Auditory Verbal Hallucination-Specic Functional Brain Networks in Schizophrenia: A Study Based on Graph Theory

Although mounting previous studies have characterized auditory verbal hallucinations (AVH) related brain network abnormalities in the patients with schizophrenia, AVH related brain network alterations based on graph theory was rarely reported. In addition, the relationship between the features of AVH related brain networks based on graph theory and clinical features of schizophrenia patients with AVH is unclear. Our study to explore associations among network metrics, and clinical features in schizophrenia patients with AVH. Thirty-one schizophrenia patients without AVH, 17 patients with AVH, and 31 healthy controls were examined by functional magnetic resonance imaging. Graph theory method was performed to analyze the topological properties of functional network in three groups. Results Our results showed that schizophrenia patients with AVH displayed decreased local network eciency, clustering coecients, and nodal eciency of the right dorsolateral prefrontal cortex. Local network eciency was positively correlated with AVH characteristics.


Introduction
Auditory verbal hallucinations (AVH) are the perceptions of voice without corresponding external stimuli, are one of typical symptoms of schizophrenia (SZ) [1], affect about 60%-90% patients with SZ [2]. AVH can also occur in other mental disorders including bipolar and major depression disorders, as well as in the general population [3,4]. At present, the neural basis of AVH is not fully clear. Neuroimaging studies have suggested abnormal brain structure and function in patients with AVH [5].
In recent years, functional Magnetic Resonance Imaging (fMRI) is commonly used to investigate the features of brain activity in schizophrenia patients with AVH. Extensive studies have shown that brain function of schizophrenia patients with AVH differ from those of without AVH and healthy controls [4,[6][7][8]. For example, a previous meta-analysis indicated that schizophrenia patients experiencing AVH showed increased activity in Broca's area, anterior insula, precentral gyrus, frontal operculum, middle and superior temporal gyri, inferior parietal lobule, and hippocampus/parahippocampus region [8]. Another literature reported that altered brain resting-state networks between default mode network and cognitive control, salience network [9]. In addition, the abnormal functional connectivity of thalamic is correlated with hallucinations, delusions, and bizarre behavior [10]. Meanwhile, Zhuo et al. found that AVH-speci c cerebral blood ow was increased in the auditory and striatal areas and decreased in visual and parietal cortices [11]. Structural neuroimaging studies have indicated grey matter volume reductions in the superior temporal gyrus, middle temporal gyrus, left postcentral gyrus, and posterior cingulate gyrus in SZ patients with AVH [5,12]. Previous study found that AVH-related critical brain regions are mainly located in prefrontal cortex, auditory cortex, superior temporal gyri, insula cortex and anterior cingulate cortex. However, the current study focuses on the interaction of several brain regions, and there are still few studies on the characteristics of the whole brain network in schizophrenia patient with AVH.
Graph theoretical analyses is a powerful tool that can be used to analyze the topological properties of complex brain networks. This method describes brain networks as graphs composed of nodes and edges [13]. Previous studies have indicated that topological properties of brain networks are disrupted in schizophrenia [14]. For example, Zhu et al. reported the "small-worldization" and network e ciency of functional networks is decreased in patients with schizophrenia [15].
In this study, to systematically investigate associations among network metrics, and clinical features, we applied the graph theory method to quantitatively analyze functional brain connectivity. Herein, we hypothesized that global and local topological properties of functional network of schizophrenia patients with AVH would be abnormal, and the features of AVH based on graph theory in the patients with schizophrenia would be related to the clinical other features of schizophrenia.

