Why are thalamo-cortical networks central to sustained attention? All incoming sensory inputs (except olfactory) pass through the thalamus, and the structure has been viewed as a “relay station” and “gateway” to subcortical and cortical regions (27, 28). Anatomically, several distinct thalamic sub-nuclei synapse onto pertinent attention-related brain cortices (29, 30) including the dorsal anterior cingulate cortex (dACC), dPFC, primary visual cortex (V1), and the superior parietal cortex (SPC). In this regard, ascending relays from the thalamus assume importance.

Thalamic transmission of sensory information can be subdivided into first-order and higher-order relays. In first-order relays, thalamic nuclei convey the initial sensory inputs from sensory organs (i.e., eyes) to the cortex. For instance, retinal inputs travel to the lateral geniculate nucleus (LGN) of the thalamus before being relayed to the V1. In higher-order relays, thalamic nuclei receive descending inputs from cortical regions before relaying and controlling the flow of information back to the cortex (31–33). These descending (cortico-thalamic) and ascending (thalamo-cortical) pathways cumulatively underpin communication in brain networks (34), and suggest an important role for thalamic engagement and dynamic regulation of attentional, decision-making, executive control processes in the cortex (35).

Extant studies suggest that networks associated with sustained attention are subsets of the overall thalamo-cortical system (36). The reliably identified constituent regions of the attention network include frontal-striatal regions such as the dACC, the dPFC and the basal ganglia (BG)—all important for executive or supervisory mechanisms of attention (37, 38). The dACC functions as a “control center” organizing and recruiting executive brain regions like the dPFC (39–42). Additionally, the dACC shares bilateral connections to the dPFC, with the latter region known to be involved in the critical functioning of vigilance, executive control, working memory, and selective and divided attention (43). The BG are a varied group of subcortical nuclei involved in the functioning of eye movement (44), working memory (45), decision making (46), and in the organization of motivations that lead to the execution of goal-directed behaviors (e.g., organizing task-driven responses) (47). In addition to these “heteromodal” regions, multiple regions have been associated with spatial attention and orientation and sensory gating including the SPC, involved in the allocation of spatial attention, which is innervated by the V1 via the dorsal stream (48, 49) as well as direct thalamic connections (27, 40, 50).

Is there evidence of disordered brain function in OCD patients during attention processing? Utilizing a cued task-switching paradigm, Gu et al. found significantly reduced levels of brain activation in cortical (dPFC, dACC, and orbitofrontal cortex) and subcortical (caudate nucleus) regions, replicating prior studies on impaired executive functioning and cognitive flexibility (51); the task-switching paradigm required subjects to disengage attention from a prior task, resolve interference, and configure to a new cognitive task-set (52). Attention-related mechanisms in OCD have also been probed using the Stroop task, that measures functions related to selective attention, interference control, and cognitive flexibility (53). Exploring inhibition control in OCD, Page et al., used three different motor and cognitive control tasks (Go/No-go, motor Stroop, Switch tasks) to assess dysfunctional brain activations. The authors found reduced activation in OCD in multiple cortical (DLPFC, OFC), striatal, and thalamic regions (54). In a motor task with substantial attention demands, Meram and colleagues demonstrated disordered network-based mechanisms of frontal control in OCD (55), a pattern that replicates prior studies using working memory (56). Finally, Woolley et al. used an inhibitory (‘stop’) task, to show reduced activation of frontal, striatal and thalamic regions in pediatric OCD, evidence of dysregulation of fronto-striatal-thalamic (FST) regions during motor inhibition (57). Given that thalamic inputs are highly salient in driving cortical networks, and given the loss of cortical network tone in OCD, Del Casale and colleagues have suggested that altered thalamic gating in OCD, results in alterations in cortical activity in the disorder, leading to the subsequent emergence of core clinical characteristics including intrusive thoughts and repetitive and ritualistic behaviors (58). Is there any evidence of altered effective connectivity of ascending and/or descending relays of the thalamo-cortical network in OCD using task-based fMRI? To our knowledge this specific issue has not been addressed with DCM.

