Resting-state network connectivity and metastability predict clinical symptoms in schizophrenia

Abstract

Background:

The functional architecture of resting-state networks (RSNs) is defined by their connectivity and metastability. Disrupted RSN connectivity has been amply demonstrated in schizophrenia while the role of metastability remains poorly defined. Here, we undertake a comprehensive characterisation of RSN organization in schizophrenia and test its contribution to the clinical profile of this disorder.

Methods:

We extracted RSNs representing the default mode (DMN), central executive (CEN), salience (SAL), language (LAN), sensorimotor (SMN), auditory (AN) and visual (VN) networks from resting-state functional magnetic resonance imaging data obtained from patients with schizophrenia (n = 85) and healthy individuals (n = 48). For each network, we computed its functional cohesiveness and integration and used the Kuramoto order parameter to compute metastability. We used stepwise multiple regression analyses to test these RSN features as predictors of symptom severity in patients.

Results:

RSN features respectively explained 14%, 17%, 12% and 5% of the variance in positive, negative, anxious/ depressive and agitation/disorganization symptoms. Lower functional integration between the DMN, CEN and SMN primarily contributed to positive symptoms. The functional properties of the SAL network were key predictors of all other symptom dimensions; specifically, lower cohesiveness of the SAL, lower integration of this network with the LAN and higher integration with the CEN respectively contributed to negative, anxious/depressive and disorganization symptoms. Increased SAL metastability was associated with negative symptoms.

Conclusions:

These results confirm the primacy of the SAL network for schizophrenia and demonstrate that abnormalities in RSN connectivity and metastability are significant predictors of schizophrenia-related psychopathology.

1. Introduction

Schizophrenia is a neuropsychiatric syndrome associated with disturbances in perception, emotion and cognition (APA, 2013). Multiple genetic, epigenetic, transcriptional and synaptic mechanisms have been implicated in the pathophysiology of schizophrenia (Lewis and Sweet, 2009; Schizophrenia Working Group of the Psychiatric Genomics Consortium, 2014; Glessner et al., 2010; Akbarian, 2014) all of which ultimately impact on the integrity of brain organization. Brain functional organization is characterised by the presence of resting-state networks (RSNs) (Doucet et al., 2011; Fox and Raichle, 2007; van den Heuvel and Hulshoff Pol, 2010; Power et al., 2011) that are defined by the temporal dependency of the spontaneous fluctuations in regional blood-oxygen-level dependent (BOLD) signals (Biswal et al., 1995; Cordes et al., 2000). Functionally, RSNs can be considered in terms of higher-order networks (e.g., default mode, salience) that are involved in sustaining diverse and complex mental operations over time and lower-order networks (e.g., auditory, visual) that rapidly respond to specialized external inputs (Power et al., 2011; Smith et al., 2009).

Dysconnectivity in brain networks is currently the best-supported model for schizophrenia (Frangou, 2014; Stephan et al., 2009; Friston, 1999; Bullmore et al., 1997). Significant disruption in functional connectivity has been demonstrated in multiple RSNs (Pettersson-Yeo et al., 2011; Fornito et al., 2012; Karbasforoushan and Woodward, 2012; Fitzsimmons et al., 2013; Kuhn and Gallinat, 2013; Kambeitz et al., 2015; Dong et al., 2017). The most consistent findings support a general pattern of reduced network cohesion coupled with aberrant integration (Cole et al., 2011; Garrity et al., 2007; Whitfield-Gabrieli et al., 2009; Woodward et al., 2011; Palaniyappan and Liddle, 2012; Venkataraman et al., 2012; Yu et al., 2012; Du et al., 2016; van de Ven et al., 2017; Alderson-Day et al., 2016; Northoff and Qin, 2011; Khadka et al., 2013; Shinn et al., 2013; Kaufmann et al., 2015). There is some evidence that dysconnectivity of the default mode (DMN), central executive (CEN) and salience (SAL) networks is associated with the severity of positive symptoms (Rotarska-Jagiela et al., 2010; Woodward et al., 2011; Khadka et al., 2013; Palaniyappan et al., 2013). However, there remains a significant knowledge gap regarding the association between the disrupted RSN connectivity and the clinical symptoms of schizophrenia.

