Introduction
Behavioral variant frontotemporal dementia (bvFTD) is characterized by early brain atrophy in the frontotemporoinsular regions (Piguet, Hornberger, Mioshi, & Hodges, Reference Piguet, Hornberger, Mioshi and Hodges2011; Rascovsky et al., Reference Rascovsky, Hodges, Knopman, Mendez, Kramer, Neuhaus and Miller2011). These regions overlap with networks that are engaged in high-level processes, such as emotion recognition, social inference [e.g., theory of mind (ToM)], and executive functions (Ibanez & Manes, Reference Ibanez and Manes2012; Kennedy & Adolphs, Reference Kennedy and Adolphs2012; Stanley & Adolphs, Reference Stanley and Adolphs2013). Several reports have associated bvFTD-specific neurodegeneration with deficits in such social-executive domains (Possin et al., Reference Possin, Feigenbaum, Rankin, Smith, Boxer, Wood and Kramer2013; Torralva, Roca, Gleichgerrcht, Lopez, & Manes, Reference Torralva, Roca, Gleichgerrcht, Lopez and Manes2009); and previous studies have shown that the disruption of long-distance networks (Pievani, de Haan, Wu, Seeley, & Frisoni, Reference Pievani, de Haan, Wu, Seeley and Frisoni2011) provides information about behavioral symptoms (Farb et al., Reference Farb, Grady, Strother, Tang-Wai, Masellis, Black and Chow2013), executive functions (Agosta et al., Reference Agosta, Sala, Valsasina, Meani, Canu, Magnani and Filippi2013), and disease progression (Day et al., Reference Day, Farb, Tang-Wai, Masellis, Black, Freedman and Chow2013) in bvFTD. However, no single study has assessed the network centrality of the frontotemporoinsular network and its potential association with social-executive impairments.
Here, we seek to determine whether connectivity properties of the frontotemporoinsular network were associated with social-executive performance. Our analysis was based on the use of Graph Connectivity Metrics, which constitute a sensitive approach to study neurodegeneration (Pievani et al., Reference Pievani, de Haan, Wu, Seeley and Frisoni2011). We selected the Network Centrality (NC), a local metric which indicates the importance of a node in the global context of a network (Freeman, Reference Freeman1977). NC is a sensitive metric for bvFTD (Agosta et al., Reference Agosta, Sala, Valsasina, Meani, Canu, Magnani and Filippi2013) and, compared to other local metrics (e.g., clustering coefficient or degree), offers rich data about the relations between a network’s properties and observed symptoms and behaviors (Goch et al., Reference Goch, Stieltjes, Henze, Hering, Poustka, Meinzer and Maier-Hein2014; Zuo et al., Reference Zuo, Ehmke, Mennes, Imperati, Castellanos, Sporns and Milham2012). Global connectivity metrics (such as characteristic path length or average clustering coefficient) were not used because they do not provide local network information (Sporns, Reference Sporns2014). Thus, they are unsuitable to evaluate whether behavioral impairments are associated with deficits in specific nodes.
We assessed whether the frontotemporoinsular network’s centrality was altered in bvFTD, and examined whether this centrality measure was associated with social-executive performance. To this end, we used three control steps. First, we included frontoinsular stroke patients as a disease control group to test whether the NC properties of the frontotemporoinsular network in bvFTD were specific to neurodegeneration. Frontal stroke patients present important similarities with the clinical symptoms of bvFTD (Mesulam, Reference Mesulam1986), such as distractibility and personality changes. However, opposed to bvFTD, frontal lobe patients show high cognitive variability, ranging from almost totally preserved to impaired performance in multiple domains, including social cognition (Ibanez & Manes, Reference Ibanez and Manes2012; Mesulam, Reference Mesulam1986). Studies comparing patients with neurodegenerative diseases and stroke lesions provide valuable insights into such common patterns (Baez et al., Reference Baez, Couto, Torralva, Sposato, Huepe, Montanes and Ibanez2014; Lambon Ralph, Cipolotti, Manes, & Patterson, Reference Lambon Ralph, Cipolotti, Manes and Patterson2010). By comparing two groups of patients with similar clinical manifestations but different neuropathology, we aimed to evaluate whether NC results are specific to bvFTD degeneration or common to a broad range of neurological conditions. Second, to determine whether NC alterations were specific to the frontotemporinsular network, we also considered the integrity of selected anatomical regions from six well-characterized resting-state networks.
Finally, to challenge the distinctive association between the NC of the frontotemporoinsular network and social-executive performance, we also considered the association between NC and a general cognitive measure, which assess other domains than social-executive performance.
In sum, our aims were (i) to assess the NC of the frontotemporoinsular network in bvFTD, (ii) to evaluate whether NC is associated with social-executive profile, and (iii) to determine the contributions of this metric (together with behavioral deficits) in identifying bvFTD. We hypothesized that the NC of the frontotemporoinsular network would discriminate bvFTD patients from controls, and from stroke patients, and that it would be associated with social-executive performance.