What insights in OCD have we gained from functional connectivity studies? In the past 20 years, task-positive and resting-state functional connectivity studies have expanded our understanding of brain circuit abnormalities in OCD (59). Chiefly, our understanding of dysfunctional cortical, striatal, and thalamic (CSTC) circuits and resting-state networks [i.e., default-mode network (DMN) circuits] has been greatly expanded. Task-based fMRI studies have revealed abnormal patterns of hypo- and hyper-activation in domains such as executive functioning, emotional processing, and symptom provocation (59) with commonly implicated brain regions including: the prefrontal cortex (PFC) (54, 60), BG and thalamus (61), orbitofrontal cortex (OFC) (62), dACC (63), amygdala (64).

What is effective connectivity and how can it expand our understanding of dysfunctional brain networks in OCD? While functional connectivity describes statistical dependencies between times series, effective connectivity rests on a mechanistic model of the neurophysiological processes that generated the observed data (65). Thus, it is plausible to arrive at an understanding of the casual interactions one neuronal unit exerts over another; the elucidation of these causal influences via effective connectivity studies may prove pivotal in the characterization of aberrant pathways in OCD. Indeed, attempts to understand directed connectivity have been conducted using approaches such as psychophysiological interaction (PPI) (66) or structural equation modeling (SEM) (67). However, these approaches rely on observed fMRI signals (65). We relied on DCM as we were specifically motivated to recover causal effects between brain regions in OCD patients and controls.

Dynamic causal modeling allows for the interpretation of causal interactions between hidden state variables (9, 26) by modeling the brain as a bilinear input-output system. In this framework, inputs to the system are the experimental conditions and stimuli, and the outputs from the system are fMRI measured hemodynamic responses. DCM evaluates and assesses competing neurobiologically plausible models to determine any model’s posterior evidence for predicting the observed fMRI data. The competing models constitute a motivated a priori hypothesis space. Posterior probabilities are used to weight parameter estimates and differences between estimated parameters can be investigated using standard inferential statistics (26, 68). The following state differential equation is implemented to measure changes in neural responses:

Where, “A” represents the task-independent intrinsic (endogenous) coupling that exists between brain regions of interest, “B^(j)” represents the task-dependent modulation of intrinsic connections via experimental manipulations, and variable “C” represents the sensorimotor driving inputs on cortical regions. We surmise that characterizing ascending thalamic-cortical relays in OCD is essential and that DCM applied to fMRI data collected during sustained attention, provides that ideal combination of method and task (4, 22, 69). Therefore, DCM allowed us to specifically assess the effects of task conditions on the ascending and descending relay and how (and why) any contextual modulation on thalamo-cortical relays might differ in OCD participants relative to controls.

Thirty OCD subjects (Age: 10.1–21.9 years; Mean Age: 16.2 years.; 10 males) and 32 healthy controls (Age: 10.9–21.1 years; Mean Age: 16.7 years; 12 males) provided informed consent or assent to participate in the fMRI study. For assenting participants, parental consent was also obtained. Participants’ IQ (89 < IQ < 143; Mean: 109.8; SD: 11.3) was assessed using either the Wechsler Intelligence Scale for Children (WISC) or the Wechsler Adult Intelligence Scale (WAIS). Healthy controls were evaluated for, and cleared of, any psychiatric illness. OCD participants were evaluated by a psychiatrist. Demographic and clinical characteristics of study participants can be viewed in Table 1.

The Schedule for Obsessive-Compulsive and Other Behavioral Syndromes and Schedule for Schizophrenia and Affective Disorders for School-Aged Children-Present and Lifetime Version were used to interview participants and their parents (70, 71). Furthermore, a modified version of the Children’s Yale-Brown Obsessive-Compulsive Disorder Scale (CY-BOCS) was used to measure the lifetime (maximum) and current severity of OCD participants (72, 73). Using DSM-5 criteria, clinicians independently confirmed the lifetime and current Axis I diagnosis. Participants with the following medical or psychiatric criteria were excluded: lifetime history of schizophrenia, bipolar disorder, substance abuse/dependence, anorexia nervosa, bulimia nervosa, epilepsy, head injuries resulting in chronic loss of consciousness, Huntington’s disease, dyskinesia, autism, IQ ≤ 80 or ≥ 15 on the lifetime version of the Social Communication Questionnaire (74, 75). The Wayne State University School of Medicine (WSUSOM) and the UM Human Investigation Committee approved our study.

