Transdiagnostic multimodal neuroimaging in psychosis: structural, resting-state, and task MRI correlates of cognitive control

Abstract

Background:

Disorders with psychotic features, including schizophrenia and some bipolar disorder, are associated with impairments in regulation of goal-directed behavior, termed cognitive control. Cognitive control related neural alterations have been studied in psychosis. However, studies are typically unimodal and relationships across modalities of brain function and structure remain unclear. Thus, we performed transdiagnostic multimodal analyses to examine cognitive control related neural variation in psychosis.

Methods:

Structural, resting, and working memory task imaging for 31 controls, 27 bipolar, and 23 schizophrenia patients were collected and processed identically to the Human Connectome Project (HCP), enabling identification of relationships with prior multimodal work. Two cognitive control related independent components (ICs) derived from the HCP using multiset canonical correlation analysis + joint independent component analysis (mCCA+jICA) were used to predict performance in psychosis. De novo mCCA+jICA was performed and results correlated with cognitive control.

Results:

A priori working memory and cortical thickness maps significantly predicted cognitive control in psychosis. De novo mCCA+jICA identified an IC correlated with cognitive control that also discriminated groups. Structural contributions included insular and cingulate regions; task contributions included precentral, posterior parietal, cingulate, and visual regions; and resting-state contributions highlighted canonical network organization. Follow-up analyses suggested correlations with cognitive control were primarily influenced by schizophrenia patients.

Conclusions:

A priori and de novo imaging replicably identified a set of interrelated patterns across modalities and the healthy-to-psychosis spectrum suggesting robustness of these features. Relationships between imaging and cognitive control performance suggest shared symptomatology may be key to identifying transdiagnostic relationships in psychosis.

Introduction

Psychosis, classically a hallmark of schizophrenia (SZ) (1), is present in several other disorders including schizoaffective and bipolar (BP) disorders. Importantly, alterations in cognition (2–4), including cognitive control (5, 6), are a key feature of psychosis. Furthermore, cognitive control alterations are observed transdiagnostically (7), including in individuals with SZ, BP (8), and other disorders (9–11). Here, we expand beyond prior work by using multimodal image analysis to examine transdiagnostic patterns of neural variation in structural, resting-state, and task imaging related to cognitive control.

Within the broad construct of cognition (12), cognitive control (13, 14) refers to the set of cognitive functions that enable and support goal-directed behavior and regulation of one’s thoughts and actions (15), including the ability to maintain information over time (e.g., working memory), protect against distraction, and combine novel inputs to provide flexibility in task execution (15, 16). Evidence in psychosis suggests that task flexibility (17) may be the source of generalized neurocognitive deficit in schizophrenia and possibly bipolar disorder (4). Within the psychosis spectrum, SZ have poorer cognitive control performance than healthy controls (HC), and BP often have intermediary performance between SZ and HC (7, 18, 19). This graded performance further supports conceptualizations of psychosis as a dimensional transdiagnostic construct.

The imaging literature has identified transdiagnostic neural alterations in psychosis related to cognitive control. For example, with structural imaging, working memory performance was inversely correlated with cortical thickness in the right rostral anterior cingulate and positively correlated with surface area in the left rostral anterior cingulate and right rostral middle frontal region in a cohort of individuals with SZ and BP (20). Recent meta-analysis of structural alterations in a broad mental illness sample identified significant psychosis-related gray-matter losses in medial prefrontal cortex (PFC), insula, thalamus, and amygdala and significant gray-matter increases in the striatum as compared to HC and non-psychotic mental illnesses (21). Finally, Shepherd et al. (22) examined a transdiagnostic psychosis cohort and dichotomized patients into high and low executive function groups based on two-back working memory performance. Compared to HC, the low executive function group exhibited decreased gray matter volume in bilateral superior and medial frontal gyri, and right inferior operculum and hippocampus.