Materials and Methods
Network Centrality analyses
Participants
We recruited 14 patients who fulfilled the revised criteria for probable bvFTD (Rascovsky et al., Reference Rascovsky, Hodges, Knopman, Mendez, Kramer, Neuhaus and Miller2011). These patients presented with prominent changes in personality and social behavior, which were verified by their caregivers. They underwent a clinical standard examination for accurate diagnosis at the Institute of Cognitive Neurology (INECO). This includes an extensive battery of neurological, neuropsychiatric, and neuropsychological assessments, and a MRI-SPECT. The diagnoses were made by a group of bvFTD experts (F.M. and T.T.). All patients showed frontal atrophy on MRI, and frontal hypoperfusion on SPECT, when available. They were all in the early/mild stages of the disease and did not fulfill criteria for specific psychiatric disorders. Patients who primarily presented with language deficits were excluded.
We also formed a control group of 12 age- and education-matched participants with no history of psychiatric or neurological disease (Table 1A). In addition, we recruited 10 frontoinsular stroke patients (Figure 2A) as a disease control group for complementary comparisons that were also assessed with the institutional standard examination. They were evaluated at least 6 month after suffering the stroke (time needed for the stability of the lesion extension and the clinical symptoms presentation).
Demographic and behavioral statistical results
Note. Mean (SD).
*Significant differences.
EF=executive functions; ER=emotion recognition; ToM=theory of mind; NC=Network Centrality; MMSE=Mini-Mental Status Examination; SCS=Social Cognition score; SEP=Social-Executive Performance.
All participants underwent a 10-min functional MRI (fMRI) resting protocol. They provided signed informed consent in accordance with the Declaration of Helsinki. The study’s protocol was approved by the institutional Ethics Committee.
FMRI preprocessing and connectivity analysis
FMRI acquisition
Functional images were acquired on a Philips Intera 1.5T with a conventional head coil. Thirty-three axial slices (5-mm thick) were acquired parallel to the plane connecting the anterior and posterior commissures and covering the whole brain (repetition time=2777 ms, echo time=50 ms, flip angle=90, image matrix=64×64 mm). The fMRI acquisition lasted 10 min and we obtained 209 functional brain images for each subject. The participants were instructed to think about their daily routines (e.g., the activities performed that day since waking or what they were going to do for the rest of the day), to keep their eyes closed and to avoid moving and falling asleep (Sedeño et al., Reference Sedeño, Couto, Melloni, Canales-Johnson, Yoris, Baez and Ibanez2014).
fMRI preprocessing
Functional data were preprocessed using statistical parametric mapping software (SPM8; http://fil.ion.ucl.ac.uk/spm). Echo-planar imaging (EPI) images were slice-time corrected, aligned to the mean volume of the session scanning, normalized (using the SPM8 default EPI template) and smoothed (using an 8-mm full-width half-maximum Gaussian kernel), following the same procedures previously described by our group (Barttfeld et al., Reference Barttfeld, Wicker, Cukier, Navarta, Lew, Leiguarda and Sigman2012, Reference Barttfeld, Wicker, McAleer, Belin, Cojan, Graziano and Sigman2013; Sedeno et al., Reference Sedeño, Couto, Melloni, Canales-Johnson, Yoris, Baez and Ibanez2014) (Figure 1A–C). The final spatial resolution of the images was 2×2×2 mm.
Functional MRI preprocessing and graph connectivity metrics. Preprocessing. A,B: Images were slice-time corrected and aligned to the mean volume of the scanning session. C: Data were normalized to a SPM8 default echo-planar imaging template and then smoothed. D: A band-pass filter was applied to correct and extract low-frequency drifts. Next, the images were regressed out by motion parameters, cerebrospinal fluid (CSF), white matter (WM), and global brain signals. E: Mean time series were extracted by averaging BOLD voxel signals in each region of interest (ROI), and then wavelet analysis was applied to construct correlation matrices of slow frequencies (0.01 to 0.05 Hz). Graph Connectivity Metrics analysis. F: Network Centrality (NC) was calculated based on a series of undirected graphs, with different numbers of positive connections (ranging from 50 to 100% of the connections of correlation matrices). G: We analyzed the average NC of a frontotemporoinsular network (and the main areas of six resting-state networks, see Figure 3 and Supplementary Data 2 for details related to the anatomical atlas and brain areas included in these networks) of the different undirected graphs in the range of 50 to 100% of positive connections with a cluster-based permutation test (see the Statistical Analysis section). H: We conducted simple linear regression analyses to explore whether social cognition and executive performances were partially associated by the averaged NC results from the 90 to 100% of positive connections (in these, differences were more consistent across comparisons).
Motion parameters showed no movements greater than 3 mm or rotation movements higher than 3° of rotation (Supekar & Menon, Reference Supekar and Menon2012). We also compared the mean translational and mean rotational parameters among groups using a mixed repeated-measures analysis of variance (ANOVA) test, with a within-subject factor (the two motion parameters) and a between-subject factor (group). No parameter effects [F(1,33)=1.12; p=.29] or parameter x group interaction [F(2,33)=.63; p=.53] were observed, indicating no significant differences in motion parameters among groups. In addition, we did not find any significant correlation between motion parameters and the main results at the group level (Supplementary Data 1A).
To partially correct and remove low-frequency drifts from the MR scanner, we applied a band-pass filter between 0.078 and 0.35 Hz using the Resting-State fMRI Data Analysis Toolkit (REST, http://resting-fmri.sourceforge.net/). Finally, applying these software, we regressed out the following items: (i) the six motion parameters, (ii) the average signals acquired form spherical ROIS in the ventricular cerebrospinal fluid (CSF) and white matter (WM), and (iii) the signal averaged over the whole brain (global signal) (Van Dijk, Sabuncu, & Buckner, Reference Van Dijk, Sabuncu and Buckner2012). This last procedural step was performed to remove the potential variance introduced by spurious sources (Figure 1D).
