Integrated metastate functional connectivity networks predict change in symptom severity in clinical high risk for psychosis

The ability to identify biomarkers of psychosis risk is essential in defining effective preventive measures to potentially circumvent the transition to psychosis. Using samples of people at clinical high risk for psychosis (CHR) and Healthy controls (HC) who were administered a task fMRI paradigm, we used a framework for labelling time windows of fMRI scans as ‘integrated’ FC networks to provide a granular representation of functional connectivity (FC). Periods of integration were defined using the ‘cartographic profile’ of time windows and k‐means clustering, and sub‐network discovery was carried out using Network Based Statistics (NBS). There were no network differences between CHR and HC groups. Within the CHR group, using integrated FC networks, we identified a sub‐network negatively associated with longitudinal changes in the severity of psychotic symptoms. This sub‐network comprised brain areas implicated in bottom‐up sensory processing and in integration with motor control, suggesting it may be related to the demands of the fMRI task. These data suggest that extracting integrated FC networks may be useful in the investigation of biomarkers of psychosis risk.

implicated in bottom-up sensory processing and in integration with motor control, suggesting it may be related to the demands of the fMRI task. These data suggest that extracting integrated FC networks may be useful in the investigation of biomarkers of psychosis risk.

K E Y W O R D S
cartographic profile, clinical high-risk for psychosis, network analysis, network based statistics, network integration, task fMRI

| INTRODUCTION
A key challenge in the clinical management of people at clinical high risk (CHR) for psychosis is that it is difficult to predict whether their presenting symptoms will improve, persist, or progress to a frank psychotic disorder, and if their overall level of functioning will improve or deteriorate Simon et al., 2013). This has led to a search for biological measures that might help clinicians to predict clinical outcomes in this group (Gifford et al., 2017). The present study aimed to achieve this by exploring the association of novel functional connectivity (FC) based biomarkers with longitudinal changes in psychosis symptoms and functioning.
The CHR state is associated with subjective impairments in cognitive function, which are often described as Basic Symptoms (Huber & Gross, 1989). These include difficulties in dividing attention between sensory modalities, and focusing attention on non-salient visual stimuli (Schultze-Lutter, Klosterkötter, Picker, Steinmeyer, & Ruhrmann, 2007). Neuropsychological assessments in CHR subjects indicate that this state is also associated with objective cognitive impairments across multiple domains , including attention and vigilance (Zheng et al., 2018).
Neuroimaging studies have shown a range of alterations in functional connectivity (FC) in CHR participants (Allen et al., 2010;Crossley et al., 2009;Schmidt et al., 2014;Winton-Brown et al., 2017). Such studies have shown alterations in 'static' FC by modelling connectivity profiles over the entirety of a scan. It has been suggested that in healthy individuals however, that the brain switches between periods of network integration and segregation, and that the state of network integration is associated with periods of attention  and task performance (Shine, Bissett, et al., 2016). Recent work has exhibited intermittent alterations in FC in CHR subjects (Du et al., 2018), raising the possibility that switching between network segregation to integration may be perturbed in this group. Such a deficit might underlie the deficits in attention and sensory processing in CHR subjects described above.
The present study aimed to address this issue by using fMRI to examine FC in CHR subjects within periods of integration (acting as a proxy for periods of high attention/vigilance) whilst they were performing a cognitive task.
We computed the 'cartographic profile' (CP) (Guimera, R., & Amaral, L. A. N., 2005;Shine, Bissett, et al., 2016;) during a novelty salience fMRI paradigm, which is a method that can be used to mark integrated and segregated 'metastates' during a scan. Network Based Statistics (NBS) (Zalesky, Fornito, & Bullmore, 2010) were then used to search for network differences during integrated metastates between groups of CHR subjects and controls. The CHR samples were then followed up clinically to determine their clinical and functional outcomes. This allowed us to investigate whether the disruption of FC networks derived from integrated metastates is related to subsequent outcomes in CHR participants. Our first hypothesis was that FC networks derived from integrated metastates would be altered in CHR subjects relative to controls. We then tested the hypothesis that this would be associated with clinical and functional outcomes in these subjects at follow-up.

