Demographic and clinical information for all participants are shown in Table 1. The Institutional Review Board at the Washington University School of Medicine approved all study protocols. Participants provided written, informed consent prior to participation in the study. Imaging data were collected from 33 healthy control (CN) and 27 SZ participants, aged 18–30 years recruited from the greater St. Louis area. Four participants were excluded from analysis due to incomplete data acquisition (two CN and two SZ). SZ participants were outpatients and met DSM-IV criteria for SZ, as assessed by both the Structured Clinical Interview for the DSM-IV Axis I Disorders (SCID-I) (29) and a clinical evaluation by a research psychiatrist (D.M.). All potential participants were excluded if they met DSM-IV criteria for substance dependence or severe/moderate abuse in the past 6 months, were at the time of recruitment clinically unstable (i.e., significantly sedated or unable to follow instructions) or had a history of head injury or loss of consciousness. Exclusion criteria for CN included a lifetime history of DSM-IV psychotic or mood disorders.

Prior to scanning, participants completed a computer-based cognitive assessment, the University of Pennsylvania Computerized Neurocognitive Battery (Penn-CNB) (30). Cognitive scores were determined as previously (31, 32), based on the results of working memory and attention related cognitive tasks including the Penn Abstraction, Inhibition, and Working Memory Task (AIM) (33), Penn Continuous Performance Test-Number and Letter Version (CPT) (34), and Letter-N-Back (LNB) (35). A composite working memory performance score was calculated by z-scoring the true positive rate for the LNB 1-back and 2-back conditions and summing with the z-score of the true positive rate of the working memory portion of the AIM task. An attention score was generated by z-scoring the accuracy (all correct responses/all trials) of the CPT task results.

The n-back task and procedure used has been previously described (36). Briefly, the task was practiced outside the scanner prior to beginning the scan session. Task instructions were presented and explained again in the scanner prior to the beginning of each task run. Two runs per participant were obtained, each 5:01 mins long and containing eight task blocks, split evenly between 0-back and 2-back conditions. Each block lasted 25 s and consisted of 10 2.5-s trials. Interspersed between task blocks were 15-s fixation blocks. Stimuli were presented using the E-prime (2.0.10.242) presentation software and back projected onto a screen outside of the scanner. Participants viewed this screen via a mirror attached to the head coil inside the scanner. Four types of stimuli were shown: faces, places, tools, and body parts. Task blocks within each run were evenly split between each of the four stimulus types. Participants responded on each trial whether the stimulus image matched the target image using one of two right-handed button presses for “match” or “no match.” In the 0-back condition, participants were asked to respond whether each presented stimulus matched a target image presented at the start of each 0-back block. The 2-back required participants to respond whether each presented stimulus matched the image presented two trials prior to the current image.

Participants who had <70% accuracy on the 0-back portion of a run, performed an additional run to replace the lowest performance scan.

The fMRI protocol mirrored the HCP Phase 2 task functional MRI acquisition (37). These acquisitions utilized a 32 channel head coil on a modified 3T Siemens Skyra with TR = 720 ms, TE = 33.1 ms, flip angle = 52°, BW = 2,290 Hz/Px, in-plane FOV = 208 × 180, 72 slices, 2.0-mm isotropic voxels, and a multi-band acceleration factor of 8 was used for whole brain acquisitions as a means of increasing signal to noise (38). Each task session contained two runs with a right-to-left and a left-to-right phase encoding.

This preprocessing pipeline package output minimally preprocessed 4D time series data consisting of motion correction, gradient unwarping, brain-boundary-based registration of EPI to structural T1-weighted scan, fieldmap-based EPI distortion correction, non-linear (FNIRT) registration into MNI152 space and grand-mean intensity normalization. From here, the data were projected onto a surface representation and smoothing was constrained to cortical surface and subcortical gray-matter parcels as part of a “grayordinate”-based approach (39).