Correlation matrices for wavelet connectivity analysis
Based on the Automated Anatomical Labeling (AAL)-Atlas (Tzourio-Mazoyer et al., Reference Tzourio-Mazoyer, Landeau, Papathanassiou, Crivello, Etard, Delcroix and Joliot2002), mean time courses were extracted by averaging the BOLD signal of all voxels contained in each of the 116 regions of interest (ROIs). Wavelet analysis was used to construct a 116-node functional connectivity network for each subject from these time series, based on slow frequency components (0.01 to 0.05 Hz) (Supekar, Menon, Rubin, Musen, & Greicius, Reference Supekar, Menon, Rubin, Musen and Greicius2008). We followed the same procedures described by Supekar et al. (Reference Supekar, Menon, Rubin, Musen and Greicius2008), which have been previously used and detailed in studies of our group (Barttfeld et al., Reference Barttfeld, Wicker, Cukier, Navarta, Lew, Leiguarda and Sigman2012, Reference Barttfeld, Wicker, McAleer, Belin, Cojan, Graziano and Sigman2013; Sedeno et al., Reference Sedeño, Couto, Melloni, Canales-Johnson, Yoris, Baez and Ibanez2014) (Figure 1E).
Graph theory analysis: Network Centrality (NC)
NC measures the number of shortest paths that pass through a node and links the other node pairs across the network (Freeman, Reference Freeman1977). It indicates the importance of a node for efficient communication and integration across a network (Freeman, Reference Freeman1977). Several studies have already used NC (also called “betweenness centrality”) to identify changed connections in disconnections syndromes (Agosta et al., Reference Agosta, Sala, Valsasina, Meani, Canu, Magnani and Filippi2013; Buckner et al., Reference Buckner, Sepulcre, Talukdar, Krienen, Liu, Hedden and Johnson2009; Goch et al., Reference Goch, Stieltjes, Henze, Hering, Poustka, Meinzer and Maier-Hein2014; Seo et al., Reference Seo, Lee, Lee, Park, Sohn, Lee and Woo2013). In our study, we calculated the average NC across regions within different networks to characterize the central role of each network in the overall system’s dynamics.
To calculate NC, we converted functional weighted correlation matrices into binary undirected ones. Because network metrics depend both on network structure and size, a group comparison of the groups should be performed on networks of equal size (de Haan et al., Reference de Haan, Pijnenburg, Strijers, van der Made, van der Flier, Scheltens and Stam2009). Thus, if the samples have metric results calculated on matrices of the same size of connections, the network differences might reflect differences in graph structure (de Haan et al., Reference de Haan, Pijnenburg, Strijers, van der Made, van der Flier, Scheltens and Stam2009). To achieve this goal, we used the number of links (ROIs that are positively correlated) in weighted matrices as a cutoff to create a series of undirected graphs with different proportions of positive connections (global network density) (de Haan et al., Reference de Haan, Pijnenburg, Strijers, van der Made, van der Flier, Scheltens and Stam2009; He, Chen, & Evans, Reference He, Chen and Evans2008; Tian, Wang, Yan, & He, Reference Tian, Wang, Yan and He2011; Yao et al., Reference Yao, Zhang, Lin, Zhou, Xu, Jiang and Alzheimer’s Disease Neuroimaging2010).
The BCT toolbox (Sporns & Zwi, Reference Sporns and Zwi2004) was used to calculate the averaged NC across nodes within the frontotemporoinsular network (bilateral as well as left and right sides). This network involves the main areas of early degeneration that are the frontal paralimbic network, which includes the anterior cingulate cortex (ACC), anterior insula, frontal pole, amygdala, and striatum (Ibanez & Manes, Reference Ibanez and Manes2012; Piguet et al., Reference Piguet, Hornberger, Mioshi and Hodges2011; Rascovsky et al., Reference Rascovsky, Hodges, Knopman, Mendez, Kramer, Neuhaus and Miller2011; Rosen et al., Reference Rosen, Gorno-Tempini, Goldman, Perry, Schuff, Weiner and Miller2002; Seeley et al., Reference Seeley, Crawford, Rascovsky, Kramer, Weiner, Miller and Gorno-Tempini2008). In addition, this early degeneration pattern have been associated with specific bvFTD social cognition impairments (Couto et al., Reference Couto, Manes, Montanes, Matallana, Reyes, Velasquez and Ibanez2013).
Then, we examined whether NC results in bvFTD were specific to its atrophy areas or represented a property of all long-range connections. To this end, we evaluated the averaged NC of the main anatomical regions from six resting-state networks (the default mode, the cingulo-opercular, the frontoparietal, the sensorimotor, the visual and the cerebellar networks). The anatomical regions corresponding to each network were selected from the AAL-Atlas according to previous reports (Beckmann, DeLuca, Devlin, & Smith, Reference Beckmann, DeLuca, Devlin and Smith2005; Damoiseaux et al., Reference Damoiseaux, Rombouts, Barkhof, Scheltens, Stam, Smith and Beckmann2006; Kalcher et al., Reference Kalcher, Huf, Boubela, Filzmoser, Pezawas, Biswal and Windischberger2012; Smith et al., Reference Smith, Fox, Miller, Glahn, Fox, Mackay and Beckmann2009; van den Heuvel, Mandl, & Hulshoff Pol, Reference van den Heuvel, Mandl and Hulshoff Pol2008) (Supplementary Data 3).