More recently, the characterisation of the brain functional organization at rest has shifted towards measures that attempt to capture dynamic network properties based on the temporal patterns of the oscillatory activity of their constituent brain regions (Cabral et al., 2011; Deco et al., 2011). Particular attention is focused on network synchrony and metastability. This is because synchrony in the oscillatory activity of network regions is thought to underpin information exchange (Fries, 2005), while metastability represents the variability in the synchronization of network regions over time that is considered important for adaptive information processing (Kelso, 2012; Tognoli and Kelso, 2014). The biological mechanisms that support these dynamic network properties occur at multiple timescales which also include the slow frequencies measured with functional magnetic resonance imaging (fMRI). We have previously shown that the RSNs differ in metastability in a manner related to the repertoire of network state required for their functional roles (Lee and Frangou, 2017). Higher-order RSNs appear to have lower metastability aligned with their function in sustaining mental operations over time while lower-order RSNs show higher metastability, suggestive of greater capacity to respond quickly to external demands (Lee and Frangou, 2017). The role of metastability in schizophrenia remains unclear; this represents a further knowledge gap in the comprehensive characterisation of schizophrenia-related dysconnectivity despite its potential to provide novel mechanistic insights.

In the present study, we aimed to address the knowledge gaps outlined above using resting-state fMRI data obtained from patients with schizophrenia (n = 85) and health individuals (n = 48). Based on their involvement in schizophrenia, seven RSNs (default mode, central executive, salience, language, auditory, visual and sensorimotor) were selected for analysis and characterised in terms of their cohesiveness (within-network functional connectivity) and integration (between-network functional connectivity). Based on our previous work (Lee and Frangou, 2017), the synchrony and metastability of each RSN were computed using the Kuramoto order parameter (Shanahan, 2010; Cabral et al., 2011). First, we investigated schizophrenia-related abnormalities in the synchrony and metastability of the selected RSNs. Our main prediction was that schizophrenia-related abnormalities in these dynamic network measures will be primarily observed for the salience network which has emerged as central to the process through which symptoms may emerge (Palaniyappan et al., 2013; Menon, 2011). Further, we tested the explanatory power of the functional connectivity and dynamic features of the RSNs for the symptom dimensions of schizophrenia.

2. Materials and methods

2.1. Participants

The study sample comprised 93 patients with recent onset schizophrenia (illness duration <5 years) and 50 healthy individuals (Table 1). Patients fulfilled criteria for schizophrenia as defined by the fifth edition of the Diagnostic and Statistical Manual of Mental Disorders (DSM-5) (APA, 2013) and were recruited via clinician referrals from the psychiatric services of the Mount Sinai Health System, New York. Healthy individuals were recruited via advertisements in the local press. All participants were screened to exclude those with IQ < 70, medical or neurological disorders, history of significant head trauma, and lifetime history of DSM-5 substance use disorder. The diagnostic status of all participants was determined using the research version of the Structured Clinical interview for DSM-5 (First et al., 2015) supplemented by information from medical records in the case of patients. The presence and severity of psychopathology were assessed in all participants using the expanded 24-item Brief Psychotic Rating Scale (BPRS) (Lukoff et al., 1986) which allows decomposition of the clinical profile of psychosis into four dimensions comprising positive symptoms, negative symptoms, anxiety/depression and agitation/disorganization (Ventura et al., 2000; Kopelowicz et al., 2008). An estimate of IQ was obtained from all participants using the Wechsler Abbreviated Scale of Intelligence, 2nd Edition (WASI-II) (Wechsler, 2011). Medication type and dose was recorded in all patients and the daily antipsychotic dose was converted to chlorpromazine equivalents (CPZE) (Gardner et al., 2010). The study was approved by the Institutional Review Board of the Icahn School of Medicine at Mount Sinai (ISMMS). All participants provided written informed consent.