Resting-state imaging has also identified transdiagnostic neural alterations (23, 24) related to general cognition (25–27), with fewer studies examining cognitive control (28). For example, trail making task performance in persons with SZ and BP positively correlated with average connectivity strength between the whole brain and left caudate, thalamus, and temporal occipital fusiform cortex/lingual gyrus (25). Further, a network connectivity approach in another transdiagnostic cohort (29) identified a significant relationship between global efficiency of the cingulo-opercular network and cognitive and executive function, along with significant decreases in cingulo-opercular network efficiency in psychosis (29). Additionally, an independent component analysis (ICA) approach to resting-state data (27) identified significant decreases in connectivity between a fronto-occipital component and a combined anterior default mode and prefrontal component in individuals with SZ, psychotic BP, and unaffected siblings. In the same study, decreased connectivity was also identified between meso/paralimbic and sensory/motor components, but this was only observed in individuals with SZ and not psychotic BP. Furthermore, recent analyses from the same dataset identified seven abnormal networks exhibiting significant correlations with cognitive control including visual, working-memory, visuomotor integration, default mode, and frontoparietal control networks (28).

Several studies have examined transdiagnostic alterations in task imaging related to cognitive control. For example, Brandt et al. (30) used ICA decompositions of two-back working memory task and identified nine task-related components. Of those, three components with spatial distribution in frontal and parietal regions corresponding to working-memory networks showed significant graded hyperactivation pattern (SZ>BP>HC). Additionally, Smucny et al. (7) used an a priori contrast approach examining activation in dorsolateral prefrontal cortex (DLPFC) and superior parietal cortex (SPC) during the AX-CPT task in a transdiagnostic population. They identified significant graded task performance (HC>BP>SZ) and BOLD responses in DLPFC and partially significant responses in SPC.

As noted, a number of studies have examined neural alterations in psychosis. However, much of this literature is unimodal (see (31) for review and fuller motivation for multimodal imaging). Thus, it remains unclear how extant results relate across modalities and across disease state. For example, are the same participants with structural correlates of poor cognitive control performance also the same participants exhibiting resting state correlates of cognitive control performance? Multimodal analyses examining structural, resting state, and diffusion imaging and relationships to a cognitive battery in SZ identified a relationship between overall cognitive impairment and variation in cortico-striato-thalamic circuitry (32). Additionally, examination of three modalities derived from structural imaging in a transdiagnostic cohort (HC, BP, and SZ) identified putative associations between gray matter alterations and processing speed, working memory, and attention (33). However, these associations were not significant across multiple diagnostic categories within the same component and did not survive multiple comparison correction. Importantly, both studies did not focus specifically on cognitive control.

We recently used multiset canonical correlation analysis with joint independent component analysis (mCCA+jICA) to study multimodal neural correlates of cognitive control in the normative population (34). The mCCA+jICA framework was chosen as it flexibly identifies patterns across modalities and decomposes them into maximally spatially independent sources of variance (35). This has proven to be a powerful analytic framework and has been used to identify abnormalities in SZ (36) and also discriminate between HC, BP, and SZ (37). Using a community sample from the Human Connectome Project (HCP) (38), we identified relationships between two mCCA+jICA-derived multimodal patterns and individual differences in cognitive control performance (34). These two components included structural and functional contributions from the anterior insula, visual, and parietal regions as well canonical resting-state network structures. Importantly, the findings were replicable in an independent sample of participants from the HCP using both predictive and independent analyses.

The goal of the present study was to examine transdiagnostic multimodal neural alterations in psychosis, using the results of our previous work to guide analyses. To accomplish this, we recruited a transdiagnostic psychosis cohort comprised of HC, SZ, and BP. Participants were imaged using the same HCP-customized scanner and performed a subset of the same HCP tasks. We first assessed the replicability of our normative findings by predicting performance in the psychosis participants using the two components from our prior work that were significantly related to cognitive control. We then performed an independent mCCA+jICA decomposition of this new transdiagnostic dataset to identify novel patterns of alteration.