There are no established criteria to select relevant undirected graphs for examining metric results. Here, we explored the networks’ configuration in the range of 50 to 100% of positive connections to allow comparability with a previous graph theory study in bvFTD (Agosta et al., Reference Agosta, Sala, Valsasina, Meani, Canu, Magnani and Filippi2013), which revealed topological abnormalities in the patients’ more densely connected networks. Note that, by establishing the 50% of connections as the lower limit, we avoided the inclusion of networks with disconnected nodes (Agosta et al., Reference Agosta, Sala, Valsasina, Meani, Canu, Magnani and Filippi2013; Supekar et al., Reference Supekar, Menon, Rubin, Musen and Greicius2008). Finally, we have also corroborated that these networks presented a small-world organization (Supplementary Data 1B).
One bvFTD patient and one stroke patient were eliminated from the NC analysis because they presented metric values 2 SDs above the mean of their respective groups.
Behavioral Assessment
Participants
A sub-sample of the participants completed general cognitive, executive function, and social cognition tasks. This sub-sample encompassed 14 bvFTD patients (nine of whom carried out the emotion recognition task), four frontoinsular stroke patients, and 12 controls. The results thus obtained, alongside the NC results from the 90 to 100% of positive connections (where differences were more consistent across comparisons, Figure 2C), were used for simple linear regression and classification analysis.
A: Frontal and insular structures that were injured in stroke patients. The colormap indicates lesions overlapping across the group: red refers to areas affected by the lesion of only one subject, while white shows injured areas shared by three patients. B: Regions of interest included in the frontotemporoinsular network were based on Tzourio-Mazoyer’s (Reference Tzourio-Mazoyer, Landeau, Papathanassiou, Crivello, Etard, Delcroix and Joliot2002) Automated Anatomical Labeling (AAL)-Atlas (see Supplementary Data 3). C: Pink boxes indicate the clusters were the bvFTD patients presented decreased NC compared to controls. Light blue boxes indicate the clusters were bvFTD patients showed decreased NC compared to the frontoinsular stroke group. No significant differences were found between controls and the last sample in the centrality of the frontotemporoinsular network. D: Compared with controls and stroke patients, bvFTD patients showed impairments in executive functions (EF), Social Cognition Score (SCS), and Social-Executive Performance (SEP) measures. No differences were found between controls and stroke patients. E: The NC of the bilateral frontotemporoinsular network was associated with participants’ performance in executive functions, SCS, and SEP.
General cognitive state
The Mini-Mental State Examination (MMSE) (Butman, Allegri, Harris, & Drake, Reference Butman, Allegri, Harris and Drake2000) is a clinical screening instrument that evaluates the general cognitive state of subjects and is used in bvFTD (Chow, Hynan, & Lipton, Reference Chow, Hynan and Lipton2006; Rascovsky et al., Reference Rascovsky, Salmon, Lipton, Leverenz, DeCarli, Jagust and Galasko2005). It comprises questions that assess orientation, memory, attention, and language.
Executive functions evaluation
The INECO Frontal Screening (IFS) (Torralva et al., Reference Torralva, Roca, Gleichgerrcht, Lopez and Manes2009) is a sensitive battery to detect executive dysfunction in patients with dementia (Gleichgerrcht, Roca, Manes, & Torralva, Reference Gleichgerrcht, Roca, Manes and Torralva2011; Torralva et al., Reference Torralva, Roca, Gleichgerrcht, Lopez and Manes2009). It includes the following subtests: motor programming, conflicting instructions, motor inhibitory control, numerical working memory, verbal working memory, spatial working memory, abstraction capacity, and verbal inhibitory control.
Social cognition
Emotion recognition
The Awareness of Social Inference Test (TASIT) (McDonald, Flanagan, Rollins, & Kinch, Reference McDonald, Flanagan, Rollins and Kinch2003) involves videotaped vignettes of everyday social interactions, which have been proven useful for detecting subtle deficits in bvFTD patients (Kipps, Nestor, Acosta-Cabronero, Arnold, & Hodges, Reference Kipps, Nestor, Acosta-Cabronero, Arnold and Hodges2009). This task introduces contextual cues (e.g., prosody, facial movement, and gestures) and additional processing demands (e.g., adequate speed of information processing, selective attention, and social reasoning) that are not taxed when viewing static displays. We only considered part 1, termed the emotion evaluation test (EET), which assesses recognition of spontaneous emotional expression (fearful, surprised, sad, angry, and disgusted). We selected this because it is affected at the initial stages of bvFTD regardless of the degree of atrophy (Kumfor et al., Reference Kumfor, Irish, Leyton, Miller, Lah, Devenney and Piguet2014). In the EET, speaker demeanor combined with the social situation indicates the emotional meaning. It comprises a series of 20 short (15–60 s) videotaped vignettes of trained actors interacting in everyday situations. After viewing each scene, the participant is instructed to choose (from a forced-choice list) the emotion expressed by the focused actor.