Open source, HCP data analysis pipelines were used to process raw data, which utilized tools from FSL (5.0.9) and AFNI FreeSurfer (5.3.0) (39). To account for the potential influence of task-related evoked responses on correlations, we modeled task events using a general linear model regression (40–43). Separate predictors were included for each stimulus type (i.e., faces, tools, places, and body parts) crossed with each working memory load level (0-back, 2-back). To form these predictors, we convolved a canonical hemodynamic response function with block predictors for each appropriate trial, creating 0- and 2-back predictors for each stimulus type. Separately, we included a model of correct and incorrect responses by creating two predictors for all stimuli with correct or incorrect responses. Additional confound regressors included in the model were the global signal, average ventricle and white matter timecourses (calculated from data in the volume space), and six estimated parameters of motion and their first derivatives. The resulting model consisted of predictors for each stimulus by n-back level (0- and 2-back), correct responses, incorrect responses, and confound regressors described above. The residuals of this regression were then specified for the five specified networks (and their constituent parcels) using the existing 333-node parcelation scheme described by Gordon et al. (27). Briefly, the Gordon parcelation scheme involved boundary-map generation identifying transitions in resting-state correlations across the cortical surface, in a previously described dataset (27). The 333-node (parcel) assignments were grouped into functional networks using a process that involved community assignments based on the maximization of within-community random walks in a functional connectivity matrix, as well as harmonization with previously reported cortical network locations.

Residuals were temporally masked, for each n-back level, at time points corresponding to correct trials. To account for potential spurious correlations produced by high-motion frames (44, 45), time points corresponding to an observed frame-wise displacement of 0.7 mm or greater were excluded (46). 0.7 mm was chosen as the threshold to maximize the number of usable frames for analysis. Mean (SD) number of frames dropped were 0.9 (2.0) for healthy controls, and 5.0 (11.3) for SZ patients. The result was a timecourse, for each n-back condition, masked to include only correct trials and excluding high-motion time points. Pearson correlations between parcels were calculated on these masked, residualized time courses. Fully connected, group and condition averaged network matrices are shown in Figure 1.

For each network of interest, we calculated the average within-network functional connectivity, with connectivity defined as the Pearson correlations between parcels comprising each network. These averages were calculated on the fully connected, symmetric, single-subject functional connectivity matrices for each n-back condition. Networks of interest included the DMN, FPN, CON, DAN, and VAN, derived from the parcelation scheme described by Gordon et al. (27). In addition, we calculated the average between-network connectivity for each of these networks with the FPN and DMN (see Figure 2).

All statistical analyses were performed using IBM SPSS Statistics, version 24 (SPSS Inc.). To test for average between- and within-network differences across n-back condition and group, we utilized a 2 × 2 ANCOVA, with diagnostic group (control vs. SZ; between subject) and n-back condition (0-back vs. 2-back; repeated-measures) as factors. All analyses included number of frames dropped due to head-motion censoring as a covariate. For within-network tests, p-values are Bonferroni corrected for five tests (one per network tested). For all between-network tests, p-values reported are Bonferroni corrected for seven between-network connectivity tests. Based on ANCOVA results, we performed follow-up analyses to determine whether network connections that showed an interaction between diagnosis and n-back condition correlated with clinical symptom severity (i.e., for positive, disorganized, and negative symptoms) or in-scanner accuracy. We also investigated correlations of network connectivities with cognitive performance (i.e., working memory and attention) outside the scanner. Connectivity variables for correlations were either the specific connectivity at either condition, or derived as the functional connectivity difference between conditions (i.e., 2-back minus 0-back). We tested the effect of typical antipsychotic drugs on connectivity at either task condition by comparing SZ patients on typical antipsychotics to those on only atypical antipsychotic drugs using a Student’s t-test.

Table 2 shows the in-scanner working memory task accuracy of CN and SZ participants. For both 0-back and 2-back, SZ participants had lower accuracy compared to CN. The table also indicates that working memory and attention performance outside the scanner was worse in SZ participants.

Estimated marginal means by group for each within-network functional connectivity are shown in Figure 3. The FPN showed a significant main effect of n-back condition [F_(1,54) = 10.629, p = 0.01] but no main effect of diagnostic group [F_(1,54) = 0.185; p > 0.99]. The FPN, also showed a significant condition-by-diagnosis interaction [F_(1,54) = 9.824, p = 0.015], driven by decreased connectivity at the 0-back condition in SZ compared to CN participants. Unlike in CN participants, SZ participants increased within-FPN connectivity with n-back difficulty. Post hoc analyses, however, did not meet significance for group effects at 0-back [F_(1,54) = 2.028, p = 0.16] and 2-back [F_(1,54) = 0.265, p = 0.6] for within-FPN connectivity.