All participants were administered a modified version of the Continuous Performance Task, Identical Pairs (CPT-IP) (Figure 1) deployed in previous studies (14, 22, 76), and required subjects to detect repeated numbers within a rapidly presented sequence. In separate blocks, two- or three-digit numbers were presented (50 ms, 250 ms SOA) over extended block lengths (120 s). The extended blocks were specifically designed to induce sustained (as opposed to transient) attention. Titrating demand in sustained attention has typically been implemented by changing the onsets between successively presented stimuli (77) or by experimentally manipulating stimulus characteristics (as undertaken here). Previous studies that have analyzed behavioral performance have demonstrated that the numerosity of the targets does evoke a lucid parameterization of task difficulty associated with attention-demand (Low vs. High); detection sensitivity (d’) for targets (repeated numbers) during three-digit epochs is lower than detection sensitivity for targets during two-digit epochs (14). Thus, attention demand constitutes a potentially critical modulator of the effective connectivity of ascending and/or descending thalamo-cortical relays (22, 78). In addition, 20 s rest epochs interspersed the experiment allowing participants some recovery between task-active epochs.

480 stimuli were shown in each block (25% were targets). To diminish figure-ground contrast (and increase attention commitments), numbers were presented in white numerical text (RGB: 255,255,255) against a grey background (RGB: 225,225,225) (79). Furthermore, to prevent selection of consecutive digits based on low-level features such as the lack of a flicker between repeated numbers, participants were presented with alternating fonts for successive digits in the sequence (78) thus ensuring a constant level of flicker throughout each task-active epoch.

Gradient echo echo-planar imaging (EPI) fMRI data were obtained using a 3 Tesla Siemens Verio system (12-channel volume coil head) at the Vaitkevicius Magnetic Resonance Center. The following parameters were used for the T₂* fMRI acquisition: TR: 2.6 s, echo time; TE: 29 ms; matrix dimensions: 128 × 128; voxel dimensions: 2 mm × 2 mm × 2 mm; field of view (FOV): 256 mm × 256 mm; 36 axial slices. The 3D matrix (128 × 128 × 36) provided very high resolution coverage of the cortex for a conventional EPI T₂* sequence. The 36 axial slices were positioned parallel to the anterior commissure/posterior commissure (AC-PC) line and provided comprehensive coverage of the cerebrum. A T₁-weighted structural image was obtained for preprocessing and co-registration using a 3D Magnetization Prepared Rapid Gradient Echo (MPRAGE) sequence with the following parameters: TR: 2,200 ms; TI: 778 ms; TE: 3 ms; FOV: 256 mm × 256 mm; matrix dimensions: 256 × 256; flip-angle: 13°; 256 axial slices, 1.0 mm thickness. All scans were reviewed by a neuroradiologist to exclude clinically significant irregularities.

Image processing was undertaken in SPM12¹ using established methods for temporal (slice timing correction) and spatial preprocessing. First, EPI images were manually oriented to the AC-PC line with the reorientation vector applied across the EPI image set, realigned to a reference image to correct for head movement, and co-registered to the anatomical high resolution T₁ image. The T₁ image was normalized to the Montreal Neurological Institute (MNI) template, with the resultant deformations applied to the co-registered EPI images. Low frequency components were removed with a lenient filter (1/256 as opposed to 1/128) and this filter accommodated longer block lengths (120 s) that were used to induce sustained, as opposed to transient attention (14, 22). Images were resliced to (2 mm³) and smoothed (8 mm FWHM). At the first level, epochs were modeled with boxcar stimulus functions convolved with a canonical hemodynamic response function to form regressors of interest, with the six motion parameters (3 for translation and 3 for rotation) from the co-registration modeled as covariates of no interest. We did not model phasic or event-related responses to targets because we were interested in the responses specifically associated with sustained attention. Images exceeding 4 mm of movement (<1% of all images) were excised from analyses. This approach to image processing and modeling is consistent with recent publications (55, 68, 80–84).