Table 1.

Sample Characteristics.

	Patients with schizophrenia (N = 85)	Healthy individuals (N = 48)
Age (years)	27.18 (7.31)	29.35 (8.28)
Sex (male: female)1	64:21	28:20
IQ2	93.85 (14.87)	115.87 (16.32)
Duration of illness (years)	5.72 (6.14)	n/a
BPRS positive symptoms1*	12.57 ± 6.17	4 (0)
BPRS negative symptoms1*	6.83 ± 3.68	3.02 (0.14)
BPRS anxiety/depression1*	9.06 ± 5.35	4.02 (0.14)
BPRS agitation/disorganization1*	9.57 ± 5.81	6.02 (0.14)
Any medication, n (%)	79 (92.94)	n/a
Antipsychotics, n (%)	75 (88.23)	n/a
Antipsychotic dose (CPZE)	277.24 (246.99)	n/a
Antidepressants, n (%)	23 (27.06)	n/a

Continuous variables are shown as mean (standard deviation); Categorical variables are shown as number of cases (n) and percentage (%); BPRS=Brief Psychiatric Raring Scale

^*In BPRS symptoms are coded as 1 (absent) to 7 (extremely severe); BPRS positive symptoms = sum of the scores for hallucination, unusual thought content, bizarre behavior subscales; BPRS negative symptoms = sum of the scores for blunted affect, emotional withdrawal, motor retardation subscales; BPRS Depression/Anxiety Scores = sum of the scores for anxiety, depression, suicidality, guilt subscales; BPRS Agitation/Disorganization Scores = sum of scores for motor hyperactivity, elevated mood, excitement, distractibility, grandiosity subscales. CPZE = chlorpromazine equivalents; n/a = not applicable

¹p = 0.01

²p < 0.001.

2.2. Neuroimaging data acquisition and quality assurance

Anatomical and resting-state fMRI (eyes open) data were acquired on a Siemens Skyra 3 T scanner (Erlangen, Germany) at the ISMMS (details in Supplementary Material). Data from the participants were preprocessed using established procedures implemented in Statistical Parametric Mapping software (SPM12; http://www.fil.ion.ucl.ac.uk/spm/software/spm12) and the Data Processing Assistant for Resting-State fMRI software (DPARSF) (Chao-Gan and Yu-Feng, 2010) available in the Data Processing and Analysis for Brain Imaging (DPABI) Toolbox (Yan et al., 2016) (details in Supplementary Material). Two healthy individuals and 8 patients with schizophrenia were excluded because of excessive head motion (linear shift >3 mm or rotation >1° across the run and on a frame-to-frame basis). After removal of these datasets, comparison between groups (patients, healthy volunteers) on multiple head motion parameters (i.e., mean and maximum scan-to-scan head motion and framewise displacement) did not yield significant difference (all p > 0.05; details in Supplementary Material).

2.3. Computation of RSN functional connectivity

In each individual resting-state fMRI dataset, we defined seven major RSNs comprising the default mode network (DMN), the central executive network (CEN), the salience network (SAL), the sensorimotor network (SMN), the visual network (VN), the auditory network (AN), and the language network (LAN). These RSNs were defined using validated network masks that are publicly available through the Functional Imaging of Neuropsychiatric Disorders (FIND) Lab, Stanford University (http://findlab.stanford.edu/functional_ROIs.html) (Shirer et al., 2012). Each of the 7 RSN masks comprises its corresponding functional regions-of-interest (ROIs) (Fig. 1A; Supplemental Table S1). We chose functionally defined RSN masks because they offer great cogency and reproducibility (Smith et al., 2009; Laird et al., 2011; Power et al., 2011). We have previously demonstrated the reproducibility of the FIND Lab masks using data from the Human Connectome Project (detailed in Supplemental Material and in Doucet et al., 2018). The median overlap between functionally defined RSN masks (Shirer et al., 2012) and data-driven RSNs (Doucet et al., 2018) was 91%. We assessed the cohesiveness and integration of each RSN by computing their within-and between-network functional connectivity. Within-network functional connectivity was estimated by averaging the fMRI time series over all the voxels within each network region and then calculating all possible pairwise Pearson’s correlation between the network regions. Between-network functional connectivity was estimated by computing pairwise Pearson’s correlation between the networks’ average time series. All correlation coefficients were transformed into Fisher’s z-score prior to further analyses. These computations resulted in 28 connectivity measures per subject which were used in further analyses. No correlation was found between any of these measures and framewise displacement (FD) even at p < 0.05, uncorrected (detailed in Supplementary Material).