Within the DMN, there was no main effect of condition [F_(1,54) = 5.139, p = 0.14], diagnosis [F_(1,54) = 0.536, p > 0.99] or condition-by-group interaction [F_(1,54) = 0.713, p > 0.99].

There was a significant main effect of n-back condition for within-CON [F_(1,54) = 14.348, p = 0.002] and within-DAN [F_(1,54) = 9.337, p = 0.015] connectivity, with connectivity within both the CON and DAN increasing with n-back difficulty in both SZ and CN participants. There was no group effect or condition-by-group interaction for either network. Within-VAN connectivity showed only a trend level effect of n-back condition [F_(1,54) = 0.083, p = 0.09], and no group or condition-by-group effect.

Estimated marginal mean connectivities between either FPN or DMN and other networks are depicted in Figure 4. DMN-to-FPN functional connectivity did not show a main effect of n-back condition [F_(1,54) = 0.278, p > 0.99] or group [F_(1,54) = 0.062, p > 0.99]. However, there was a significant condition-by-group interaction [F_(1,54) = 8.041, p = 0.042]. Notably, DMN-to-FPN connectivity with n-back difficulty, decreased in CN participants and increased in SZ participants. Post hoc analyses did not meet significance for group effects at 0-back [F_(1,54) = 2.072, p = 0.16] and 2-back [F_(1,54) = 0.042, p = 0.84].

FPN functional connectivity with either CON, DAN or VAN did not show any significant effects of condition or group, and there were no condition-by-group interactions.

DMN-to-FPN functional connectivity results are described above. DMN-to-CON connectivity showed a significant main effect for n-back condition (F_(1,54) = 13.786, p = 0.003), but there was effect of group [F_(1,54) = 6.355, p = 0.11] or a significant condition-by-group interaction [F_(1,54) = 0.274, p > 0.99]. Similarly, there was a significant effect of n-back condition [F_(1,54) = 10.093, p = 0.014] for DMN-to-DAN connectivity, without group [F_(1,54) = 3.544, p = 0.46] or interaction [F_(1,54) = 0.741, p > 0.99] effects. Connectivity of the DMN with either the CON or the DAN decreased with n-back difficulty, in both SZ and CN groups.

There were no significant effects of n-back condition or group, or any condition-by-group interaction for DMN-to-VAN functional connectivity.

Table 3 shows the correlations of in-scanner accuracy of 0-back and 2-back tests with within-FPN connectivity and DMN-to-FPN connectivity, in both subject groups. Accuracy at 0-back correlated with DMN-to-FPN connectivity significantly (r = 0.41; p = 0.04) and at a trend level with within-FPN connectivity (r = 0.4; p = 0.051). Accuracy at 2-back did not correlate significantly with either FPN connectivity.

We also investigated the relationship of connectivities with working memory or attention performance outside the scanner. Across groups, significant correlations were observed for working memory performance with the within-FPN functional connectivity difference (r = −0.3; p = 0.04) and a trend level correlation with the DMN-to-FPN functional connectivity difference (r = −0.29; p = 0.06). However, there was no significant relationship when only SZ participants were analyzed. Attention performance did not correlate significantly with either the FPN or DMN-to-FPN connectivity difference, in either all participants or SZ participants.

We tested the relationship of FPN connectivity and DMN-to-FPN connectivity, with clinical symptoms (i.e., positive, disorganized, and negative symptoms). Across groups, significant correlations were observed for negative symptoms with the within-FPN functional connectivity difference (r = 0.35; p = 0.01), and a significant correlation of positive symptoms with the DMN-to-FPN functional connectivity difference (r = 0.30; 0.03).

As seen in Table 3, there were no significant relationships of any clinical symptom with FPN connectivities at 0-back or 2-back in either group.