The regressors of interest representing each of the levels of demand (Low:2-digit; High:3-digit) were submitted to a second-level mixed random effects model (85), with condition and group respectively modeled as repeated and independent factors. To identify an appropriate conjunction of co-activated clusters across groups and conditions, we used Nichols et al. version of the minimum-inference statistic (86)(minimum statistic compared to the conjunction null). These co-activations (depicted in Figure 2) conform to activation profiles identified in previous studies using identical and related paradigms (14, 22, 87). Clearly identifiable peaks are noted in regions including the V1, SPC, thalamus, BG, dACC, and the dPFC. These regions of interest provided a thalamo-cortical network of interest wherein inter-group differences in effective connectivity could be investigated without the confound of activation-differences between groups (41).

A total of 162 competing network models were configured in the hypothesis/model space. Hypotheses of interest heavily focused on evaluating the contextual modulation of specific pathways during sustained attention with varying levels of demand. Across models eight hypothesis-neutral intrinsic connections were fixed, depicted in Figure 1C by a solid black (curved) line. Four cortical-thalamic intrinsic connections were permuted across the model space represented by solid white lines (Figure 1D): (1) Thal → dPFC; (2) dPFC → Thal; (3) Thal → dACC; and (4) dACC → Thal. We tested whether these four cortical-thalamic connections when present, were contextually modulated by the task (represented by black circles in Figure 1E). Finally, the dPFC → BG connection was permuted across models and when present we tested if it was contextually modulated by the task. This 162-model space was therefore a combination of (a) three possible contextual modulation effects on four permuted intrinsic connections plus, (b) two possible effects on the dPFC → BG connection (3⁴ × 2 = 162 models).

Across each the 62 participants, time series were extracted using spheres (5 mm radius) centered on the peak of the “effects of interest” F-contrast (p < 0.05, adjusted for “effects of no interest”) within each region. These time series were submitted to model estimation, and the cumulative estimated model space (10,044 models, i.e., 162 models across 62 participants) was then submitted to random effects (RFX) Bayesian model selection (BMS), with an initial goal of determining the most likely generative model in each of the OCD and HC groups (Figure 3). A variational Bayes method using the RFX procedure was employed to estimate the posterior probabilities of competing models, resulting in posterior likelihoods associated with each model in the space in each group. RFX is an ideal approach for identifying model likelihood under conditions of sensorimotor or cognitive tasks (14, 26, 41).

Parameter estimation (for values associated with intrinsic connectivity and contextual modulation) was conducted using Bayesian model averaging (BMA) which generates a weighted average of the parameter estimates of each model within a group, where the weight is determined by the posterior probability of each individual model. BMA is well-optimized for reliably estimating parameter values in relatively large model spaces (88), and in a human psychological context has been hypothesized as being a generally rational approach to decision making (89).

A perusal of the exceedance probabilities (relative likelihood) for each of the 162 competing models in Figure 3 is illustrative. In each group, no single model was characterized by substantially larger likelihood relative to its competitors and the likeliest generative models in the OCD and the HC groups differed slightly. This difference suggests that the effective network structure of task implementation in OCD differed from HC, but the relative proximity of posterior likelihoods motivate the subsequent reliance on BMA for inference (26).

The statistical significance of intrinsic connectivity parameters was comprehensively evaluated using a combination of one-sample t-tests (to identify pathways where parameters in either group were significantly different from zero), and independent samples t-tests (to identify pathways with significant inter-group differences in intrinsic connectivity). For ease of access, the single full matrix in each group is decomposed into a) no significant effects (top row), significantly different (from zero) connectivity parameters (middle row) and significant inter-group differences (bottom row) in Figure 4. Each asymmetric connectivity matrix represents intrinsic connectivity parameters from a source region (column) to a target region (row) with the colors calibrated by the estimated mean value (see color bar). One-sample t-tests indicated that in both OCD and HC groups, four intrinsic connectivity parameters were significantly different from zero: (1) V1 → Thal, (2) Thal → dACC, (3) Thal → dPFC, and (4) V1 → SPC (Φ in Figure 4). Two other intrinsic connectivity parameters were also significantly different from zero, but only in the OCD group: (1) BG → dPFC and (2) dACC → Thal (†in Figure 4). Finally, the results of the independent samples t-tests revealed significant differences in the parameters for the V1 → SPC pathway, indicating that V1 → SPC connectivity was lower in OCD. Statistical information is provided in Table 2.