Fig. 1. Temporal synchronization in resting-state networks.

(A) Spatial maps of the 7 resting-state networks, (B) Temporal fluctuations of the order parameter amplitude R(t) of the functional magnetic resonance imaging (fMRI) signal for each RSN. Color-coded thick lines and shaded bands indicate, respectively, the mean and standard error of mean (SEM) of the order parameter amplitudes across subjects for healthy individuals (blue) and patients with schizophrenia (red). Network synchrony and metastability were computed as the mean and standard deviation of the order parameter R(t) over time, respectively.

2.4. Computation of network synchrony and metastability

The voxel-wise BOLD data for each RSN were filtered to extract frequency-band fMRI signals using the maximal overlap discrete wavelet transform (MODWT) (Percival and Walden, 2000). Because the repetition time of the fMRI data acquisition was 1 s, the frequency bands extracted were: scale 1 = 0.25–0.5 Hz; scale 2 = 0.13–0.25 Hz; scale 3 = 0.06–0.13 Hz; scale 4 = 0.03–0.06 Hz; scale 5 = 0.015–0.03 Hz. We used a wavelet filter of Daubechies Least Asymmetric with a wavelet length of 8 (Lee and Frangou, 2017). We focused on scale 4 wavelet coefficients, because the most relevant information for resting-state fMRI data is mainly contained within this scale (Achard et al., 2006; Biswal et al., 1995; Fornito et al., 2010; Glerean et al., 2012). We then extracted the phases of the fMRI time series for each RSN in each participant by applying the Hilbert transform to the wavelet-filtered fMRI BOLD signals to obtain the associated analytical signals. The analytic signal represents a narrowband signal, s(t), in the time domain as a rotating vector with an instantaneous phase, φ(t), and an instantaneous amplitude, A (t), i.e., s(t) = A(t) cos (φ(t)). The phase and the amplitude are given by the argument and the modulus, respectively, of the complex signal z(t), given by z(t) = s(t) + iH[s(t)], where i is the imaginary unit and H[s(t)] is the Hilbert transform of s(t) (Glerean et al., 2012; PonceAlvarez et al., 2015). Subsequently, the first and last 10 time steps were discarded to avoid border effect inherent to the Hilbert transform (Ponce-Alvarez et al., 2015). To evaluate the dynamic properties of each RSN, we computed the Kuramoto order parameter R(t), defined as

$$R(t)=\left|\frac{1}{N}{\sum }_{n=1}^N{e}^{\text{i}{\phi }_n(t)}\right|$$

where N is the total number of voxels within each RSN and φ_n(t) is the instantaneous phase of each fMRI BOLD signal at voxel n of each RSN. This parameter measures the level of synchronization between brain regions across time (Fig. 1B). For each network, synchrony (mean synchronization) and metastability (the variation in synchronization over time) were defined as the mean and the standard deviation of the Kuramoto order parameter over time, respectively (Cabral et al., 2011; Lee et al., 2017; Shanahan, 2010; Váša et al., 2015). These computations resulted in 14 measures per subject that were used in further analyses.

