Neurophysiologic Characterization of Resting State Connectivity Abnormalities in Schizophrenia Patients

Background: Patients with schizophrenia show abnormal spontaneous oscillatory activity in scalp-level electroencephalographic (EEG) responses across multiple frequency bands. While oscillations play an essential role in the transmission of information across neural networks, few studies have assessed the frequency-specific dynamics across cortical source networks at rest. Identification of the neural sources and their dynamic interactions may improve our understanding of core pathophysiologic abnormalities associated with the neuropsychiatric disorders. Methods: A novel multivector autoregressive modeling approach for assessing effective connectivity among cortical sources was developed and applied to resting-state EEG recordings obtained from n = 139 schizophrenia patients and n = 126 healthy comparison subjects. Results: Two primary abnormalities in resting-state networks were detected in schizophrenia patients. The first network involved the middle frontal and fusiform gyri and a region near the calcarine sulcus. The second network involved the cingulate gyrus and the Rolandic operculum (a region that includes the auditory cortex). Conclusions: Schizophrenia patients show widespread patterns of hyper-connectivity across a distributed network of the frontal, temporal, and occipital brain regions. Results highlight a novel approach for characterizing alterations in connectivity in the neuropsychiatric patient populations. Further mechanistic characterization of network functioning is needed to clarify the pathophysiology of neuropsychiatric and neurological diseases.


INTRODUCTION
Neurophysiologic abnormalities are commonly studied in patients with schizophrenia in response to experimental stimuli, cognitive, tasks, and even at rest. Neural oscillations play an essential role in cortico-cortical transmission and the integration of information across neural networks supporting critical brain functions, including perception, attention, and other higherorder cognitive functions (1)(2)(3)(4)(5)(6)(7).
Neural oscillations can be measured in the scalp electroencephalogram via a variety of analytic and experimental settings [e.g., spontaneous, evoked, induced, and emitted (8)(9)(10)], which have productively resulted in the identification of abnormalities across a broad range of conditions in schizophrenia patients. Task-related (i.e., evoked and induced) high frequency oscillatory abnormalities in schizophrenia patients, especially for gamma band oscillations (i.e., above 30 Hz), have been consistently reported among the myriad neurophysiological abnormalities seen in schizophrenia (8,9,(11)(12)(13)(14)(15)(16)(17), and are associated with multiple cognitive deficits in patients (18). In contrast to the widely studied stimulusor task-evoked gamma oscillations, spontaneous oscillatory abnormalities in schizophrenia, particularly in the gamma band, have been relatively less studied.
Resting-state EEG does not require behavioral responses to stimuli or cognitive tasks for elicitation and is already widely used as part of routine neurologic and psychiatric assessments (19). Spontaneous oscillations arise from the synchronous firing of neurons in distributed neuronal networks and are characterized at broadband frequency ranges detectable via scalp sensors. Such oscillations can also be characterized via the flow of spectral information among their calculated neural sources. Identification of the primary contributing neural sources as well as the dynamic interactions among sources of spontaneous EEG activity may elucidate fundamental pathophysiologic abnormalities associated with the illness which may ultimately yield clinically relevant applications (biomarkers of illness, risk of illness, or sensitivity to therapeutic interventions).
A recent review article reported that schizophrenia patients showed increases in the canonical theta, alpha, and beta bands, but with no difference in the delta band activity in scalp level responses (20). Despite enthusiasm for measures of gamma band phase locking and synchronization to steady-state stimulation (8,21), resting state gamma band activity has not been commonly studied. Moreover, while such scalp-level responses have been extensively described, the spatial information of neural network dynamics underlying frequency-band specific resting-state EEG activity in schizophrenia patients is largely unknown. To our knowledge, only one paper, by Andreou et al. (22) reported increased resting-state gamma-band functional connectivity across the Rolandic operculum, a region that includes superior temporal and inferior frontal gyri, in schizophrenia patients.
In this study, a novel multivector autoregressive modeling method was developed and applied to assess the effective connectivity of resting-state EEG activity among cortical sources in schizophrenia patients and healthy comparison subjects. This data-driven approach enables an analysis of cortical network dynamics with directed information flow [e.g., Granger causality (23); increased or decreased EEG phase coherence between two cortical regions] using a correlation with a time delay. We hypothesized that patients with schizophrenia would show abnormal increased frequency-specific oscillations (e.g., gamma-band activity) across frontotemporal cortical networks. Furthermore, we aimed to characterize the networks associated with other frequency bands in schizophrenia patients and healthy comparison subjects.