Social inferences (Theory of Mind, ToM)
The Reading the Mind in the Eyes Test (RMET) assesses emotional inference aspects of ToM (Baron-Cohen, Jolliffe, Mortimore, & Robertson, Reference Baron-Cohen, Jolliffe, Mortimore and Robertson1997) and is a sensitive task used to evaluate bvFTD patients (Torralva et al., Reference Torralva, Roca, Gleichgerrcht, Lopez and Manes2009). It is a computerized and validated test that consists of 36 pictures of the eye region of a face. Given four words, the participants are asked to choose the best word that describes what the person in each photograph is thinking or feeling.
Global scores
Based on a similar strategy of previous studies that have found bvFTD patients to be impaired in emotion recognition and ToM (Kipps et al., Reference Kipps, Nestor, Acosta-Cabronero, Arnold and Hodges2009; Torralva et al., Reference Torralva, Roca, Gleichgerrcht, Lopez and Manes2009), we constructed a global Social Cognition Score (SCS) to evaluate the global performance of participants and also to analyze the association of this performance with the NC results. The SCS combines the percent of correct answers from the TASIT and RMET.
In addition, given that the interrelationship between social cognition and executive functions plays an important role in the clinical presentation and symptomatology of bvFTD (Eslinger, Moore, Anderson, & Grossman, Reference Eslinger, Moore, Anderson and Grossman2011; Possin et al., Reference Possin, Feigenbaum, Rankin, Smith, Boxer, Wood and Kramer2013), we derived another score that indexes the global Social-Executive Performance (SEP) of the participants and combines the IFS, TASIT, and RMET scores. We also tested whether the SEP was associated with NC results.
Statistical Analysis
Demographic information was compared among groups using ANOVA tests, and Pearson chi-square (χ2) was used for gender.
To reduce the impact of the multiple comparison problem on the analysis of NC, we have used a modified version of the cluster-based permutation test proposed by Maris et al. (Maris & Oostenveld, Reference Maris and Oostenveld2007). This analysis was implemented using the FieldTrip Toolbox (Oostenveld, Fries, Maris, & Schoffelen, Reference Oostenveld, Fries, Maris and Schoffelen2011) and it has been previously applied to analyze multiple thresholds in graph theory (Sanz-Arigita et al., Reference Sanz-Arigita, Schoonheim, Damoiseaux, Rombouts, Maris, Barkhof and Stam2010). In this, the statistical metric of the original data was computed with two-tailed independent samples t tests. Afterward, the t-values were combined into connected sets based on their adjacency, and cluster-level statistics were calculated by taking the sum of the t-values within each cluster. The data were later permutated by applying 5000 permutation draws to generate a histogram. Then, we used the Monte-Carlo estimation of the permutation p-value, which is the proportion of random partitions in which the observed test statistic is larger than the value drawn from the permutation distribution. If this p-value is smaller than the critical alpha-level of .05, then the data can be concluded to reveal significant differences. This method offers a straightforward solution to the multiple comparisons problem and does not depend on multiple comparisons correction or assumptions about the normal distribution of the data (Nichols & Holmes, Reference Nichols and Holmes2002).
Given age differences between groups (Table 1A), we decided to perform an analysis of covariance test adjusted for age for the analyses of NC (Supplementary Data 1D), as well as for the social-executive comparisons (regarding the last, we reported only those effects that remained significant after covarying).
Simple linear regression analyses were used to explore whether the behavioral tasks and the global scores were partially associated by NC in the frontotemporoinsular, and whether global scores were associated with the main areas of the resting-state networks.
A k-means-like (MacQueen, Reference MacQueen1967) analysis (a vector discretization method) was used to test whether NC and SEP discriminated the bvFTD patients from controls and from stroke patients. This involves computing a centroid for each group by averaging corresponding data. Centroids were calculated for two groups: one encompassed by bvFTD patients and the other composed by controls and stroke patients. Then, a predicted group for an individual would be given by the closest centroid. This conservative approach (it only considers averages, disregarding information about cluster shapes) is appropriate because, given our sample size, averages should be reasonably robust, but cluster shapes may not be so. Note that this observation concerns the predictive power of the aforementioned variables, not the properties of an optimal classifier, which could be the object of further research. Once again, because of sample size, a leave-one-out cross-validation approach proved reasonable to determine whether classifier performance would generalize well to new data.
To corroborate classification results, we applied a different and independent method: the nearest neighbors’ classification method (Altman, Reference Altman1992), selecting three neighbors as parameter for the analysis. In this, a data point is compared to its three closest neighbors and is assigned to the most common class among them (in our case we had two: bvFTD patients and the other two groups). From the outputs of this classification method, we calculated the sensitivity and specificity for bvFTD from this combination of the NC of the frontotemporoinsular network and the SEP.
Results
Network Centrality
Compared to controls and stroke patients, the bvFTD group exhibited significantly decreased NC in the bilateral and right side of the frontotemporoinsular network. Significant differences were also observed on the left side, but only relative to controls (Table 2A; Figure 2C). Differences among groups remained the same after adjusting for age (Supplementary Data 1D).
NC and regression analysis
Note. Mean (SD).
a Tendency differences.
*Significant differences.
EF=executive functions; ER=emotion recognition; ToM=theory of mind. NC=Network Centrality; MMSE=Mini-Mental Status Examination; SCS=Social Cognition score; SEP=Social-Executive Performance.