2.5. Statistical analysis

Group differences in demographic characteristics and IQ were examined using t- and χ² tests as appropriate. We detected an effect of group for sex and IQ and we therefore conducted collinearity diagnostics using regression models. These variables were not collinear as their variance inflation factor (VIF) ranged between 1 and 1.4 which is below the traditional threshold of 2 for collinearity. We performed three multivariate analyses of covariance (MANOVAs) to examine the effect of diagnosis separately on measures of network dynamics (synchrony and metastability), measures of within-network functional connectivity and measures of between-network functional connectivity. Sex and IQ were entered as covariates in all analyses. For each MANOVA, when the overall model was significant at p < 0.05, we conducted post-hoc pairwise comparisons for which the threshold of statistical significance was adjusted using Bonferroni correction based on the number of variables included in each model. In addition, the effect size of case-control differences in neuroimaging measures was calculated using the Cohen’s d (Nakagawa and Cuthill, 2007). The predictive value of the neuroimaging variables for clinical symptoms was only tested in the patient group. The four BPRS symptom scores were not collinear (|r| < 0.4) and were used as dependent factors in separate stepwise regression analyses. This approach operates by successively adding and removing predictor variables based on their squared t-statistic to select the best grouping of predictor variables that accounts for most of the variance (R²) in the outcome. The assumptions, advantages and limitations of this approach are well understood (e.g. Derksen and Keselman, 1992; Burnham and Anderson, 2002; Cohen et al., 2013). Stepwise regression models allow a principled and data-driven selection of predictor variables, that is largely independent of sample size, and are best suited for testing prespecified hypotheses about the predictor and outcome variables. This is the case here as our aim was to test the predictive value of predefined connectivity features for symptom severity in schizophrenia. Stepwise regression analyses are prone to R² values inflation and are sensitive to the presence of collinearity amongst the predictor variables. To address these concerns, we undertook the following three steps: (a) Curve estimation and collinearity assessments were conducted to select the neuroimaging variables entered as independent predictors. Variables were considered co-linear if their pairwise correlation was |r| > 0.4. The synchrony and metastability of each network were the only highly correlated (r values ranged between 0.63 and 0.72) neuroimaging variables and therefore only metastability was entered in the regression analyses. (b) We confirmed that none of the regression results showed collinearity based on VIF <2 and tolerance >0.2. (c) We examined the reliability of the R² values of the regression models using a leave-one-out method that enables testing the reliability of these values while preserving statistical power.

3. Results

The descriptive statistics for all neuroimaging parameters are shown in Supplemental Table S2 and univariate correlations between neuroimaging measures and symptom dimensions are shown in Supplemental Figs. S1 and S2. None of the neuroimaging parameters showed significant correlations with medication dose expressed in daily CPZE (all |ρ| b 0.17, p > 0.07, uncorrected).

3.1. Effect of diagnosis on network dynamic and connectivity measures

The effect size of the case-control differences for within- and between-network functional connectivity, network metastability and synchrony are shown in Supplemental Table S2. We found a significant effect of diagnosis on the dynamic network measures (F_14,113 = 1.91; p = 0.03). There was no significant effect of sex (F_14,113 = 0.12; p = 0.35) or IQ (F_14,113 = 0.11; p = 0.37). No interactions were significant either (p > 0.05). Post-hoc pairwise comparisons showed that the SAL network had higher metastability (p = 0.042, Bonferroni corrected) and synchrony (p = 0.028, Bonferroni corrected) in patients than healthy participants. We also found a significant effect of diagnosis on the within-network functional connectivity (F_7,122 = 2.11; p = 0.04). There were no other significant effects (sex: F_7,122 = 1.35; p = 0.23; IQ: F_7,122 = 1.14; p = 0.34) or interactions (p > 0.05). Post-hoc pairwise comparisons showed that the SAL within-network connectivity was higher in patients than healthy participants (p = 0.024, uncorrected) although this finding was below the Bonferroni-adjusted threshold. The multivariate analysis of covariation for between-network functional connectivity did not reach statistical significance for diagnosis (F_19,109 = 1.53; p = 0.08), sex (F_19,109 = 1.50; p = 0.11) or IQ (F_19,109 = 1.08; p = 0.37). The effect of diagnosis remained below the conventional threshold for statistical significance (F_19,113 = 1.59; p = 0.06) even after removing the covariates.