Participants
Forty-eight patients with schizophrenia (17 with AVH, 31 without AVH) and 31 age and gender matched healthy controls were included. All patients were diagnosed using the Structured Clinical Interview for DSM-IV (SCID). Healthy controls were screened using the non-patient version of the SCID to rule out lifetime psychosis. Symptoms of psychosis were assessed using the Positive and Negative Syndrome Scale (PANSS).
[16] The Auditory Hallucination Rating Scale (AHRS) was used to evaluate the severity and characteristics of AVH. [17] Exclusion criteria consisted of: (1) a history of alcohol or substance abuse; (2) MRI contraindications; (3) pregnancy; and (4) a history of brain injury, epilepsy, glaucoma, or diabetes. Our study was approved by the Ethics Committee of Tianjin Anding Hospital, and written informed consent was obtained from all subjects.
Subjects were scanned using a 3T Magnetom Skyra scanner (Siemens, Erlangen, Germany). During the scans, participants were required to relax and to close their eyes and without falling asleep. In addition, sponge pads were used to limit head motion. Resting-state functional magnetic resonance imaging (fMRI) was conducted using an echo-planar imaging sequence (repeat time = 750 ms, echo time = 30 ms, slice thickness = 3, ip angle = 54°, eld of view = 222×222 mm, voxel size = 3×3×3 mm 3 ; 480 volumes). All images were visually examined by an experienced radiologist to exclude visible artifacts. fMRI data preprocessing was performed by Statistical Parametric Mapping software (SPM12, https://www. l.ion.ucl.ac.uk/spm) and GRETNA.
[18] Preprocessing included removal of the initial ten images; slice timing correction; realignment to the middle image; spatial normalization of the image to the Montreal Neurological Institute template; re-sampling to 3×3×3 mm 3 ; spatial smoothing with 6 mm 3 Gaussian kernel; and band-pass ltration (0.01-0.08 Hz). Finally, white matter, cerebrospinal uid, and 24-parameters of head motion were regressed out. To minimize the effect of head motion, subjects with a maximum head motion > 2 mm or > 2° were excluded.

Network construction
Construction of functional brain networks and graph theoretical analyses were performed using GRETNA [18]. We utilized the atlas of Dosenbach to parcel the brain into 160 cortical and sub-cortical regions of interest (ROI) [19] that act as network nodes. The atlas has been widely used in the study of brain networks [20,21]. Averaging time series were extracted from each ROI. Pearson correlation coe cients between each pair of ROI time courses were considered as the edges of functional networks, and represented inter-ROI functional connectivity strengths. This process produced a 160×160 correlation matrix for each participant (Fig. 1).
Brain networks of different subjects differ in the number of signi cant edges [22]. To ensure that each graph had the same number of edges, we de ned a wide range of network cost thresholds. Sparsity is the actual number of edges divided by the maximum possible number of edges in the network [23]. The selection of sparsity is based on the following criteria: (1) the average degree (the degree of a node is the number of edges linked to the node) of overall nodes in each threshold network was larger than 2 × log(N), where N is the number of nodes; and (2) the small-worldness of the threshold networks was larger than 1.1 for each subject [24]. According to these criteria, cost thresholds ranged from 0.03 to 0.30, with step = 0.01.

Network metrics
We calculated both global and local characteristics of functional networks at each sparsity threshold. The network metrics included: (1) small-world characteristics involving small-worldness (σ), clustering coe cient (C P ), characteristic shortest path length (L P ), normalized clustering coe cient (γ), normalized characteristic pathlength (λ), (2) network e ciency related global e ciency (E gloabl ), and local e ciency (E local ). The local characteristics included node e ciency (E i ), nodal clustering coe cient (C i ), and node degree (S i ) [25,26]. C P measures the local cliquishness of network, and quanti ed the local interconnectivity of a network. L P represents the overall routing e ciency of a network. γ is C P of real network divided by C P of random network; λ is L P of real network divided by L P of random network. The global e ciency measures the ability of parallel information transmission over the network. The local e ciency indicated the network fault tolerance, re ecting the communication e ciency between the rst adjacent nodes when it is eliminated. Ei characterizes the e ciency of parallel information transfer of that node in the network. C i measures the likelihood its neighborhoods are connected to each other. S i re ects its information communication ability in the functional network. Finally, we calculated the area under the curve (AUC) for all network metrics at each cost threshold, providing an overall value for the topological characterization of functional brain networks independent of the selected sparsity threshold.

Statistical analysis
One-way analysis of variance was used to examine signi cant AUC differences between the three groups, and to regress out gender and age. Post hoc analyses were performed to test inter-group differences. Multiple comparison correction was used by false discovery rate (FDR, p = 0.05). We further evaluated the association between network metrics and AVH by using Spearmen's correlation analysis. The above steps were completed with GRETNA[18] and SPSS 23.

Demographic and clinical features
Demographic and clinical features of the three groups are summarized in Table 1. There were no signi cant differences in age (F = 2.903, p = 0.061) and gender (χ 2 = 0.407, p = 0.816) between the three groups. The AVH and non-AVH (nAVH) groups did not differ signi cantly in duration of illness (t = − 1.161, p = 0.252) and PANSS total scores (t = − 0.366, p = 0.716).