We investigated the effects of the major classes of antipsychotics among SZ patients, by comparing FPN connectivities in those on typical antipsychotics (N = 12) and those on only atypical antipsychotics (N = 13). There were no significant effects observed for either within-FPN connectivity (0-back: F = 1.1, p = 0.3; 2-back: F = 1.4, p = 0.2) or DMN-to-FPN connectivity (0-back: F = 0.02; p = 0.9; 2-back: F = 1.3; p = 0.3).

Our study found that across subjects from both groups, functional connectivity within the frontoparietal (FPN), cingulo-opercular (CON), and dorsal attention (DAN) network increased with cognitive load. However, increased connectivity within the FPN was driven by the SZ group, with control participants showing similar connectivity during the 0-back and 2-back tests. This was in contrast to within the CON and DAN, where connectivity in both control and SZ participants increased with cognitive load. The CON and DAN also showed increased decoupling from the task-negative default mode network (DMN) with increased cognitive load, suggesting that the CON and DAN are involved in task processing. Both the CON and DAN have known roles in cognition. The CON includes brain regions that include the anterior insula/operculum and dorsal anterior cingulate cortex. Unlike the FPN that has been ascribed roles in the initiation and modulation of cognitive control abilities, the CON is thought to facilitate the maintenance of task-relevant goals and the incorporation of error information to adjust behavior (47, 48), and are co-activated during cognitive control tasks (49, 50). On the other hand, the DAN is an intrinsic network which comprises of regions of the intraparietal and superior frontal cortices, and has been reported to be involved in top-down, goal-directed selection deployment of attention (51). Our findings indicate that carrying out a working memory task requires increased neural processing within these three brain networks (FPN, CON, and DAN). The results further suggest that increased task-related neural processing within the FPN may be more important for those with psychopathology, than for healthy individuals, possibly due to decreased efficiency of the other cognitive networks.

Our results are supported by reports from other studies. While negative findings have been reported (19), the CON has been found to flexibly link with other brain networks during cognitive tasks, and to be involved in a broad range of cognitive processes (52). Functional connectivity or increased coherence within the CON have also been reported to increase with tasks requiring more tonic alertness (53), word recognition (54), visuospatial attention and episodic memory, memory encoding of speech in noise (55) and slow reveal task (56). In addition, general cognitive ability has been linked to the local and global efficiency of the CON in both control and SZ participants, and those with psychotic-like experiences (57, 58). It has been suggested that the CON plays a more downstream role in cognitive control, compared to the FPN, possibly associated with output gating of memory. For example, Wallis et al. reported that when spatial cues occurred indicating the relevant item in a working memory array, the FPN was activated following the cue (59). However, when cues occurred during the maintenance period, FPN activation was transient and succeeded by CON activation. In our study, the association of increased within-DAN connectivity with cognitive load was consistent with the role of the DAN in controlling the spatial orientation of attention. Activation patterns of specific DAN regions have been found to predict both verbal and visual working memory load (60). Also, increased DAN activation has also been reported with increasing short-term memory load, in contrast to the ventral attention network activity, which was decreased (61).

We found two significant connectivity differences between SZ patients and control participants in our study. First, there was decreased within-FPN connectivity during the 0-back condition in SZ patients, compared to controls; however, during the 2-back condition, FPN connectivity was similar across groups. Second, connectivity between the FPN and the default mode network (DMN) decreased with greater working memory load in healthy participants, but increased in SZ patients. The connectivity of other cognitive networks studied tended to be lower in SZ than controls, consistent with a generalized hypoconnectivity that has been widely described (62, 63), although these group differences did not meet statistical significance. These findings suggest that in SZ patients, the FPN is hypoconnected during with a lower cognitive load, but patients are capable of achieving the needed connectivity to process more complex tasks. At the same time, in patients, the FPN does not uncouple from the DMN with high cognitive load as it does in healthy participants; rather, they become more strongly connected. While speculative, such a finding may suggest that in SZ patients, low connectedness of the FPN at low cognitive load or at rest, may be overcome by recruitment of compensatory pathways with increasing task complexity.