3.2. Neuroimaging predictors of symptom dimensions in schizophrenia

3.2.1. Positive symptoms

The best model explained 14% of the variance in symptom severity (Table 2; Supplementary Fig. S3). Measures of functional connectivity between the DMN-SMN, DMN-CEN and SMN-AN made a statistically significant contribution to the model, such that for any unit increase in the connectivity between these networks, the model predicted lower BPRS positive symptom score (Table 2, Fig. 2).

Table 2.

Summary of Stepwise Regression Analyses for Neuroimaging Predictors of Symptom Severity in Patients with Schizophrenia.

	Positive symptoms
model	Variable	B (SE B)	β	Adjusted R2	F	P-value
1	DMN-SMN	−7.30 (3.33)	−0.23	0.04	4.80	0.03
2	DMN-SMN	−10.38 (3.47)	−0.33	0.10	5.54	0.006
	DMN-CEN	−6.79 (2.77)	−0.27
3	DMN-SMN	−11.29 (3.43)	−0.36	0.14	5.27	0.002
	DMN-CEN	−6.06 (2.74)	−0.24
	SMN-AN	−7.02 (3.40)	−0.21
	Negative symptoms
1	SAL-meta	49.26 (18.88)	0.27	0.07	6.80	0.01
2	SAL-meta	54.46 (18.23)	0.30	0.14	7.61	0.001
	AN-LAN	4.70 (1.68)	0.28
3	SAL-meta	55.84 (17.88)	0.31	0.17	6.72	0.0004
	AN-LAN	4.83 (1.64)	0.29
	SAL	−23.24 (11.17)	−0.20
	Anxiety/depression
1	SMN-AN	−6.92 (3.01)	−0.24	0.05	5.29	0.02
2	SMN-AN	−7.73 (2.98)	−0.27	0.08	4.74	0.01
	SAL	−34.65 (17.31)	−0.21
3	SMN-AN	−8.00 (2.92)	−0.28	0.12	4.76	0.004
	SAL	−36.42 (16.98)	−0.22
	SAL-LAN	−6.06 (2.89)	−0.22
	Agitation/disorganization
1	SAL-CEN	6.18 (2.71)	0.24	0.05	5.20	0.02

AN = auditory network; DMN = default mode network; CEN = central executive network; LAN = language network; SAL = salience network; SMN = somatosensory network; Meta = abbreviation for metastability; when the name of a network is used alone it denotes the within-network functional connectivity of that particular network; between-network connectivity is denoted by joining the names of the respective networks with a hyphen.

B: unstandardized B coefficient, SE B: standard-error of the B coefficient.

Fig. 2. Schematic presentation of the results of regression analyses.

(A) Positive symptoms, (B) Negative symptoms, (C) Anxiety/Depression, (D) Agitation/Disorganization. Each network is presented as a simplified diagram of its key nodes for ease of visualization. Lines connecting the nodes of each network represent its within-network functional connectivity. Lines connecting two networks represent their between-network functional connectivity. Signal representation over the key nodes of a network represents the network’s metastability. For each significant model for each symptom cluster the standardized coefficients of the significant predictor variables (β). The details of the regression analyses are shown in Table 2. AN: auditory network, CEN: central executive network, DMN: default mode network, LAN: language network, SAL: salience network, SMN: sensorimotor network.

3.2.2. Negative symptoms

The best model explained 17% of the variance in symptom severity (Table 2). The neuroimaging variables that made a statistically significant contribution to the model were the metastability and within-network connectivity of the SAL network and the connectivity between AN-LAN. For each unit increase in SAL metastability and in AN-LAN connectivity, the model predicted higher BPRS negative symptom score while for any unit increase in the within-network connectivity of the SAL network, the model predicted lower BPRS negative symptom score (Table 2, Fig. 2).