Methods:

Participants:

Participants were recruited from a broader study of neural alterations in psychosis and included healthy controls (HC), schizophrenia or schizoaffective disorder (SZ), and bipolar disorder (BP). Prior to image pre-processing, there were n=35 HC, n=36 BP, and n=31 SZ available. Of this, n=31/30* HC, n=27 BP, and n=23 SZ had the requisite data for inclusion in this study (*see supplement). Participants were recruited from clinical and community settings in Saint Louis. Participants had no substance use disorder in the prior six months, no clinically significant head trauma, and no neurological diseases. Patient participation criteria included: DSM-IV diagnosis of BP or SZ, age 18–30, and stable outpatient or partial hospital status. HC were recruited to have similar demographics (age, gender, parental level of education) as patients. HC participation criteria included: no history of DSM-IV psychotic disorder and no cognitive enhancing or psychotropic medication for prior three months. Study procedures were approved by the Washington University Institutional Review Board and all participants gave written informed consent.

Behavioral assessment:

A composite measure of cognitive control was generated for each participant from their performance on four tasks (see supplement for details on each task): (1) In-scanner N-Back task; 2) Out of scanner letter N-Back task from the Penn Computerized Neuropsychological battery (PennCNP) (39); 3) Progressive matrices from the PennCNP; and, 4) Penn Conditional Exclusion Task accuracy. Performance on each task was individually Z-scored and averaged to generate the composite measure. This composite had a Cronbach’s alpha of 0.75, suggesting good internal consistency.

Neuroimaging collection and pre-processing

Participants were scanned at Washington University in St. Louis using a customized Siemens “Connectome” scanner developed for the HCP (38). T1 and T2 were acquired at 0.7mm isotropic resolution and BOLD imaging were collected at 2mm isotropic resolution with 720ms TR. Data were pre-processed using the HCP pipelines (40) and further processed to generate the final imaging measures used in the present study as was done in (34). Three imaging modalities were used: 1) cortical thickness (sMRI); 2) resting state functional connectivity correlation matrices (rsfcMRI) generated using cortical (41), cerebellar (42), and subcortical parcels from Freesurfer; and, 3) the HCP N-back working memory task fMRI (tfMRI) – activation in the 2-back condition (see supplement).

Relationship to a priori normative multimodal correlates of cognitive control:

Our previous work using mCCA+jICA identified replicable multimodal patterns from two mCCA+jICA components that were significantly related to cognitive control in healthy participants in the HCP (34). We performed an analysis to determine whether patterns from these two components also predicted cognitive control in the present participants. Given that data were collected and processed identically, we directly applied the three relevant modalities from these two components from the prior study to the source data from the present cohort (referred to as HCP_gC1_IC3 and HCP_gC1_IC7, see supplement). This generated a set of subject-specific weights for the present cohort corresponding to the extent to which a priori components reflected the present data (see supplement). Resultant weights were correlated with the composite cognitive control measure.

mCCA+jICA multimodal imaging analysis:

We also identified mCCA+jICA components de novo in the psychosis data set. Multiset canonical correlation analysis with joint independent component analysis (mCCA+jICA) is an unsupervised analysis framework that identifies relationships across modalities and decomposes data to reveal maximally independent latent sources of variance (31, 35) (see supplement). Briefly, mCCA (43, 44) first aligned the three imaging modalities in order to simplify the correlational structure and maximize inter-subject covariation (32, 34). Next, jICA maximized spatial independence (45) yielding a set of 9 independent components (ICs). Each IC contained a set of linked modalities including maps of cortical thickness (sMRI) and working memory task activation (tfMRI), and a parcel-wise correlation matrix (rsfcMRI). Each IC had a corresponding set of subject-specific weights that reflected the extent to which a given IC comprised the participant’s original data. These weights were then used for statistical analyses to identify brain-behavior relationships and assess group discriminability (see supplement).

Statistical analyses:

Statistical analyses were performed in SPSS and MATLAB with multiple comparison correction using FDR (46). For correlations between subject-specific weights and cognitive control performance, partial correlation was used to correct for differences in group means. We assessed group discrimination performance of each IC using multivariate analysis of variance (MANOVA) in which group was used to predict all three imaging weights (one per modality) related to each IC. Significant MANOVA omnibus tests were followed up with planned contrasts.