Our study results support the role of FPN cortical regions, including the DLPFC, in SZ. Meta-analytic and review studies of working memory have generally found reduced activation of the DLPFC in most studies of SZ patients compared to healthy controls (11, 12, 64). Group differences in other areas are also seen, including increased activation of anterior cingulate and the left frontal pole (11). This have been suggested to represent a dysfunctional brain network supporting working memory in SZ, which involves impaired attention control by the DLPFC, with associated increases in error monitoring involving the anterior cingulate (11), functioning patients also showed increased activation in some prefrontal regions (65). Results of studies involving FPN functional connectivity in SZ have been variable, and may be related to methodological differences and biology heterogeneity of the disorder. Nielsen et al. reported that the increased FPN connectivity modulated by working memory, was decreased in first-episode SZ patients (66). Repovs and Barch however, found increased FPN connectivity in both control and SZ populations with increased working memory load (19). Similar to our findings, Eryilmaz et al. reported normalization of FPN functional connectivity during a working memory task, despite impaired resting-state connectivity (24). These authors further found that working memory deficits were related to limbic and thalamic dysconnectivity and altered connectivity between FPN and the thalamus.

The underlying neurobiology of decreased functional connectivity in SZ with working memory may be related to decreased integrity of key white matter tracts within the brain. Functional connectivity is not isomorphic with structural connectivity (at least in terms of single synapse connections) and thus one cannot directly interpret alterations in functional connectivity as reflecting alterations in structural connectivity. White matter abnormalities in SZ have, however, been well described from diffusion studies (67, 68), including decreased white matter integrity in the superior longitudinal fasciculi, the main frontal-parietal white matter connection. Superior longitudinal fasciculi integrity has also been related to working memory performance in SZ (69). Such findings are consistent with an abnormal neurodevelopment in SZ, with regional synaptic deficit and intact long-distance connections (70, 71). Our findings point to a need to more directly examine the degree to which changes in functional connectivity are reflective of changes in white matter integrity in SZ.

Despite the relevance of the results presented, our study does have some limitations that may influence our findings. First, as with other functional connectivity studies, potential motion-related effects may have affected data quality. While our motion-censoring is expected to minimize such influences, motion-related differences across groups may have confounded our results. The amount of data available could also have influenced the reliability of our connectivity results (72–74). While the data volume used in our study is comparable to other task-related connectivity studies in SZ, it is possible that the available data after motion-censoring may have become insufficient to accurately estimate functional connectivity. Second, the network parcelation scheme used in our study was derived from resting-state data. While the spatial locations of task-based networks in the brain largely overlap with that of resting-state networks, differences may also be present. Some differences in the spatial location of networks are also expected across working memory conditions. Thus, no parcelation scheme can be universally applicable. This should be considered in interpreting results; as cortical parcels may not fully delineate each specified network in each individual. Potential inaccuracies in network parcels would, however, be expected to affect individuals in each group similarly. Future studies using alternative parcelation schemes will be useful to validate our findings. Thirdly, our study was not designed to investigate differences in regional blood oxygen level dependent activation during the working memory task between SZ and control groups. Such differences, however, would present a more complete picture of the brain regions recruited during task performance, and could influence the interpretation of the connectivity findings. For instance, decreased activation of DLPFC (a major component of the FPN) in SZ during working memory, as seen in several studies (11, 12, 64), may contribute to the observed decreased FPN connectivity in our study. Future investigations would, therefore, benefit from including information on brain activation alongside connectivity results. Finally, the role of antipsychotic treatment of SZ also may influence neuroimaging findings during working memory tasks. For example, Wolf et al. (75) found improved frontotemporal function after several weeks of antipsychotic treatment, together with improved working memory performance in SZ patients. Others report normalization of frontal cortical activation only specific antipsychotic drugs, including aripiprazole (76, 77), Quetiapine (78) or risperidone (79). Normalizing of functional dysconnectivity in SZ participants have also been reported (80, 81). Thus, it is likely that functional findings in SZ patients are underestimated, and medications likely confound the understanding of the pathophysiology of working memory deficits in SZ patients. Studies involving medication naïve patients, unmedicated high-risk individuals and longitudinal treatment studies would, therefore, shed light on specific medication effects on task-based connectivity.