| Participants
One hundred and sixteen participants consented to the present study.
Four CHR participants were excluded from analysis because they did not fully complete the three task fMRI runs, and one was excluded because a macroscopic abnormality was detected on MRI scanning. A further 30 participants (23 CHR/7 HC) were removed from the analysis due to the stringent motion criteria (see pre-processing section and Supplementary Material Section 2), leaving 81 participants (24 HC/57 CHR).
All participants were studied using fMRI at the Centre for Neuro-  (Bell, 1994). HC participants were excluded if they had a personal or familial history of psychiatric/neurological illness. All participants reported no illicit substance use in the past week and no alcohol use in the 24 hour period before scanning.
Participant demographic information is shown in

| Clinical scales
The CHR state was diagnosed using the CAARMS. Pre-morbid IQ was measured using the National Adult Reading Test (Nelson & Willison, 1991), variation in functioning was measured the functional subscale from the Global Assessment of Functioning scale (Karterud, Pedersen, Løvdahl, & Friis, 1998), and psychotic symptoms were assessed using the positive and negative syndrome scale (PANSS) (Kay, Flszbein, & Opfer, 1987).

| Code sharing
The present study used Matlab scripts made available for modelling Cartographic Profiles  Multiplication of Temporal Derivatives (Shine et al., 2015) and multilayer community structure (Jutla, Jeub, & Mucha, 2011;Mucha, Richardson, Macon, Porter, & Onnela, 2010). NBS was performed using the available software (Zalesky et al., 2010). Post hoc tests and statistics were performed using tailor-made scripts in R 3.5.1 and Python 3.6.8.

| Task fMRI paradigm
Each participant was administered a novelty salience task derived from that described by Bunzeck and Düzel (2006). Each block consisted of 80 standard images, 10 neutral oddballs, 10 novel oddballs, and 10 target images. Participants were asked to respond to target images with a button press. Images were presented for 100 ms followed by a fixation cross, with an inter-image interval of 2,700 ms.
Each run lasted for 6 min (179 volumes, TR 2 s). Reaction time and error rate from this task indicates performance on target button presses and not instances of novelty stimulus presentation. Data from the present study using this task has previously been published using Dynamic Causal Modelling (Modinos et al., 2020).

| Pre-processing
Functional MRI scans were pre-processed using a standard fMRIPREP pipeline (Esteban et al., 2019). A full description of the pre-processing used is given in the Supplementary Materials Section 1. A high pass cut off of 128 s was included in this pipeline (discrete cosine basis during CompCor extraction) (Behzadi, Restom, Liau, & Liu, 2007), which is equivalent to bandpass filtering that had been used in previous research using the same task paradigm (Bunzeck & Düzel, 2006;Modinos et al., 2020).
Head motion has a problematic effect on FC-derived network measures (Satterthwaite et al., 2012;Van Dijk, Sabuncu, & Buckner, 2012). In the present study, participants were excluded from analysis if more than 25% of volumes contained a framewise displacement (FD) > 0.25 mm or DVARS >4%, or they had a mean FD > 0.2.
This was done in addition to an ICA-based removal of motion artefacts (Pruim et al., 2015). In the final cohort mean FC remained significantly different between the two groups (mean FD (SD): HC = 0.10 (0.03), CHR = 0.11 (0.03); t (47.68) = 2.05, p = .046) and so mean FD was controlled for in group comparisons and in associations with changes in functioning/symptom scores. As we were using dynamic FC methods, volume censoring was not performed, in order to preserve the temporal structure of scans. A full description of motion correction procedures is given in the Supplementary Materials Section 2.

| Region of interest definition
Time-series were averaged within cortical/subcortical regions of interest using 5 mm spheres centred at 264 coordinate locations (cortical and subcortical) from Power et al. (2011). To negate the effects of inflated connectivity due to task events, HRF-convolved task event data were regressed out of the time-series, the same process used in another study using the cartographic profile with task fMRI data (Shine, Bissett, et al., 2016). Nodes with high signal drop out were removed from the analysis, leaving 204 of the original 264 regions. An area of signal drop out was defined as an area with a mean time-series T A B L E 1 Demographics including N, median age, sex, ethnicity, handedness, mean years in education, mean IQ, medication, mean GAF symptom and function scores at baseline, mean PANSS positive and negative symptom scores at baseline signal intensity z score (across nodes) less than or equal to −1.64 in any subject, corresponding to outliers <5th percentile.

| Multiplication of temporal derivatives
In order to provide an estimate of FC that is robust against task-based inflation of connectivity between regions, multiplication of temporal derivatives (MTD) was used (Shine et al., 2015). To calculate the first temporal derivative for node i at time-point t (dt it ) of a time series (ts), the bold intensity at time point t − 1 was subtracted from that at time point t (dt it = ts it − ts it − 1 ). Each node was given a vector of dts across the time series (t − 1), which were then normalised by the SD across that vector (σ). A node x node MTD matrix for each time point was calculated by multiplying each dt for each pair of nodes ij (Shine et al., 2015). We used non-overlapping windows of 15 TR (30s), which produced a total of 33 windows (11 windows including 15 volumes in each of 3 runs).