Subjects
EEG data from n = 147 healthy comparison subjects and n = 159 schizophrenia patients was processed. Recordings from n = 2 healthy comparison subjects and n = 5 schizophrenia patients were dropped in the quality control step in the pre-processing of EEG. In the sample of n = 145 healthy comparison subjects and n = 154 schizophrenia patients, age and sex were significantly different between the groups. Therefore, we removed the subjects of extreme value of age and sex, and used a final sample of n = 126 healthy comparison subjects and n = 139 schizophrenia patients in the effective connectivity analysis (Supplementary Method 1, Supplementary Table 1). Resting-state spectral characteristics assessed at a single principal component analysis (PCA)-based composite scalp sensor level were previously reported (24). Antipsychotics, anxiolytics, and anticholinergics were prescribed for 125, 27, and 42 schizophrenia patients, respectively. Since anxiolytics and anticholinergic medications are known to have potential impacts on resting state scalp EEG (25,26), separate analyses of schizophrenia patients who did not have either anxiolytics nor anticholinergics (N = 80) were also conducted. Written informed consent was obtained from each subject. The Institutional Review Board of University of California San Diego approved all experimental procedures (071831, 170147).

Electroencephalography Recording and Pre-processing
Participants sat in a comfortable chair in a quiet room and were instructed to relax and with their eyes open. Subjects were closely monitored via a one-way mirror throughout this brief 5 min session. The recording could be paused if subjects appeared to be drowsy either by direct observation or as indicated in their EEG/EOG recordings. The recording would then be resumed after the subject was reminded to keep their eyes open.

Source Localization Using an Equivalent Current Dipole Model
The values in the column of the mixing matrix derived from ICA were mapped on to the scalp electrodes to obtain IC scalp topography, which represents scalp projection of ICAderived effective EEG sources inside the brain (36). A previous study showed that this scalp topography is modeled well by an equivalent current dipole model, and in fact the "dipolarity" of IC scalp topography correlates with the mutual information reduced by ICA (37). Thus, even though ICA is agnostic on spatial information (electrode locations, electric forward model of the brain, or spatial information about location of the EEG generators), minimizing the mutual information in the decomposed signals naturally achieves a physiologically valid dipolar spatial projection pattern. These findings are often taken as evidence of physiological validity of ICA when applied to scalp-recorded EEG data (independence-dipolarity identity). This estimation of equivalent current dipoles was performed using Fieldtrip functions (38). Two symmetrical dipoles were estimated for scalp topographies (39).

Selection of Independent Components Representing EEG
To select brain ICs among all types of ICs, EEGLAB plugin ICLabel() was used (40). The inclusion criteria were (1) "brain" label probability > 0.7 and (2) residual variance i.e., var[(actual scalp topography) -(theoretical scalp projection from the fitted dipole)]/var(actual scalp topography) < 0.15. Seven subjects were removed because they did not have minimum of 4 brain ICs. The mean number of ICs remained was 12.5 (SD 4.5, range 4-25). To ensure consistency across computations, recordings longer than 300 s were truncated to 300 s. Mean data length was 297.7 s (SD 8.9, range 202-300).

