functional connectivity patterns. The first was represented by State 3 and mirrored the general cognition‐emotion‐interoception division found in previous studies. The second type was represented by States 1 and 4 that were sensory integration states. Finally, States 2 and 5 represented states where the dAI had unique connections with DMN and frontal components in a possible “network switching” state driven by the dAI. This suggests that the insula can branch out beyond its traditionally assumed cognition‐emotion‐interoception role found in s‐FNC approaches, and that such a tripartite‐division is only one of several possible frameworks uncovered in a d‐FNC analysis.

The within‐cluster differences in overall connection strength between insula subdivisions mirrored that of the significant positive connections showing that it was not only the positive connections that were similar within a state, but it was in fact the general functional profile of the subdivisions that were similar. States 1 and 4, generally representing sensory and motor processing, showed no significant differences in overall connections while also showing significantly fewer differences in overall connections than States 3 and 5. This suggests that insula subdivisions tend to have overlapping profiles in states that integrate and/or evaluate external sensory and motor information. State 3, representing the cognition‐emotion‐interoception division, had significantly more connection differences than State 1 while State 5, representing independence of the dAI, had significantly more connection differences than State 4. This suggests that insula subdivisions tend to have unique profiles in states contributing to the cognition‐emotion‐interoception division and independence of the dAI.

These findings are relevant to previous research exploring the unique and overlapping functional profiles of insula subdivisions. It has been demonstrated that when exploring connections of insula subdivisions in isolation, each presents with a unique functional profile [Cauda et al., 2011; Deen et al., 2011]. However, considerable overlap of functional profiles of insula subdivisions has also been demonstrated [Uddin, et al., 2014]. The current study demonstrates that unique or overlapping functional profiles of insula subdivisions may also correspond to specific brain states.

The convergence and divergence of the functional connections of insula subdivisions suggests a mechanism by which the insula can be related to specific processes such as cognition, emotion, and interoception, while also functioning as a “network hub” that facilitates network switching and information coordination across networks. State 3 showing subdivision divergence of functional connections represents functional specialization related to cognition, emotion, and interoception processing while diverging States 2 and 5 represent communication and coordination between different neural networks. Finally, States 1 and 4 represent network communication regarding evaluation of external sensory information. In this way, connectivity divergence may facilitate functional specialization and network switching, while connectivity convergence may enable communication across sensory and motor networks.

Static Versus Dynamic Insula Profiles

An important aspect of the current study is the ability of a d‐FNC analysis to reveal information masked in traditional s‐FNC analysis. For example, d‐FNC analyses enable the quantification of the amount of time spent in a dynamic state that mirrors the profile of the static state, the differences in connections that are revealed across dynamic states, and the identification of functional connections not found in the static state.

As previously mentioned, the d‐FNC analysis showed that State 3 was most similar to the static analysis by replicating the cognition‐emotion‐interoception divisions of insula subdivisions. Accordingly, State 3 exhibited the strongest correlation with the static insula profile. Interestingly, State 3 was the most frequently occurring insula state, occurring 38% of the time. Thus, the dynamic state the insula spends the most time in mirrors the connections of the static insula profile. However, taking into account the other four insula states, the insula actually spends more time overall (62%) in a state different from the dominant cognition‐emotion‐interoception profile. This suggests that a large amount of information regarding the insula's connectivity profile is not captured in a strictly s‐FNC approach. The insula may actually switch between up to three types of general connectivity profiles, as observed in the d‐FNC analysis.

The dAI showed two important differences between the s‐FNC and d‐FNC analyses that support its hypothesized role as the most functionally flexible insula subdivision [Uddin et al., 2014]. First, the dAI had the weakest correlations between its static and dynamic connectivity profiles and between its own dynamic profiles across states. That is, the dAI showed the most variation in its connections across the dynamic and static analysis when contrasted against the other insula subdivisions. Second, the dAI was the only subdivision to show a connection with each brain system across the dynamic insula states. These findings provide evidence from rsfMRI data that further support task‐based fMRI findings that show the dAI has the most functionally diverse connections when compared with other insula subdivisions [Kurth et al., 2010; Uddin et al., 2014; Yeo et al., 2014].

Additionally, the dAI was the only insula subdivision to exhibit dynamic functional connections with the DMN (States 2 and 5), reinforcing its proposed function of network switching within the context of the SN [Menon and Uddin, 2010; Sridharan et al., 2008]. Typically, the DMN shows strong anti‐correlations with the SN and CEN, where activations in those networks are related to deactivation of the DMN and vice versa [Fox et al., 2005]. The dAI showed dynamic connections with the CEN in all states except in States 2 and 5. This provides evidence for infrequent connections between the networks, supporting the idea that although the SN and CEN are typically anti‐correlated with the DMN, dynamic connections with the dAI can serve to create a links with the CEN and DMN, perhaps to coordinate network switching.