Results:

Behavioral and demographics

There were no significant group differences in gender or parental education\SES. Groups differed significantly on age, ethnicity, education, symptoms, and cognitive control performance. SZ had significantly impaired cognitive control performance as compared to HC and BP, with no difference between HC and BP (table 1). Dichotomizing BP participants by psychosis severity revealed that low-psychosis BP performed significantly better than SZ and high-psychosis BP exhibited performance indistinguishable from SZ (table S13).

Table 1:

Demographics, Clinical symptoms, and Behavioral Performance

	Healthy Controls (HC)	Bipolar (BP)	Schizophrenia (SZ)	Omnibus test statistic	p-values
Age	24.6±3.1 (30)	26.7±3.0 (27)	24.9±3.7 (23)	F(2,77)=3.21	Omnibus p=0.046 HC vs SZ p=0.943 HC vs BP p=0.049 BP vs SZ p=0.145
Gender	14F,16M	13F,13M	4F,18M	χ2=6.034	Omnibus p=0.197
Ethnicity (%Caucasian)	57%	80%	29%	χ2=0.006	Omnibus p=0.002 HC vs SZ p=0.047 HC vs BP p=0.066 BP vs SZ p<0.000
Years of education	16.2±2.4 (30)	15.3±2.3 (26)	13.0±1.4 (22)	F(2,75)=14.9	Omnibus p<0.000 HC vs SZ p<0.000 HC vs BP p=0.264 BP vs SZ p=0.001
Parental level of education/SES proxy	2.04±1.3 (27)	2.28±1.3 (25)	2.76±1.1 (17)	F(2,66)=1.71	Omnibus p=0.189
SAPS	0.1±0.7 (30)	2.1±2.9 (27)	4.6±3.2 (23)	F(2,77)=21.03	Omnibus p<0.000 HC vs SZ p<0.000 HC vs BP p=0.008 BP vs SZ p=0.003
SANS	0.8±1.7 (29)	3.3±3.5 (27)	6.3±3.0 (23)	F(2,76)=24.67	Omnibus p<0.000 HC vs SZ p<0.000 HC vs BP p=0.003 BP vs SZ p=0.002
WERCAP Psychosis history	0.90±3.2 (30)	8.78±7.4 (27)	31.83±11.59 (23)	F(2,77)=106.76	Omnibus p<0.000 HC vs SZ p<0.000 HC vs BP p=0.001 BP vs SZ p<0.000
Cognitive control composite	0.40±0.46 (30)	0.16±0.58 (27)	−0.36±0.85 (23)	F(2,77)=9.579	Omnibus p<0.000 HC vs SZ p<0.000 HC vs BP p=0.332 BP vs SZ p=0.013

Text: mean +/− std (n). N-values vary slightly due to missing data. F=Female; M=Male.

Level of education scale: 1=Graduate professional training; 2=Completed undergraduate; 3=Partial college; 4=High school graduate; 5=Partial high school; 6=Junior high school; 7=less than seven years of education. Tukey post-hoc tests used to assess for differences across groups. WERCAP - Washington Early Recognition Center Affectivity and Psychosis screen. SAPS – Scale for the Assessment of Positive Symptoms. SANS – Scale for the Assessment of Negative Symptoms.

Prediction using a priori ICs

Two ICs from (34) (figs. S1–2) were applied to the present dataset and resultant weights partially correlated with cognitive control in order to identify whether a priori ICs predicted cognitive control in the present dataset (table 2, fig. 1, and fig. S3; see supplementary results for individual group correlations). This identified significant predictions from a priori HCP_gC1_IC3 working memory tfMRI and HCP_gC1_IC7 cortical thickness. Follow-up analyses (figs. 1, S3; tables S3–S8) identified a significant interaction driven by correlations between these weights and the cognitive control composite for SZ in HCP_gC1_IC3 tfMRI. There were no significant interactions with group for HCP_gC1_IC7 sMRI. Neither HCP_gC1_IC3 or HCP_gC1_IC7 were group discriminative (tables S16–S17).