3.2.3. Anxiety/depression

The best model explained 12% of the variance in symptom severity (Table 2). The neuroimaging variables that made a statistically significant contribution to the model were the within-network connectivity of the SAL and the connectivity between SAL-LAN and SMN-AN. For any unit increase in any of these connectivity measures, the model predicted lower scores in BPRS anxiety/depression symptoms (Table 2, Fig. 2).

3.2.4. Agitation/disorganization

The best model explained 5% of the variance in agitation/disorganization (Table 2). Only the connectivity between SAL-CEN made a statistically significant contribution such that increased connectivity between these networks predicted higher BPRS agitation/disorganization scores (Table 2, Fig. 2).

4. Discussion

The present study evaluated the explanatory value of fMRI-derived measures of RSN functional organization for the symptoms of schizophrenia. Importantly, we did not limit our investigation to positive symptoms. Instead, we considered the entire spectrum of psychopathology in schizophrenia in accordance with previous studies which have shown that positive symptoms, amotivation/negative symptoms, depressive/anxious symptoms and agitation/disorganization represent separable factors (Peralta et al., 2013; Russo et al., 2014). This factor structure is present at the first psychotic episode; it is longitudinally stable (Russo et al., 2014) and is captured by the BPRS (Lukoff et al., 1986). Our results provide further empirical support to the notion that brain dysconnectivity underlies symptomatic expression in schizophrenia. The parameters of RSN functional organization considered here explained a sizable proportion of the variance in symptom severity with the exception of disorganization.

To our knowledge, this is the first study on RSN synchrony and metastability in schizophrenia and the results reinforce the importance of the SAL network. The SAL network is anchored in the anterior insula and dorsal anterior cingulate cortex and has prominent prefrontal, limbic and subcortical connections (Seeley et al., 2007). It is widely accepted that the primary function of this network is to integrate sensory, visceral and affective data in order to direct attention and shape cognitive and behavioral responses (Seeley et al., 2007; Sridharan et al., 2008; Medford and Critchley, 2010; Menon and Uddin, 2010). Meta-analyses of neuroimaging studies in schizophrenia have reported reductions in the volume of the constituent regions of the SAL network (Haijma et al., 2013), in its internal coherence and its integration within the brain functional connectome (Dong et al., 2017). It has been proposed that the SAL network disruption impairs the appropriate attribution of salience to internal and external events and leads to reality distortion (Menon, 2011; Northoff and Qin, 2011; Alderson-Day et al., 2016) and to anxious/depressive symptoms (Kaiser et al., 2015). Disruption in the SAL network may also contribute to negative symptoms as difficulties in allocating motivational salience and initiating appropriate responses may present as lack of motivation and reduced emotional responsivity (Menon, 2011). The results of this study partly confirmed these predictions as parameters of the SAL functional organization contributed to negative, anxious/depressive and disorganization symptoms.

Importantly, the findings of this study offer novel putative mechanistic insights into the SAL network dysfunction in schizophrenia based on the dynamic features of the network. Synchrony between network regions is required for the exchange of information; the degree of synchrony and its variability (i.e., metastability) over time define different network states; weakly synchronized states are dominated by noise while highly synchronized states prevent information flow (Fries, 2005; Kelso, 2012; Tognoli and Kelso, 2014). The increase in SAL network synchrony in patients compared to healthy participants implies impediments in the propagation of salience-related information within the network. The metastability of the SAL network was also increased, indicating greater variability in SAL network states in patients. This finding could be interpreted as a reduction in the capacity of the SAL network to maintain contextually-appropriate network states but may also indicate abnormally increased flexibility of the SAL network that allows it to adopt contextually-inappropriate configurations. Future work involving a broad-range of behavioral and cognitive tasks in the scanner will be required to better characterize the patterns of impairment in salience monitoring and allocation associated with altered SAL network synchrony and metastability in schizophrenia. We also note that changes in metastability and synchronyof moderate effect size were also present in patients in the VN, AN, and LAN networks. These initial results underscore the value of larger studies on network dynamics in schizophrenia.