BMA was used to weight parameters over the 162 generative model architectures based on the posterior probability of models across a group. Thus, the likelihood of each hypothesis is represented in a quantitative estimate of the parameters for variables associated with that hypothesis. For an initial assessment, histograms were generated from extracting 2 × 10⁴ BMA samples in Occam’s window. These were used to highlight speculative trends associated with within or between group parameter estimates of contextual modulation of the relevant pathways by each of the attention conditions. These histograms are presented in Figures 5, 6. In each figure, data from the OCD group are represented in red, while data from the HC group are in blue. Data from the more demanding attention condition are represented by opaque colors, while data from the less demanding condition in transparent colors. Figure 5 represents ascending and descending fronto-thalamic pathways and is organized in four panels: each column represents pathways involving either the dACC or the dPFC, and each row represents the ascending (from the thalamus) or descending (to the thalamus) pathway. In each panel, histograms are presented twice; initially to depict within group trends in each of the task conditions (top row) and then to depict between group trends in each of the conditions (bottom row).

Figure 5A (top left panel) depicts two interesting trends associated with the ascending Thal → dACC pathway: (1) As seen, within the HC (but not OCD) group, the parameter histogram associated with high demand is shifted to the right. This suggests that in healthy controls, an increase in attention demand effects an increase in connectivity of this ascending pathway; (2) Between groups, the histogram for HC is shifted to the left of the OCD group. This effect is a between-group extension of the previously noted within-group effect. In Figure 5B (top right panel), two interesting trends characterize the Thal → dPFC pathway: (1) Within OCD, the histogram of values for the high demand attention condition is shifted to the right; (2) When inspecting data between groups, we observe that for both the low and the high demand conditions, the parameter histograms in HC are shifted to the right of those in OCD. This effect suggests that there is a general increase in the contextual modulation of this ascending relay in HC compared to OCD (or a decrease in OCD compared to HC). In Figure 5C (bottom left), the histograms for the dACC → Thal descending relay suggests that the high demand condition induces an overall larger, positive contextual modulation in the sample means in HC. Finally, in Figure 5D (bottom right panel), parameter histograms for the dPFC → Thal descending relay suggest that for the low demand condition, the parameter histogram for HC is shifted to the right of that for the OCD group. However, for the high demand condition, we observed a reversal of this effect. Figure 6 depicts parameter histograms for the dPFC → BG pathway revealing that for both low and high levels of demand, the parameter histograms for the HC group are shifted to the right of the OCD group.

The effects depicted in Figures 5, 6 are suggestive, but motivated more robust parametric analyses in the service of inference. Accordingly, parameter estimates across participants and groups were submitted to independent sample t-tests to identify pathways with significant inter-group differences. Figure 7 (top) uses a heat map to visualize parameter means in each of the OCD and HC groups. Data are separated into each of the low and high demand conditions. Cells marked by symbols constitute a member of a pair with significant differences in contextual modulation (see Table 3 for details). Two effects were significant, both observed on the ascending Thal → dPFC relay. Both the low and the high attention demand conditions induced greater contextual modulation in HC compared to OCD. These significant effects were pulled from the heat maps with the group means visualized in the bar graphs in Figure 7 (bottom) (±sem).

The dorsal visual pathway has long been presumed to play a key role in the transmission of spatial aspects of visual processing in the service of higher sensori-motor function (49). The V1 receives visual sensory information via the lateral geniculate nucleus (LGN) of the thalamus before sending projections to subsequent brain regions (95). Likewise, the SPC is also innervated by thalamic connections (27, 40, 50) and functions in the allocation of spatial attention. How does visual information interface between occipital and parietal cortices? As noted, current understanding rests on the two-streams (dorsal and ventral) as being the purported mechanism of action (96). The dorsal visual stream (i.e., V1→ SPC) functions in computing spatial relationships and orienting of visuospatial attention, often referred to as the “where” pathway (48, 97, 98). In contrast, the ventral visual stream functions in the recognition of objects and faces, often described as the “what” pathway (48, 97, 98).