Table 2:

Partial correlation results for application of a priori components and de novo component

Imaging modality	A priori HCP_gC1_IC3 application	A priori HCP_gC1_IC7 application	De novo IC3
Cortical Thickness	r = 0.133	r = 0.233*	r = 0.287*
	p = 0.244	p = 0.040	p = 0.011
	n=80	n=80	n=80
Resting State Connectivity	r = 0.016	r = 0.117†	r = 0.258*
	p = 0.887	p = 0.308	p = 0.022
	n=80	n=80	n=80
2 Back Working memory tfMRI	r = 0.438**	r = 0.078	r = 0.363**
	p < 0.000	p = 0.498	p = 0.001
	n=80	n=80	n = 80

^*p<0.05,

^**p<0.01.

All correlations are partial Pearson correlations corrected for group mean differences.

^†Resting state connectivity data from HCP_gC1_IC7 were not significantly correlated with cognitive control performance in the normative data (Lerman-Sinkoff, Sui et al. 2017) and thus were not predicted to be significantly predictive of cognitive control in the present dataset. Correlations for a priori ICs are not FDR corrected as we had strong predictions from our prior work that these modalities would predict cognitive control performance in psychosis. p-values listed for de novo IC3 are the original uncorrected p-values; all three p-values meet the FDR-determined critical p-value of 0.022.

Figure 1: Scatter plots of cognitive control and a priori ICs applied to psychosis cohort.

Two out of six modalities in HCP_gC1_IC3 and HCP_gC1_IC7 significantly predicted cognitive control performance in the psychosis cohort. Scatter plots of the other four modalities are available in fig. S3.

De novo mCCA+jICA

mCCA+jICA was also performed de novo to identify novel domains of variation in the psychosis cohort. This identified a single component, IC3, that significantly correlated with cognitive control for all three modalities after FDR correction. Follow-up partial correlations correcting for group were performed, indicating that all three modalities still significantly correlated with cognitive control (table 2). Follow-up analyses (Figs. 2, S3; tables S9–11) identified significant interactions in all three modalities such that SZ were responsible for the significant correlations between imaging and cognitive control.

Figure 2: Scatter plots of cognitive control and de novo IC3 imaging weights.

Partial correlations between de novo IC3 imaging weights and cognitive control for all three groups pooled were significant for all three modalities. Scatter-plots by group suggested that this may be driven by SZ participants, which was formally tested using regression analyses (tables S6–S8).

Spatial distributions of de novo IC3

The modalities in de novo IC3 (figs. 3–5) were visually inspected and bore strong visual resemblances to a priori HCP_gC1_IC3 (fig. S1). Algorithmic matching using η² confirmed visual similarity by pairing IC3 with HCP_gC1_IC3 as the strongest match across all ICs with 91.5% shared variance across the two ICs (table S1). Matching analyses using individual modalities were also performed (see supplement). For sMRI data in IC3 (fig. 3), the strongest positive contributing areas were in bilateral anterior and posterior insula, temporal pole, cingulate, frontal superior cortex, and temporal gyrus. Thus, improved cognitive control performance was related to greater thickness in these areas.

Figure 3: Psychosis IC3 – sMRI modality.

Given that mCCA+jICA generates component maps with a value at every vertex/voxel/pairwise-correlation, maps were thresholded at |Z|>2 to simplify interpretation of the spatial pattern of results in de novo IC3 (Sui, Pearlson et al. 2015, Lerman-Sinkoff, Sui et al. 2017). Thus, the maps presented in (Fig, 3–5, S1, S2) highlight those elements in the map that were strongest relative to all other elements in the given map (unthresholded maps are available in the supplement, figs. S4–S6). Top row: Thresholded IC3 sMRI map displayed on the HCP Q1–Q6 440 Subject midthickness surface map. Bottom row: same sMRI map displayed on the HCP Q1–Q6 440 Subject inflated surface map.

Figure 5: Psychosis IC3 – working memory tfMRI modality.