The DMN integration emerged as the primary contributor to positive symptoms. The DMN is implicated in self-referential and integrative processes that do not require attention to external stimuli or events (Raichle et al., 2001; Greicius et al., 2003). In the current study, greater functional integration of the DMN with the CEN and SMN predicted lower positive symptom severity. Abnormalities in the functional integration of the SMN are likely to disrupt the orderly processing of sensorimotor information and may thus contribute to reality distortion. The CEN is involved in goal-directed and task-related selection of stimuli and responses (Corbetta and Shulman, 2002). The two networks typically have an antagonistic relationship and a successful suppression of the DMN by the CEN is considered crucial for optimal cognitive task performance (Sridharan et al., 2008).

Of note, higher integration between the SMN-AN predicted lower severity for positive symptoms. This finding is aligned with the impaired corollary discharge (or efference copy) hypothesis of schizophrenia which posits that positive psychotic symptoms reflect impairment in somatosensory and auditory integration (Feinberg, 1978; Frith et al., 2000; Mathalon and Ford, 2008). These corollary signals accompany motor action and enable prediction of the sensory consequences of self-initiated action and thus their differentiation from externally triggered events (Von Holst and Mittelstaedt, 1950; Sperry, 1950; Blakemore et al., 1998). Multisensory prediction deficits in schizophrenia, and their relation to hallucinations and delusions, have been reported in multiple electrophysiological and imaging studies (Ford and Mathalon, 2012; Picard and Friston, 2014; Shergill et al., 2014).

In addition to its association with positive symptoms, greater SMN-AN integration also predicted lower levels of anxious/depressive symptoms. This observation is relevant to on-going debates about the association between multisensory prediction errors, positive symptoms and anxious/depressive symptoms. It has been suggested that multisensory prediction errors giving rise to positive symptoms may also trigger distress and depression either because of their unpredictable nature (Gallagher, 2005; Fletcher and Frith, 2009) but also because they often have unpleasant, persecutory and derogatory content (Kjelby et al., 2015). Conversely, heighted levels of anxiety/depression may give rise to multisensory prediction errors since such feelings often precede the onset or recurrence of positive symptoms (Freeman and Garety, 2003; Allen et al., 2005).

Some limitations to this work should be noted. First, the sample size was moderate and we may have restricted power to detect some effects, particularly in connection to network dynamics and integration (Supplemental Table S1). At the same time, those findings that were statistically significant are likely to reflect core pathophysiological processes. Moreover, we provided the first estimates of the effect size of casecontrol differences in RSN synchrony and metastability that would be useful in guiding future study design in schizophrenia. Second, the two diagnostic groups were not matched for sex. Analyses of large fMRI datasets have shown that sex is relevant for resting-state functional connectivity (e.g., Biswal et al., 2010) but not for synchrony or metastability (Lee and Frangou, 2017; Lee et al., 2018). In any case, sex did not make any statistically significant contribution to the results of the current study. Third, we used masks to define the RSNs in preference to other methods for partitioning the resting-state connectome. The most widely used alternative is independent component analysis. This approach may be more sensitive to individual-level variability but functionally defined masks such as the one employed here have the advantage of greater cogency and reproducibility (Laird et al. 2011; Smith et al., 2009; Power et al., 2011; Doucet et al., 2011). Fourth, the severity of the clinical symptoms varies across samples and within the same patient over time. The patients included in this study were in the early stages of the disorder and were selected to minimize potential confounders associated with chronicity. Confirmation of the robustness of our findings requires replication in larger samples and in longitudinal studies. Fifth, medication status was based on patients’ self-report and on information from medical records regarding their prescriptions. The majority of patients were on regular antipsychotic medication. Although the daily antipsychotic dose did not correlate with any neuroimaging measure, the effect of medication cannot be conclusively excluded.

Overall, the results reported here demonstrate that abnormalities in RSN connectivity and metastability explain a significant proportion of the symptom dimensions of schizophrenia. They also reinforce the importance of the SAL network for schizophrenia and point to novel mechanisms that may explain the abnormalities in salience processing observed in schizophrenia.