| Cartographic profile
To define whether time windows represented integrated or segregated metastates, a procedure for defining the 'cartographic profile' (CP) was used (Guimera et al., 2005;Shine, Bissett, et al., 2016). The CP is a 2D histogram of the frequency distribution of two graph metrics, the Participation Coefficient (PC) and the Module Degree Z-score (MDZ). Both the PC and MDZ rely on a network being segregated into distinct communities, as the PC is a measure of between-community connectivity, whilst the MDZ a measure of within-community connectivity (for formula see Guimera et al., 2005). The PC and MDZ were determined using the Brain Connectivity Toolbox (http://www.brainconnectivity-toolbox.net/). To find the community structure of all time windows, a multilayer modularity maximisation algorithm was used (γ/ɷ = 1) (Jutla et al., 2011;Mucha et al., 2010). Once this had been found for each time window, the frequency distributions of node wise PC and MDZ were used to create one CP for each time window.
A study of how methodological choices affected the distribution of the CP in this study is described in the Supplementary Materials Section 3. In addition, an exploration of the effects of window-length, window offset, and CP resolution was performed, which supported the parameters used in the current study in the context of extracting periods of highly integrated FC (Supplementary Materials Section 5).
Such methodological choices may not apply to other forms of dynamic functional connectivity analysis where, for instance, higher temporal resolution may be beneficial.

| Creating one integrated network for each participant
We aimed to create one integrated network for each participant, to use in within-and between-group comparisons. This involved taking the CP of each time window, using a k-means clustering algorithm to label time-windowed CPs as integrated/segregated (within participant), then using these labels to create an average integrated FC network for each participant. Participant integrated FC networks were then used in within-and between-group comparisons. This procedure is illustrated in Figure 1.
A deviation in our methodology from that in previous studies (Shine, Bissett, et al., 2016;  give the significance of these sub-networks (Zalesky et al., 2010). It is typical to test the NBS procedure across a range of initial thresholds (Nelson, Bassett, Camchong, Bullmore, & Lim, 2017;Zalesky et al., 2010). In the present study, a range of F threshold from 6 to 28 was used. For any sub-networks found to be significantly associated with change in symptom/functioning scores, we also tested whether baseline scores would yield similar results, to confirm that results were specific to a change in scores. In addition, we tested whether excluding participants taking antipsychotic medication at the time of the MRI scan or at follow up would result in similar findings.

| Task performance
There was no difference between the two groups in hit rate (

| Cartographic profile: Group differences
There were no significant group differences in the median number of

| Between/within group comparisons (NBS)
There were no significant group differences in sub-networks found using either static or integrated connectivity matrices (one-way analy- Using identical multiple regression models it was also possible to find the mean Cohen's f 2 to indicate the effect size of models with the response variable (PANSS positive symptom scores). This was done across all brain network edges (upper triangular of the matrix) giving a mean Cohen's f 2 of 0.46 (SD = 0.30). Restricting this to edges of the sub-network found in the NBS procedure resulted in a mean Cohen's f 2 of 1.84 (SD = .59). Additionally, we examined the correlation between the mean FC of the whole sub-network and changes in PANSS positive scores, in order to test whether the results were due to participants who became more or less symptomatic in terms of the positive relationship with FC. This indicated that the CHR sample included both subjects who became more symptomatic and subjects who became less symptomatic, and that both contributed to the negative relationship with FC in the significant sub-network. Results are shown in Figure 2 and Table 3.
We also tested whether using the baseline PANSS positive symptom scores (as opposed to the change in these scores) as the dependent variable would produce similar results. No significant sub-networks were found, suggesting the findings were specific to longitudinal changes in symptomatology. In addition, the procedure was repeated excluding those who were taking antipsychotic medication at the time of MRI, which suggested the found association with PANSS positive symptom scores to be independent of current antipsychotic medication use.
There were no sub-networks associated with longitudinal changes in GAF or PANSS negative scores, using either static or highly integrated connectivity matrices (multiple linear regression; FWE p ≤ .025; 5,000 permutations). When changes in PANSS positive symptom scores were analysed using static connectivity matrices there were no significant sub-networks.
A supplementary analysis was performed in order to also explore regional variation in integration and segregation (Supplementary Materials Section 7). This was done using the Recruitment (R i ) and Integration Coefficients (I i ) (Mattar, Cole, Thompson-Schill, & Bassett, 2015). Whilst there were no regional integration/segregation values (I i /R i ) found to be significantly different between groups, or associated with change in positive symptom scores in CHR participants, there was a whole brain trend for higher I i in the CHR group.
Notably, using the CP procedure CHR participants spent more time in an integrated metastate (Table 2) and whole brain I i across the whole sample was strongly correlated with time spent in an integrated metastate (R = .76).