Five of the six resting state-networks used as control comparisons presented no group differences. The only exception was the cingulo-opercular network, which revealed significant decreased NC in stroke patients relative to controls (Figure 3; Supplementary Data 1C).
NC of the main anatomical areas from six resting-state networks. Brown boxes indicate the clusters were the frontoinsular stroke patients presented decreased Network Centrality (NC) compared to controls. Significant differences were found only in the cingulo-opercular (CON) between these two samples. No significant differences were observed in the main anatomical areas of the other resting-state networks among groups (see Supplementary Data 1C).
Finally, the effect sizes of all significant differences reported in NC were above 0.8, indicating large differences among groups.
Behavioral Assessment
Relative to controls and stroke patients, bvFTD patients obtained significantly lower scores in their general cognitive state, executive functions, emotion recognition and ToM. The same was true of global SCS and SEP scores (Table 1B; Figure 2D; Supplementary Data 1E).
NC Contribution to Behavioral Performance
NC in the bilateral frontotemporoinsular network was associated with SCS and SEP (Table 2B; Figure 2E). The right hemisphere nodes (but not the left ones) were also related to performance in both the SCS and SEP. With regard to behavioral tasks, the bilateral and right NC significantly contributed to emotion recognition, whereas the right side was also associated with executive functions performance. Left nodes of this network were marginally related to ToM accuracy (Table 2B; Supplementary Data 1F). These results were significant even when a stroke patient that presented extreme values (Figure 2E) was excluded from the analysis (Supplementary Data 1F).
To evaluate the association between social-executive impairments and specific frontotemporoinsular network hubs in bvFTD, we conducted additional regression analyses considering only this group. The main network was divided into frontal, temporal, and insular regions. We found that (i) increased NC in the left insular nodes was related to impairments in emotion recognition and SCS, and (ii) the right frontal nodes were marginally associated with ToM impairments (Supplementary Data 1F).
No associations were found between the NC of the frontotemporoinsular network and MMSE results. This was true when considering bilateral regions as well as left and right sides alone (Table 2B). Moreover, none of the six resting-state networks analyzed was associated with the subjects’ social-executive profiles (SCS and SEP) (Supplementary Data 1F).
Group Discrimination Based on NC and SEP
The k-means-like model had a 100% correct classification rate (24 of 24). Thus, it should generalize well to new data because of two factors: the parameter count was low and a leave-one-out cross-validation yielded 95% correct classification rate [23 of 24 models, although over-fitting limitations should be considered (Nestor, Reference Nestor2013)].
In addition, the nearest neighbors’ classification analysis yielded a high sensitivity (100%) and high specificity (100%) for discriminating bvFTD from controls.
To establish whether our classification model was biased by the inclusion of both the stroke and the control groups, we re-ran these discrimination analyses excluding the stroke patients. The results remained the same, that is, bvFTD and controls were successfully discriminated (Supplementary Data 1G)
Finally, as patients were assessed with sensitive behavioral tasks, we performed a logistic regression only with NC to evaluate the classification power of this individual variable (as in the classification methods, we considered controls and stroke patients as a single group). In this, NC of bilateral frontotemporoinsular network was found to be a remarkably good predictor (pseudo-R2=.40) of bvFTD, with a reduction of 1 point in NC being associated to a 1.40 increase in the odds of FTD (Supplementary Data 1H).
Discussion
To our knowledge, this report is the first to show abnormal NC in the frontotemporoinsular connectivity of bvFTD patients and its association with social-executive performance.
Frontotemporoinsular Centrality in bvFTD
First, as compared with controls and stroke patients, bvFTD patients showed reduced NC of the frontotemporoinsular network. This finding aligns with previous evidence of centrality alterations in frontoinsular hubs (Agosta et al., Reference Agosta, Sala, Valsasina, Meani, Canu, Magnani and Filippi2013) and confirms the sensitivity of Graph Connectivity Metrics for bvFTD (Pievani et al., Reference Pievani, de Haan, Wu, Seeley and Frisoni2011). Moreover, such centrality alterations were absent in stroke patients. Thus, both groups of patients presented a similar frontoinsular affectation but with a different impact in the network centrality. The neurodegenerative process in bvFTD shows a specific alteration of frontotempoinsular connections that is consistent with its pattern of atrophy. Brain lesions, on the other hand, present centrality deficits circumscribed to the specific injured areas (cingulo-opercular network, see below), probably triggered by hypoconnectivity of the affected regions (Garcia-Cordero et al., Reference Garcia-Cordero, Sedeño, Fraiman, Craiem, de la Fuente, Salamone and Ibanez2015).
This difference in the involvement of brain networks is associated with distinct groups’ social-executive profile. While bvFTD exhibits larger deficits in all the behavioral tasks, the stroke patients, according to the variability of their performance (Ibanez & Manes, Reference Ibanez and Manes2012; Mesulam, Reference Mesulam1986) have similar results than the controls. In this way, despite that some similar areas are compromised in both neurological diseases (as the insular regions), the particular pathogenic processes of each one generates different patterns of connectivity alterations. Thus, our results suggest that the centrality alterations of the frontotemporoinsular network are a distinctive connectivity hallmark of neurodegeneration in bvFTD.