Effective Connectivity Analyses
To calculate the grand-mean effective connectivity across ICs for each group, we applied EEGLAB plugin groupSIFT, which recently demonstrated successful application in other neuropsychiatric disorders (41). Renormalized partial directed coherence [RPDC (42)] was calculated across ICs (single window, logarithmically distributing 50 frequency bins from 2 to 55 Hz). This generated a connectivity matrix with the dimension of IC × IC for each participant. The grand-average optimum model order determined via the elbow detection method was 7.1 (SD 0.6) i.e. delayed effective connectivity up to about 64 ms was utilized. An autocorrelation function (ACF) test showed that probability for the residual to be white was 0.81 (SD 0.04). Data consistency (43) was 88.2 % (SD 4.3). The estimated equivalent dipole locations of the corresponding ICs were convolved with 3-D Gaussian kernel with 20 mm full width at half maximum (FWHM) to obtain probabilistic dipole density (truncated at 3 σ). The dipole density inside the brain space is segmented into anatomical regions defined by custom automated anatomical labeling [AAL (44)]; the original 88 anatomical regions in AAL were reduced to 76 by summarizing basal and deep limbic regions into two umbrella regions, upper and lower basal. The labels "upper basal" and "lower basal" were originally matched to ventral mid-cingulate, "mid-cingulate" as dorsal mid-cingulate, and "insula" as inferior frontal. The individual IC × IC connectivity matrix was thus mapped to a 76 × 76 connectivity matrix, on which RPDC was also mapped as a weighting factor to modulate pairwise dipole density to calculate graph edges. For both groups (healthy comparison subjects and schizophrenia patients), including a minimum of 70% of unique subjects was set to be an inclusion criterion for each graph node to be analyzed in the next stage. Also, for the group comparison (healthy comparison subjects and schizophrenia patients), 48/76 graph nodes showed overlap between the groups, which explained 82.3% of total dipole density, consistent with findings from Loo et al. (41). For the statistics of RPDC in the frequency domain, a weak family-wise error rate control was applied (45,46). The brain graphs were visualized using BrainNet Viewer software (47).
The matrix of the group-difference [healthy subjects (N = 126) and schizophrenia patients who were not treated with anxiolytics or anticholinergics (N = 80)] of each EEG band activity (p < 0.0001, corrected; two-tailed) is also shown in Supplementary Figure 3. We revealed 10 graph edges for delta band, 14 for theta band, 17 for alpha band, 17 for beta band and 6 for gamma band activity (Supplementary Figures 4, 5).

Delta Band Activity (1-4 Hz)
Decreased effective connectivity from a region near the calcarine sulcus to the fusiform, temporal and middle cingulate gyri was detected in delta band in schizophrenia patients compared to healthy subjects (Figures 1-3). A bidirectional increased interaction between the right middle temporal gyrus and the right middle cingulate gyrus was also observed. These connectivities were more prominent in the right hemisphere.

Theta Band Activity (4-8 Hz)
Theta band effective connectivity demonstrated a similar rightsided asymmetry centered on the temporal and middle cingulate gyri in schizophrenia relative to healthy subjects (Figures 1-3). Decreased effective connectivity from a region near the calcarine sulcus to the fusiform, temporal and middle cingulate gyri was also detected in theta band activity in schizophrenia patients compared to healthy subjects. The bidirectional increased interaction between the right middle cingulate gyrus and right middle temporal gyrus was also seen in theta band activity in schizophrenia patients relative to healthy subjects. Increased effective connectivity from the right fusiform gyrus to the right middle frontal gyrus was seen in schizophrenia relative to healthy subjects.

Alpha Band Activity (8-14 Hz)
The overall pattern of alpha connectivity is similar with those observed in theta band activity (Figures 1-3). Increased effective connectivity from the right middle cingulate gyrus to the Rolandic operculum (a region that includes auditory cortex and spans Broadmann areas 41 and 42) was detected in schizophrenia relative to healthy subjects. Increased effective connectivity from the right fusiform gyrus to the right middle frontal gyrus and the bidirectional increased interaction between the right middle cingulate gyrus and the right middle temporal gyrus were also seen in schizophrenia relative to healthy subjects.