| DISCUSSION
The present study has demonstrated how a process for extracting periods of high brain network integration from an fMRI scan can be used to search for novel network biomarkers in a CHR population.
The main finding was that a sub-network identified using this approach was negatively associated with changes in PANSS positive T A B L E 2 Significance tests, medians/means, and SD/interquartile range (IQR) for the proportion (%) of time spent in an integrated metastate, and the number of metastate switches symptom scores. In contrast, analyses using static FC networks did not yield significant results.

| Dynamic functional connectivity network organisation
CHR participants were suggested to have less frequent switching between integrated and segregated metastates, though not when controlling for multiple comparisons. Less switching between integrated and segregated metastates in the CHR group could reflect relatively less dynamic functional brain organisation in the CHR group.
However, metastate switching was not related to mean RT, whereas flexibility, a measure that has been related to task performance in multiple studies (Bassett et al., 2011;Braun et al., 2015;Pedersen, Zalesky, Omidvarnia, & Jackson, 2018;Telesford et al., 2016), did not differ between groups and was negatively correlated with mean RT, suggesting it was a better index of dynamic brain organisation related to task performance. In addition, several studies of flexibility have suggested that schizophrenia is associated with higher, not lower, dynamic network switching (Braun et al., 2016;Gifford et al., 2020).
We summarise that this study does not suggest less dynamic functional brain organisation in CHR populations.
Time spent in an integrated state was also not correlated with any measures of task performance, in contrast to data from a previous study (Shine, Bissett, et al., 2016). However, levels of integration and segregation are dependent on task demands (Shine, Bissett, et al., 2016), and the network dynamics in the present study may be specific to the cognitive demands of the task we used. The high accuracy rate for most participants in the present study suggests the task used required low cognitive effort. Similar to Shine, Bissett, et al. (2016), we found that participants spent more time in an integrated metastate, suggesting some validity of the present methodology.

| Discovery of psychosis-related sub-networks
A sub-network was found to be negatively associated with changes in PANSS positive scores, although only when using FC matrices derived from integrated (as opposed to static) metastates. This suggests FC patterns during periods of network integration may provide a useful suggest that using the framework of separating FC profiles into integrated and segregated metastates may produce a more granular signal than those found using an entire scan.
The present study attempted to use task fMRI data in a manner that disregarded all task events, whilst evaluating underlying functional activity. However, the sub-network associated with change in PANSS positive symptom scores could be related to task demands, specifically the integration of bottom-up sensory processing with required task motor behaviour. A central node in this network covered the right V area of the cerebellum. This area is implicated in taskrelated hand movements (King, Hernandez-Castillo, Poldrack, Ivry, & Diedrichsen, 2019;Stoodley & Schmahmann, 2009. Moreover, another central node was found in the intracalcarine cortex, which plays a key role in processing visual information and is thought to be involved in a bottom-up process of attributing salience to visual information (Koene & Zhaoping, 2007;Li, 2002;Zhang & Li, 2012). Furthermore, connections between visual areas and the parietal lobe could be interpreted as part of the 'dorsal visual stream', which is involved in planning action based on visual information (Galletti & Fattori, 2018;Rizzolatti & Matelli, 2003). In addition, default mode nodes within this sub-network included the precuneus, a centrally important node in the default mode network (Fransson & Marrelec, 2008), and implicated in visuomotor learning (Kawashima, T A B L E 3 Regions, MNI Coordinates (X, Y, Z), resting state network affiliation (Power et al., 2011), and degree (within sub-network) of nodes in a sub-network found to be significantly associated with longitudinal change in PANSS positive scores in CHR participants, using an NBS procedure (multiple linear regression; FWE p ≤ .025; 5,000 permutations) on FC matrices derived from integrated metastates  Roland, & O'Sullivan, 1995;Wenderoth, Debaere, Sunaert, & Swinnen, 2005). Finally, the sub-network included several areas of the cingulo-opercular task control network, thought to be involved in goal-directed behaviour that remains stable across task sets (Dosenbach et al., 2007).
A relationship of the findings to task demands does not necessarily mean that they are not relevant to psychotic symptoms. For example, the severity of positive psychotic symptoms has previously been associated with dysfunction in visual motion integration (Carter et al., 2017) and 'soft' neurological signs (Buchanan & Heinrichs, 1989), which are thought to reflect sensory integration abnormalities, are more frequent in CHR subjects (Lawrie et al., 2001) and patients with psychosis (Dazzan & Murray, 2002)  The stronger FC of the sub-network negatively associated with positive symptoms might reflect compensatory effort in visual sensory motor integration, which was protective against the later development of positive symptoms. This would be consistent with the notion that network integration reflects periods of high cost brain activation, as shown by associations with attention/vigilance (Shine, Bissett, et al., 2016). The sub-network associated with positive symptoms involved highly spatially distributed connections, suggesting a high wiring cost, though alternatively could reflect less efficient brain circuitry.