Importantly, this NC decrease in bvFTD was specific to the frontotemporoinsular network. This is consistent with previous studies reporting connectivity abnormalities of the salience network (SN) in bvFTD, although none of them applied a graph theory approach (Day et al., Reference Day, Farb, Tang-Wai, Masellis, Black, Freedman and Chow2013; Filippi et al., Reference Filippi, Agosta, Scola, Canu, Magnani, Marcone and Falini2013; Whitwell et al., Reference Whitwell, Josephs, Avula, Tosakulwong, Weigand, Senjem and Jack2011; Zhou et al., Reference Zhou, Greicius, Gennatas, Growdon, Jang, Rabinovici and Seeley2010). The impaired areas included in these studies as part of the SN involved the insular cortex, the ACC, the right superior temporal pole, the dorsolateral frontal lobe, the hypothalamus, the amygdala, and the striatum. Some of these regions are part of the frontotemporoinsular network that was found altered in our bvFTD sample. Thus, our results support the potential biomarker status of these networks. Indeed, connectivity alterations in bvFTD revealed by NC analyses engaged a widespread frontotemporoinsular network that overlaps with its pattern of early atrophy. Thus, our study illustrates the benefits of using graph theory analyses to examine the neurological correlates of cognitive performance in clinical populations.
We also found a significant NC alteration of the cingulo-opercular network in stroke patients relative to controls. While ours seems to be the first connectivity report of patients with frontoinsular lesions using graph methods, such alteration was expected given that the cingulo-opercular network comprises regions which are mainly damaged in this stroke sample (namely, insula and ACC). Future research focused on stroke patients’ deficits could shed broader light on the sensitivity of this network. A promising avenue is the exploration of possible alterations in long-range coupling among networks due to post-lesion compensatory effects and readjustments of functional connections in remotes sites (Grefkes & Fink, Reference Grefkes and Fink2014; Sporns, Reference Sporns2014).
To summarize, the selective alteration of frontotemporoinsular NC, only present in bvFTD and restricted to this network, is consistent with (i) several volumetric studies that have described a fronotemporoinsular pattern of atrophy in this disease, and (ii) their association with specific social cognition impairments (Couto et al., Reference Couto, Manes, Montanes, Matallana, Reyes, Velasquez and Ibanez2013; Rosen et al., Reference Rosen, Gorno-Tempini, Goldman, Perry, Schuff, Weiner and Miller2002; Seeley et al., Reference Seeley, Crawford, Rascovsky, Kramer, Weiner, Miller and Gorno-Tempini2008).
Social-Executive Performance and Long-Distance Networks in bvFTD
The frontotemporoinsular NC was associated with the participants’ social-executive profiles. This supports the view that high-level cognitive domains, particularly social cognition and related executive functions, depend on distributed frontotemporoinsular regions (particularly in right-sided areas) (Ibanez & Manes, Reference Ibanez and Manes2012; Kennedy & Adolphs, Reference Kennedy and Adolphs2012; Stanley & Adolphs, Reference Stanley and Adolphs2013). The specific involvement of this network in social-executive performance is further underscored by the null association among these behavioral domains and resting-state networks. Additionally, the lack of associations between the frontotemporoinsular NC and the MMSE (which assesses basic-level cognitive processes, such as orientation, attention, and memory) supports the specific involvement of this network in high-level social-executive performance (Ibanez & Manes, Reference Ibanez and Manes2012). Thus, by showing that similar network activity contributed to performance in both executive functions and social cognition, our results also corroborate the relationship between such domains.
Several studies have demonstrated this link between executive functions and social cognition (Decety, Reference Decety2011; Singer, Reference Singer2006; Singer & Lamm, Reference Singer and Lamm2009). Working memory, selective attention, and inhibitory control (Decety, Reference Decety2011; Rankin, Kramer, & Miller, Reference Rankin, Kramer and Miller2005; Singer, Reference Singer2006; Singer & Lamm, Reference Singer and Lamm2009) are particularly associated with the cognitive aspects of ToM. Specifically, inferring the intentionality of others requires the inhibition of one’s own perspective and the simultaneous appraisal of contextual cues (Rankin et al., Reference Rankin, Kramer and Miller2005). Additionally, brain regions that are relevant for executive functions, such as the prefrontal dorsolateral cortex, ACC, premotor cortex, parietal inferior cortex, orbitofrontal cortex, partially overlap and interact with areas involved in socio-affective responses (e.g., the ACC cortex, insula, and amygdala) (Singer & Lamm, Reference Singer and Lamm2009). Thus, the intertwining of executive functions and social cognition is not unexpected in bvFTD patients given that both domains are usually affected (Possin et al., Reference Possin, Feigenbaum, Rankin, Smith, Boxer, Wood and Kramer2013). This is in the same vein that the association we found between the frontotemporoinsular NC and the performance in both executive functions and social cognition.