Beta Band Activity (14-30 Hz)
Abnormal patterns of connectivity were observed among temporal, middle cingulate and occipital regions. These networks overlapped across beta and alpha band activity in schizophrenia patients compared to healthy subjects (Figures 1-3). Increased effective connectivity from the right middle cingulate gyrus to the Rolandic operculum was also seen in schizophrenia relative to healthy subjects. Increased bidirectional information flows between the right middle frontal gyrus and the right fusiform gyrus were detected in schizophrenia patients compared to healthy subjects.

Gamma Band Activity (30-50 Hz)
The increased bidirectional information flows between the right middle frontal gyrus and the right fusiform gyrus were also detected in gamma band activity in schizophrenia patients compared to healthy subjects (Figures 1-3). Although the abnormal neural network was overlapped across gamma and beta band activity in patients compared to healthy subjects, the overall structure was simpler and more localized for gamma vs. beta band activity. This relatively simpler structure for gamma band suggests that higher frequency abnormal networks in schizophrenia compared to healthy subjects consisted of more independent local networks that therefore did not connect with other regions.

DISCUSSION
Schizophrenia patients showed broad and widespread hyperconnectivity of cortical networks underlying resting-state EEG activity. Specifically, the following findings were detected; (1) decreased information flows from a region near the right calcarine sulcus to the right fusiform gyrus in delta band activity, and bidirectionally increased interactions between the right fusiform gyrus and the right middle frontal gyrus in beta and gamma band activity (i.e., "visual network"; Figure 4); (2) Increased information flow from the right middle cingulate gyrus to the Rolandic operculum across alpha and beta bands in schizophrenia patients compared to healthy subjects (i.e., "auditory network"; Figure 4); With few minor exceptions, these results were largely confirmed in a subgroup of schizophrenia patients who were not on anxiolytics or anticholinergics.
The present results replicate abnormal effective connectivity between frontotemporal regions in schizophrenia patients (22). Increased functional connectivity of alpha band activity at the superior parietal and the occipital lobe area at scalp levels of EEG in schizophrenia patients (N = 28) were previously reported by Liu et al. (49); we successfully replicate and extended the results showing the involvement of frontotemporal regions. We also previously reported that abnormal spontaneous gamma band activity measured via a spatial PCA of scalp channel data was associated with verbal memory performance (24). Although the PCA method used in the prior study provides a data-driven approach for characterizing macroscale/global oscillatory effects at the scalp, the neural interactions among sources were not assessed. The current results suggest that the spontaneous global gamma band abnormalities observed in schizophrenia patients at the scalp level appear to be generated by interactions between prefrontal and temporal regions.
A previous study by Andreou reported increased thetaband resting-state connectivity across midline, sensorimotor, orbitofrontal regions and the left temporoparietal junction in schizophrenia patients (N = 19) (50), consistent with our findings of right dominant increased effective connectivity among the temporal and middle cingulate gyri in broad band EEG activity including theta band activity. Inconsistencies in the laterality of effects, however, may be due to the difference of age or clinical severity. For example, the mean age of schizophrenia patients in the Andreou et al. study (50) was 23.5 vs. 44.6 years in the current study. Despite this difference in theta, our finding of increased alpha connectivity at the right temporal region at scalp levels of EEG in schizophrenia patients is fully consistent with the findings of Liu et al. (49).
Of note, despite the temporal differences between neural activity detected by very low frequency blood oxygenationdependent (BOLD) hemodynamic responses and EEG, our resting-state EEG connectivity findings are also fully compatible with functional magnetic resonance imaging study (fMRI) findings of default mode network abnormalities in schizophrenia patients (51). The present findings of greater connectivity between the middle frontal, anterior cingulate and middle temporal gyri regions, is consistent with greater default mode network activation in schizophrenia (52).