Intergroup differences in network metrics
Signi cant alterations of functional network metrics occurred primarily in E local , C net , and in the nodal e ciency of the right dorsolateral prefrontal cortex (DLPFC) (FDR corrected, p < 0.05) (Fig. 2). Furthermore, the AVH group showed lower E local and nodal e ciency of the right DLPFC compared to nAVH groups. Compared with healthy controls, the AVH group displayed decreased E local , C net , and e ciency of the right DLPFC. In addition, decreased E local and nodal e ciency of the right DLPFC was a common abnormality in patients with and without AVH. Reduced C net was the only speci c change in patients with AVH. Other network metrics were not signi cantly different among the three groups.

Relationship between network metrics and AVH severity in AVH group
There were no statistical associations between network metrics and AVH total scores. We further examined the correlation between network metrics and the AHRS checklist, and found that low nodal e ciencies of the right DLPFC correlated with increased amount of distress (r = − 0.743, p = 0.001) and intensity of distress (r = − 0.571, p = 0.017) (Fig. 2).

Discussion
Our study explored the topological organization of whole-brain level functional networks of schizophrenia patients with AVH. Our results indicated that: (1) network metrics including E local , C net , and nodal e ciency of the right DLPFC were decreased in the AVH group; and (2) local network e ciency is negatively correlated with the severity of distress caused by AVH.
Graph theory provides a powerful paradigm for the analysis of the topological organization of complex brain networks in health and in psychiatric diseases [13]. We computed Pearson correlations between different brain regions to act as undirected and unweighted binary graphs for each participant, and compared the network metrics of functional brain networks between the three groups. Schizophrenia patients with AVH had lower C P and E local than healthy controls. C P was used to measure the extent of the local connection or cliquishness in a network, thus representing network segregation [27]. The E local indicated the network fault tolerance, re ecting the communication e ciency between the rst adjacent nodes when it is eliminated [28]. Our results are consistent with a previous study that showed that schizophrenia patients with and without AVH displayed disruption of "small-world" characteristics and lower network e ciencies relative to healthy controls [15]. A task-related EEG study demonstrated decreased clustering coe cients in patients with schizophrenia, which correlated with increased negative and cognitive symptom scores [29]. Similarly, Liu et al. reported decreased C net and E local of functional networks in schizophrenia patients [30]. In contrast, Yu et al. found that C net and E local are increased in patients with schizophrenia [31]. In this study, group-independent component analyses were used to deconstruct the brain into independent components and to treat the components as network nodes. The partial correlation coe cients between different components were considered as network edges. Different methods of constructing functional networks may lead to the opposite result. In addition, our study showed that E local is correlated with AVH characteristics. This study further proved that schizophrenia is a disorder of brain connectivity [14,32].
Our results revealed that schizophrenia patients with AVH exhibited decreased nodal e ciency of the right DLPFC. These ndings support previous reports that the functional connectivity of the DLPFC is impaired in schizophrenia [33,34]. DLPFC is an important part of PFC and plays a role in monitoring speech production in language processing, especially in schizophrenia patients with AVH [35]. Numerous neuroimaging literature demonstrated that the resting-state brain function abnormalities of DLPFC is observed in schizophrenia patients with AVH[36-38]. Cui et al. reported that schizophrenia patients with AVH showed increased regional homogeneity in the right DLPFC, and increased in functional connectivity (FC) between left DLPFC and right putamen [39]. Moreover, a previous study indicated that reduced FC between left DLPFC and superior temporal cortex in schizophrenia patients with AVH[36]. Consequently, our results revealed that DLPFC abnormalities might be the underlying neural mechanism of AVH.

Conclusions
Graph theory analysis were performed to examine abnormalities of functional networks in schizophrenia patients with AVH. Schizophrenia patients with AVH exhibited reductions of "small-worldization," local network e ciency, and nodal e ciency of the right DLPFC. Importantly, the nodal e ciency of the right DLPFC was positively correlated with C net only in schizophrenia patients with AVH. These ndings suggest that abnormal functional brain networks constitute the neural basis of AVH in schizophrenia patients.

Limitations
Several limitations of this study should be considered. First, because most patients were medicated, the potential effects of antipsychotic medications on functional brain networks have not been ruled out. Second, because we conducted a cross-sectional study, the evolution of functional network abnormalities over time is unclear. Third, because the duration of illness of all patients was more than two years, some patients may have had AVH that had resolved before study enrollment. Finally, in the sub-group analysis, the number of patients with AVH was small. The sample size should be increased in future studies to improve statistical power. Design ow chart. First, whole brain was divided into 160 ROIs, and each ROI was considered as a node.

Abbreviations
Second, we computed the Pearson correlation coe cients between each pair of ROIs, resulting in 160 × 160 correlation matrices. Finally, the binarization of the correlation matrix was used for graph theoretical analysis. Figure 2