Working memory tfMRI from psychosis IC3, thresholded at |Z|>2 (see figure 3 caption for threshold rationale). Top row: Thresholded IC3 tfMRI map displayed on the HCP Q1–Q6 440 Subject midthickness surface map. Bottom row: same tfMRI map displayed on the HCP Q1–Q6 440 Subject very inflated surface map.

For rsfcMRI data in IC3 (Fig. 4), the strongest positive contributions were concentrated in within-network connectivity, the diagonal of the correlation matrix, with improved cognitive control performance related to stronger within network connectivity. Negative contributions were predominantly concentrated off-diagonal and between the default mode network (DMN) and the task positive networks, including the fronto-parietal, cingulo-opercular, parietal encoding and retrieval, dorsal attention, and ventral attention networks. Thus, improved cognitive control performance was related to stronger negative connections between networks.

Figure 4: Psychosis IC3 – rsfcMRI modality.

rsfcMRI correlation matrix from psychosis IC3, thresholded at |Z|>2 (see figure 3 caption for threshold rationale). Network labels: A= None, B= Default Mode, C= Context, D= Fronto-Parietal, E= Salience, F= Cingulo-Opercular, G= Parietal Encoding and Retrieval, H= Dorsal Attention, I= Ventral Attention, J= Visual, K= Somatomotor Hand, L=Somatomotor Mouth, M= Auditory.

For tfMRI data in IC3 (Fig. 5), the strongest positive contributing areas were in bilateral visual cortex, intraparietal sulcus, superior parietal gyrus, precentral sulcus and right middle cingulate, and inferior frontal sulcus, with improved cognitive control performance related to greater activation in these regions. The strongest negative contributions were seen in bilateral isthmus of the cingulate, left supramarginal gyrus, precuneus, and inferior parietal gyrus. Thus, improved cognitive control performance was related to lower activation in these regions.

Group-discrimination

MANOVAs were performed for each IC to assess whether imaging weights were group discriminative. Omnibus MANOVA results were significant with FDR correction only for IC3, the only component significantly related to cognitive control (table 3); all other ICs were not group-discriminative (table S2). Between subjects effects were significant for tfMRI and trended significant for sMRI. Post-hoc tests between groups were significant for group differences between HC and SZ for sMRI and tfMRI, trend-level significant for differences between BP and SZ for all three modalities, and no differences between HC and BP. Although groups did not significantly differ on parental SES, inclusion of parental SES as a covariate improved the model such that there were trend or significant effects for all modalities (table S9).

Table 3:

IC3 group-discrimination MANOVA results

	Item	Value	Test Statistic	p-values	Partial η2
Omnibus tests	Pillai’s Trace	0.219	F(6,154)=3.15	p=0.006	0.109
	Wilk’s Lambda	0.78	F(6,152)=3.21	p=0.005	0.113
Between Subjects Effects	sMRI	n/a	F(2,78)=2.52	p=0.087	0.047
	rsfcMRI	n/a	F(2,78)=1.91	p=0.156	0.061
	tfMRI	n/a	F(2,78)=4.59	p=0.013	0.105
Post-hoc tests	Item	Mean Difference	Standard Error	p-values
	sMRI: HC - BP	0.001	0.018	p=0.936
	sMRI: HC - SZ	0.038	0.018	p=0.044
	sMRI: BP - SZ	0.036	0.019	p=0.060
	rsfcMRI: HC - BP	−0.008	0.019	p=0.679
	rsfcMRI: HC - SZ	0.030	0.020	p=0.132
	rsfcMRI: BP - SZ	0.038	0.020	p=0.066
	tfMRI: HC - BP	0.025	0.022	p=0.255
	tfMRI: HC - SZ	0.070	0.023	p=0.003
	tfMRI: BP - SZ	0.045	0.024	p=0.066

Discussion:

The goal of the present study was to examine transdiagnostic multimodal neural alterations in psychosis related to cognitive control. Application of a priori ICs identified in the HCP to the psychosis dataset significantly predicted cognitive control performance in a cortical thickness map for all groups and in a working memory map for SZ participants. De novo mCCA+jICA analysis of the psychosis dataset identified a single IC that significantly correlated with cognitive control performance in SZ. This IC exhibited highly similar spatial distribution to an a priori IC and was the sole group-discriminative IC from the de novo mCCA+jICA results. Importantly, as described below, our multi-modal analyses illustrate the ways in which there are both deficits in the same regions/networks that cut across modalities, as well as deficits that are modality specific, though contributing to joint prediction of cognitive control. Such results help to link findings in individual modalities of neural structure/function into an associated pattern that is correlated with a central domain of cognitive impairment in psychosis.

Our previous multimodal work studying cognitive control in healthy individuals identified a working memory map in HCP_gC1_IC3 which exhibited predominantly posterior cortical contributions in the tfMRI data from the fronto-parietal (FP), dorsal attention (DA), and visual networks which are hypothesized to support rapid-timescale cognitive control functionality (47). It was surprising that this pattern significantly predicted cognitive control in SZ but not BP or HC even though it was derived from and significantly correlated with cognitive control in healthy participants in the HCP (see limitations). A second component identified in our prior work, HCP_gC1_IC7, contained an sMRI map with positive contributions in the cingulo-opercular, salience, and ventral attention networks and negative contributions in the default mode and visual networks, which are hypothesized to support stable-timescale cognitive control functionality (47). Interestingly, this component significantly predicted cognitive control in the psychosis cohort for all groups. Together, these a priori patterns were partially predictive of cognitive control in the general population as well as individuals with varying levels of psychosis and may provide clues towards localization of deficits in psychosis.

De novo application of mCCA+jICA to the psychosis dataset identified an IC that was both 91.5% similar to a priori HCP_gC1_IC3 and also significantly correlated with cognitive control performance. However, it was surprising that the de novo correlations with cognitive control were significant only for SZ participants even though the a priori IC was significant in healthy controls. Thus, we discuss below the interpretations of these findings in SZ and address the group-specificity of correlations later in the discussion. Nonetheless, the independent identification of two highly similar ICs using data-driven methods suggests these imaging patterns may be robust features in the broader population, though identifying relationships between these patterns and cognitive control performance requires further study.

For tfMRI data, a priori HCP_gC1_IC3 and de novo IC3 were both predictive of cognitive control in SZ. We previously postulated that strong visual contributions may be due to top-down modulation of visual regions, especially given that the working memory tfMRI task was highly visually demanding (48). Independent identification of this pattern de novo may provide intriguing clues towards the source of cognitive control dysfunction in SZ. A number of reports have identified visual system dysfunction in SZ (49–52), though the literature is more mixed for BP (53–58). Theories generated from these lines of research postulate that alterations in visual system functionality leads to aberrant integration of information in higher order cortical areas and lead to dysfunction in cognitive control (49, 52, 59). Thus, the present work could be seen as consistent with the hypothesis that impairments in the function and/or structure of visual cortex may disrupt higher order processing leading to deficits in cognitive control. In the present data, SZ participants who exhibited greater visual region contributions in IC3 tfMRI data had significantly better cognitive control performance. Importantly, the de novo data for HC and BP trended in the same direction as SZ, though this trend was not significant (fig. 2). While the present work is not a direct assessment of the relationship between visual system integrity and cognition, the findings are consistent with these models of psychosis in SZ and thus warrant further study.

For sMRI data, de novo IC3 significantly correlated with cognitive control in SZ and exhibited contributions similar to HCP_gC1_IC3 including the insula, medial prefrontal cortex, and cingulate. Transdiagnostic gray matter volume variability in these regions has been identified in these regions (21), but relationships with cognitive control in psychosis were not tested. However, (21) did identify a positive relationship between volume in these regions and cognitive control in HC. Interestingly, small portions of these regions were also present in a priori HCP_gC1_IC7, although contributions were more scattershot with smaller clusters exceeding the visualization threshold. Importantly, correlations between HCP_gC1_IC7 and cognitive control were significant for all three groups. Thus, the data suggest this finding may indeed be transdiagnostic, though further study is clearly warranted.