In addition, increased NC in the left insular and right frontal hubs in the bvFTD group was associated with the patients’ social cognition impairments. Disease-specific compensatory or abnormally increased activity of these regions may modify the network’s centrality and compromise social cognition processes. Although speculative, this interpretation aligns with the increased connectivity observed in the bvFTD patients in the left insular (Day et al., Reference Day, Farb, Tang-Wai, Masellis, Black, Freedman and Chow2013; Farb et al., Reference Farb, Grady, Strother, Tang-Wai, Masellis, Black and Chow2013) and right frontal (Rytty et al., Reference Rytty, Nikkinen, Paavola, Abou Elseoud, Moilanen, Visuri and Remes2013) hubs. Moreover, it clarifies the elusive association between bvFTD-specific atrophy and social cognition impairments. Thus, the present findings confirm executive functions and social cognition impairments in bvFTD (Possin et al., Reference Possin, Feigenbaum, Rankin, Smith, Boxer, Wood and Kramer2013; Torralva et al., Reference Torralva, Roca, Gleichgerrcht, Lopez and Manes2009) while showing that these deficits are associated with frontotemporoinsular NC.
Finally, both frontotemporoinsular NC and social-executive performance were able to distinguish bvFTD patients from the other two groups (with a high classification rate). In addition, we have shown that NC discriminates patients individually (Supplementary Data 1H). Although behavioral measures seem enough to classify patients in our sample, it must be considered that: (i) these measures were selected “a priori”, based on their sensitivity for bvFTD; (ii) social-executive performance is strongly associated with NC; and (iii) this centrality measure also has a high classification ratio on its own. These findings highlight the potential contributions of combining behavioral and connectivity measures in future studies with larger samples (Pievani et al., Reference Pievani, de Haan, Wu, Seeley and Frisoni2011).
Limitations and Further Assessment
Although our patient sample size was larger than those in other bvFTD connectivity reports (Day et al., Reference Day, Farb, Tang-Wai, Masellis, Black, Freedman and Chow2013; Garcia-Cordero et al., Reference Garcia-Cordero, Sedeño, Fraiman, Craiem, de la Fuente, Salamone and Ibanez2015), future studies should include even larger groups. While the sample of vascular patients was also small, we considered it only for complementary comparisons. Note, however, that smaller group sizes have been used in recent functional connectivity studies (Day et al., Reference Day, Farb, Tang-Wai, Masellis, Black, Freedman and Chow2013; Farb et al., Reference Farb, Grady, Strother, Tang-Wai, Masellis, Black and Chow2013; Sajjadi et al., Reference Sajjadi, Acosta-Cabronero, Patterson, Diaz-de-Grenu, Williams and Nestor2013).
In addition, bvFTD is not an anatomically homogeneous syndrome (Kril, Macdonald, Patel, Png, & Halliday, Reference Kril, Macdonald, Patel, Png and Halliday2005; Rascovsky et al., Reference Rascovsky, Hodges, Knopman, Mendez, Kramer, Neuhaus and Miller2011; Whitwell et al., Reference Whitwell, Przybelski, Weigand, Ivnik, Vemuri, Gunter and Josephs2009). We have overcome this issue by analyzing the principal atrophy areas reported, thus maintaining consistency across subjects. Further studies with larger sample sizes should consider: (i) perform functional connectivity analyses for bvFTD patient samples featuring distinct atrophy patterns, and (ii) other FTD subtypes to disentangle whether each subtype presents a particular pattern of connectivity deficits. Future research should likewise consider additional social-executive measures.
Another limitation is that we focused only on NC from binary matrices. While this measure may be implemented considering weighted graphs (Brandes, Reference Brandes2001), algorithms used to such end do not measure the same network properties (Opsahl, Agneessens, & Skvoretz, Reference Opsahl, Agneessens and Skvoretz2010) as the ones presently considered. While the algorithm for binary matrices highlights the number of connections between nodes, the other ones ascribe more importance to the ties’ weight (a node with few connections and with high weights would have greater NC than a node with more connections but with low weights). Currently, most graph theory studies assessing NC in dementia are based on the binary approach (Agosta et al., Reference Agosta, Sala, Valsasina, Meani, Canu, Magnani and Filippi2013; Baggio et al., Reference Baggio, Sala-Llonch, Segura, Marti, Valldeoriola, Compta and Junque2014; Brier et al., Reference Brier, Thomas, Fagan, Hassenstab, Holtzman, Benzinger and Ances2014; Li, Qin, Chen, & Li, Reference Li, Qin, Chen and Li2013; Liu et al., Reference Liu, Zhang, Yan, Bai, Dai, Wei and Tian2012; Xiang, Guo, Cao, Liang, & Chen, Reference Xiang, Guo, Cao, Liang and Chen2013). Future studies should analyze the impact of each method on network properties of bvFTD and other neurodegenerative diseases.
Finally, to corroborate the discrimination power of NC, it would be useful to compare this metric with node-segregation and network-integration metrics.
Conclusion
The combination of theoretical models (social cognition network approaches), clinical evidence (bvFTD brain abnormalities and specific impaired performance), and recent mathematical developments (network science) represents a promising approach to increase our understanding of the neural networks engaged in social-cognitive process affected by bvFTD.
Acknowledgments
This work was partially supported by the National Commission for Scientific and Technological Research of Chile/FONDECYT Regular (1130920 and 1140114), National Health and Medical Research Council of Australia Career Development Fellowship (APP1022684 to O.P.), Foncyt-PICT 2012-0412, Foncyt-PICT 2012-1309, National Council of Scientific and Technical Research and the Institute of Cognitive Neurology Foundation. The authors have no competing interests to declare.
Supplementary Material
To view supplementary material for this article, please visit http://dx.doi.org/10.1017/S1355617715000703.