| Strengths and limitations
Strengths of the present study include the use of a relatively large sample of CHR subjects for an fMRI study, combined with the presence of longitudinal outcome data. The logistical difficulties of recruiting, scanning, and following up very large samples would require multi-site collaboration, which would introduce the confound of site effects (Gifford et al., 2017). Most of the CHR sample (83.87%) were naïve to antipsychotic medication at the time of scanning. Moreover, we controlled for medication exposure by entering it as a covariate in the analyses. Excluding those taking antipsychotic medication at the time of MRI or at follow up also resulted in similar results. Therefore, although antipsychotic medication might influence FC (Bolding et al., 2012;Dandash et al., 2014;Fornito et al., 2013;Nejad, Ebdrup, Glenthøj, & Siebner, 2012), the results from the present study are not suggested to reflect the effects of antipsychotic medication. A caveat however, is that the amount and duration of medication use between time-points was not available to account for historical medication use between visits. Though medication use at the time of assessment was accounted for, medication use between visits may have influenced results.
A key limitation in the present study is that the negative relationship between positive psychotic symptoms and the identified subnetwork appeared to be related both to participant increase and decrease in symptoms. Whilst the findings may suggest possible neurobiology protective of psychosis development, it therefore presents no clinical usage. Additionally, the use of task-based fMRI data adds complexity to the replication and interpretation of the sub-network we found to be associated with changes in psychotic symptoms. The cartographic profile has previously been compared across resting state and various task paradigms, showing increased integration in integrated states with more cognitively demanding tasks (Shine, Bissett, et al., 2016). This suggests there may be a benefit in using a task paradigm to extract highly integrated brain states. Future studies may wish to use more cognitively demanding tasks to extract periods of high integration in exploring psychosis risk.
The results of the present study may be dependent on multiple hyper-parameters (Supplementary Materials Section 2), as well as the tuning of the NBS procedure according to the statistical threshold for suprathreshold links (Zalesky et al., 2010). A high number of impactful methodological choices is a barrier to reliability (Simmons, Nelson, & Simonsohn, 2011;Wicherts et al., 2016). It should be noted that choice of the statistic threshold in NBS does not affect the family wise error rate, rather the sensitivity to sub-network discovery (Zalesky et al., 2010).
Classifying windows based on global cartographic profiles may not take into account regional variation in integration/segregation. A supplementary analysis explored dynamic regional variation in integration/segregation (Supplementary Materials Section 7), suggesting no localised dynamic integration/segregation signals to significantly associate with CHR status or change in symptom scores. As such, we may suggest the whole brain CP procedure to be preferable in searching for clinically relevant neural substrates such as those associated with changes in positive psychosis symptoms in the current study, as compared with mass-univariate testing of regional signals.
An assumption of this study is that the brain switches between whole brain epochs of integration and segregation. Although a k = 2 solution was suggested to be appropriate in the current study (Supplementary Materials Section 4) and in previous similar studies (Fukushima et al., 2018;Shine, Bissett, et al., 2016;Shine et al., 2015;, it is important to highlight that other methodological approaches have resulted in higher numbers of distinct intermittent brain states when clustering time windowed FC profiles Damaraju et al., 2014).

| CONCLUSIONS
The use of methods for defining intermittent states of integration and segregation (Guimera et al., 2005;Shine, Bissett, et al., 2016; may provide a useful framework for future fMRI studies searching for biomarkers in psychosis.