Previously, it has been demonstrated that the dAI communicates with the DMN through effective connectivity—activity in the dAI precedes activity in the nodes of the DMN [Sridharan et al., 2008]. The current study shows another way that the dAI can communicate with the DMN—by dynamic direct connections rather than consistent effective connections. Both types of connections could serve to coordinate information between the two networks. For example, an effective connection could serve to prepare the DMN to receive important information directly from the dAI with a direct connection following closely after. Future research could explore the interplay between effective and dynamic connections to examine how they may work together to facilitate network information transfer.

The dAI was also the only subdivision to show consistent connections to the frontal brain system in both the static and dynamic analyses. This may explain why the dAI is implicated in more cognitive processing tasks than other insula subdivisions, as attention and inhibition processes are vital to most types of exogenous processes [Kurth et al., 2010; Uddin et al., 2014]. Interestingly, the MI and VI subdivisions exhibited connections to the frontal pole in State 4, demonstrating that insula subdivisions other than the dAI can exhibit dynamic connections with frontal brain areas. The PI was the only subdivision that did not exhibit any connections with frontal brain areas across both the s‐FNC and d‐FNC analysis. Thus, while the dAI has the most consistent connections to frontal areas in both the s‐FNC and d‐FNC analyses, the MI and VI can demonstrate dynamic connections to frontal areas, with the PI having no static or dynamic connections to frontal areas.

These results contrasting the s‐FNC and d‐FNC approaches highlight some of the unique findings that emerge from consideration of functional connectivity dynamics. These findings contribute to an emerging literature suggesting that it is important to take into account the nuanced changes that occur across the span of a rsfMRI scan, rather than assuming that functional connections are static and unchanging (Calhoun et al., 2014; Hutchison et al., 2013a).

Limitations

A limitation of the current study is that each insula subdivision component was bilateral as a result of the ICA decomposition. Previous research suggests that the right dAI plays the most prominent role in network switching and salience detection [Menon and Uddin, 2010; Sridharan et al., 2008]. Thus, it is difficult to contrast the current study with previous research showing hemispheric specialization of insula subdivisions. One way to address this in future work is to use a lateralized component‐based approach using spatially constrained ICA [Agcaoglu et al., 2015; Lin et al., 2010].

Additionally, little is known about how dynamic functional connections in the resting state are related to cognitive processing in task‐based fMRI experiments. Although previous research shows that networks decomposed from s‐FNC analyses of rsfMRI data represent networks instantiated during task‐based fMRI studies [Calhoun et al., 2008; Smith et al., 2009], there have been few studies exploring dynamic functional connections during task states [Chen et al., 2015; Gonzalez‐Castillo et al., 2012; Hutchison and Morton, 2015]. Future work should investigate how dynamic connections of insula subdivisions function within the context of task‐based fMRI investigations.

Another consideration is that the analysis of participant motion on dynamic states revealed minor differences between some states with regards to FD measures. It has been shown that motion artifacts can affect the strength of both short‐range and long‐range correlations [Power et al., 2012]. While the current work utilized a sample of high quality data with minimal motion artifact and stringent analytic motion control approaches, future work is needed to further understand the effects of these potential confounds.

Finally, the current d‐FNC approach is only one of several ways to map the dynamic functional connections of different brain areas. Other methods to mapping time‐varying functional connections between brain areas such as graph theoretical approaches [Braun et al., 2015], test‐statistics tracking time‐course fluctuations [Zalesky et al., 2014], co‐activation pattern identification [Chen et al., 2015], and using time‐frequency information [Chang and Glover, 2010; Yaesoubi et al., 2015] have all shown promise in identifying changes in functional connections that s‐FNC methods fail to capture. Future research should explore how these various measures can further elucidate the dynamic functional connections of insula subdivisions.

CONCLUSIONS

The results of the current study highlight how a d‐FNC analysis can better characterize functional connections of insula subdivisions. These results are important for creating more nuanced models of insula function as a major network hub both within and outside the context of the SN. As the insula has recently been identified as a possible biomarker in autism (Nomi and Uddin, 2015; Uddin, 2013) and a number of other clinical disorders such as schizophrenia, depression, and fronto‐temporal dementia (Uddin, 2015), mapping static and dynamic functional connections in the neurotypical population may help elucidate insula dysfunction in these clinical disorders.

Abbreviations

CEN

Central executive network

d‐FNC

Dynamic functional network connectivity

DMN

Default mode network

FD

Framewise displacement

IC

Independent component

MI

Middle insula

MNI

Montreal neurological Institute

PI

Posterior insula

s‐FNC

Static functional network connectivity

SN

Salience network

VI

Ventral insula

WASI

Wechsler Abbreviated Scale of Intelligence

Supporting information

Supporting Information Figure 1.

Supporting Information Figure 2.

Supporting Information Figure 3.

Supporting Information Figure 4.

Supporting Information Figure 5.

Supporting Information Figure 6.