Networks Centered at the Visual Cortex
The effective connectivity networks centered in the calcarine sulcus and the fusiform gyrus in broad band EEG activity FIGURE 4 | Two major abnormal networks associated with either visual and auditory information processing detected in schizophrenia patients at rest.
were unexpected. The calcarine sulcus is a deep fissure that starts in the temporal lobe that continues into the occipital lobe with the primary visual cortex (V1) centered in this region. The fusiform gyrus is large region in the inferior temporal cortex that also has a functional role in visual information processing (53), including object and face recognition, and the recognition of facial expressions (Figure 4). Indeed, despite these unexpected associations, results are consistent with Morita et al. (54) findings of associated eye movement impairments with gray matter cortical thickness in schizophrenia patients in the middle frontal and fusiform gyri and the lateral occipital cortex. Although speculative, patients with schizophrenia may show abnormal excessive simultaneous activation of various perceptionrelated brain regions, which may ultimately contribute to clinical symptoms such as hallucinations, aberrant salience, and delusions.

Networks Centered at the Auditory Cortex
Increased information flows were detected in schizophrenia patients from the right middle cingulate gyrus converging on the right Rolandic operculum (Figure 4). In the current analysis, primary auditory cortex is located in the region labeled the Rolandic operculum. Previous studies have demonstrated that deficits in early auditory information processing in schizophrenia patients as indexed by mismatch negativity (48,55) and gammaband auditory steady-state responses (9,56,57) are supported by distributed networks where the genesis of the responses are detected in the superior temporal gyrus (a region that includes auditory cortex) which subsequently propagate across other temporal and frontal brain regions. The present results suggest that resting state abnormalities in schizophrenia patients are present across multiple frequency bands and over relatively large spatial networks. Measures of network connectivity from cingulate gyrus to the auditory cortex may be therefore account for multiple neurophysiologic biomarkers and show promise as a future candidate biomarker of abnormalities in schizophrenia patients.

Limitations
Results of this study should be considered in the context of several limitations. First, this is a cross-sectional cohort study of a heterogeneous sample of schizophrenia patients, the majority of whom were prescribed complex medication regimens. While comparisons of patients prescribed vs. notprescribed medications that are known to impact resting state scalp responses (i.e., anxiolytics or anticholinergics) and healthy subjects showed similar patterns of results, it is possible that other medications including antipsychotics or symptoms may contribute to the observation of abnormal network dynamics. Carefully controlled prospective randomized controlled trials are needed to disentangle medication effects. Despite efforts were made to obtain medical/prescription records for all subjects, self-and informant reports of medication compliance, ultimately medication compliance could not be confirmed for the majority of patients in this study. As such, more rigorous analyses of medication doses and connectivity analyses were not pursued. Second, only 40 EEG channels were used for the analyses in this study. Future studies may benefit from the use of highdensity EEG recordings with at least 64 channels (58), individual MRI data, and digitized scalp sensor locations rather than template head models and reliance on standardized electrode locations for potentially improved accuracy of source dynamics. Third, while we believe that elaboration of neural system dynamics reported here will be broadly applicable to multiple neuropsychiatric disorders, we acknowledge the possibility that results from schizophrenia patients with an established illness may not generalize to other populations like clinical high risk or first episode psychosis. Nonetheless, given improvements in medical care and life expectancy, patients with more chronic schizophrenia are likely to represent an increasing proportion of the total schizophrenia population; characterization of abnormal network dynamics among real-world patients via data-driven approaches for assessing network dynamics may ultimately be useful for application as biomarkers the development of therapeutics for this largely underserved population.

CONCLUSIONS
Results of this study provide evidence that abnormal resting-state EEG oscillations are driven by patterns of hyper-connectivity across multiple frequency bands and a distributed network of the frontal, temporal and occipital brain regions that are involved in visual and auditory information processing in schizophrenia patients. Future studies of the neural mechanisms underlying the networks detected in this study, in both future clinical and animal studies, are needed to clarify the pathophysiology of neuropsychiatric and neurological diseases in support of the development of novel therapeutic interventions.

DATA AVAILABILITY STATEMENT
The datasets generated for this article are not readily available to third parties. Requests to access the datasets should be directed to Gregory A. Light, glight@health.ucsd.edu.

ETHICS STATEMENT
The studies involving human participants were reviewed and approved by Institutional Review Board of University of California San Diego. The patients/participants provided their written informed consent to participate in this study.