For rsfcMRI data, de novo IC3 significantly correlated with cognitive control in SZ and exhibited similar contributions to a priori HCP_gC1_IC3. Both a priori HCP_gC1_IC3 and de novo IC3 rsfcMRI maps exhibited modular network structures comprised of high within-network connectivity and anti-correlated activity between the task-positive and task-negative networks, and resembled a canonical resting state matrix. Thus, we draw the same general conclusion as in our prior work, namely that there is some evidence that greater presence of canonical resting state networks may be associated with improved cognitive control.

It was surprising that correlations between imaging and cognitive control were significantly driven by SZ in all analyses except a priori HCP_gC1_IC7 sMRI. This may be due to a variety of factors. First, there was a graded, trend-significant inhomogeneity of variance in cognitive control performance, such that SZ had greater variance than BP, who had greater variance than HC. This may have hampered our ability to identify relationships with cognitive control within the other groups. Indeed, for all correlations between imaging and cognitive control (except HCP_gC1_IC7 rsfcMRI, which was not significant in original HCP data), the direction of trendlines were concordant for patients, suggesting the presence of a relationship that we were underpowered to detect. Second, BP inclusion was not limited to individuals with psychosis and BP endorsed significantly less history of psychotic symptoms than SZ (table 1). Importantly, the literature suggests variability in cognitive performance in BP (60) with poorer working memory and executive function performance in psychotic BP, which was also observed in the present study (table S10). Thus, future work with larger BP samples with psychosis are needed to better determine similarity to schizophrenia in multimodal correlates of cognitive control.

There are several additional limitations to the present study. First, de novo mCCA+jICA was a data-driven explorative analysis and must be interpreted with caution. However, given similarity between a priori ICs and the present results, the differences across datasets provide intriguing possibilities for further hypothesis-driven study. Second, the present cognitive control metric was not identical to the HCP metric due to study design differences, potentially limiting our ability to detect all of the same relationships identified in our prior work (see supplement). Third, while IC3 was group discriminative overall, we were unable to detect significant post-hoc differences between HC and BP, though differences between HC and SZ were significant and differences between SZ and BP were trend-level, which may be due to medications or a relatively unimpaired BP group. Fourth, it was surprising that a priori HCP_gC1_IC3 rsfcMRI did not significantly predict cognitive control in the entire psychosis sample or any of the individual subgroups. This could be due to differences in the behavioral composite across the studies or, alternatively, it is possible that sum of numerous small but subtle differences in the two maps were sufficient to reduce predictive performance. Finally, sample size may have limited our ability to detect group discrimination of a priori ICs, relationships between imaging and cognitive control given that we only had behavioral data for 30 HC, 27 BP, and 23 SZ participants. However, the de novo identification of IC3 which was 91.5% similar to HCP_gC1_IC3 suggests that this pattern is indeed a durable feature in the data and that limited variance in the cognitive control composite was the primary driver of our inability to detect some effects.

In conclusion, the present study employed multimodal methods to examine transdiagnostic alterations in cognitive control in psychosis. Two modalities from a priori ICs from the normative population significantly predicted cognitive control in psychosis for two of six modalities tested. De novo mCCA+jICA identified a group-discriminative IC that significantly correlated with cognitive control for sMRI, rsfcMRI, and tfMRI data. De novo analyses suggest joint associations between cognitive control and tfMRI contributions from the posterior frontoparietal, dorsal attention, and visual networks, sMRI contributions from the insula, medial PFC, and cingulate, and rsfcMRI contributions from canonical resting-state network organization. However, significant effects were predominantly driven by SZ with little evidence for effects in BP or HC. Given psychotic symptom heterogeneity in BP, results suggest that shared symptomatology, e.g. psychosis, may be key to identification of transdiagnostic relationships with cognitive control. Together, these results identified significant and replicable relationships across modalities and the psychosis spectrum, providing targets across modalities of neural structure and function for future research.