Evidence from in vivo MRI studies tends to be consistent with these theoretical motivations. Resting state functional connectivity studies have shown that connectivity between the V1 and superior parietal lobe (SPL) is predictive of visual search efficiency which relies heavily on the spatial components of vision (99). Thus, visual search tasks induce greater engagement of brain regions that are part of the dorsal visual stream including the occipital cortex, the SPL and occipital cortex, and beyond including the dACC, and inferior frontal cortex. Moreover, participants who are more effective in visual search show stronger functional connectivity between the SPL and visual cortex. Thus, a finding of weaker intrinsic connectivity of the V1 → SPC pathway implies that in OCD there may be reduced ability of the dorsal visual stream to effectively transmit spatial information from the V1. Is there evidence for visuo-spatial processing deficits in OCD?

Indeed, existing studies have frequently described problems with visuo-spatial processing in OCD patients (100, 101). For example, poor fine-motor and visuospatial skills in pediatric OCD appear to predict the presence of the disorder well into adulthood (102). In fact, deficits in visual memory, especially when complex visual stimuli are involved, is one of the most commonly reported dysfunctional cognitive domains in OCD (103–105). One meta-analysis of cognitive functioning in OCD showed that OCD patients experience impairments in visuospatial memory, visual organizational skill, and executive functions such as planning ability (105). Other studies on OCD have also shown deficits in spatial working memory and verbal/nonverbal recall (104, 106, 107). A few early studies on youth with Tourette’s syndrome and/or ADHD disorders revealed slowed reaction times on continuous performance tasks, implying a reduced ability of subjects to remain vigilant compared to controls (108, 109). Employing numerous neurocognitive tasks (e.g., Finger Windows task), Chang et al showed spatial attention deficits in OCD youth (110); this result was consistent with prior adult OCD studies on memory, spatial attention, and other executive functions (106, 111, 112). Our observations of reduced intrinsic connectivity of the V1→ SPC are indeed consistent with these results.

Sensory inputs and outputs to the thalamic nuclei can be subdivided into first-order relays and higher-order relays. In first-order relays, thalamic nuclei receive input from sensory organs (i.e., the eye) before relaying information to the cortex. For instance, visual information from the retina is relayed from the LGN of thalamus to the V1 before further processing of visual information can occur. In higher-order relays, thalamic nuclei receive descending inputs from cortical regions before relaying and controlling the flow of information back to the cortex, forming cortico-thalamo-cortical pathways (31–33). Studies on the lateral pulvinar nucleus’ role in cortical processing of visual information further supports the expanded modulatory role of some thalamic nuclei; removing lateral pulvinar input to the V1 nearly extinguishes “visual responses in the primary visual cortex,” suggesting that the pulvinar nucleus plays a critical and integral role in the functioning of the visual cortex (31). Due to relaying more “processed” inputs flowing back to the cortex, higher-order relays have been viewed as taking a more active and modulatory role in routing inputs (113).

Because groups of thalamic nuclei are involved in relaying and controlling the flow of sensory and motor information to the cortex, the thalamus has been dubbed the “relay station” or “gatekeeper” of cortical-bound information (27, 28, 114, 115). Thalamic inputs to the dPFC assume particular importance, given this heteromodal structure’s role in a multiplicity of processing domains including working memory, selective, divided, and sustained attention, and executive function (43, 116). Far from being a passive relay in the bottom-up gating of attentional (and other sensory) information to the cortex, the thalamus is now thought to exert more constructive influences on the dynamic upward flow of sensory information. In addition to a vast array of cellular studies (117–120), this framework has also been confirmed by DCM studies showing that ascending relays from the thalamus to the dPFC are contextually modulated by sustained attention (22). Such DCM based studies affirm the importance of ascending relays in the functional innervation of cortical networks at the macroscopic scale. Our observations of reduced modulation of this ascending relay in OCD is suggestive of a contextually induced thalamo-cortical deficit in network function. What might the impact of such a deficit be? Before addressing this question, we briefly review extant understanding of the functional role of thalamo-cortical relays.

Both ascending and descending thalamo-cortical relays are necessary to balance the flow of ascending sensory inputs to the cortex against the descending modulation of these inputs. Ascending relays regulate the flow of sensory information to the cortex, whereas descending relays (dPFC → Thal) can directly excite or inhibit thalamic relay neurons via GABAergic neurons of the thalamic reticular nucleus (TRN), thus dynamically shaping the ascending inputs (121–123). The TRN provides inhibitory innervation to fellow thalamic nuclei while modulating the flow of sensory information to the cortex, as well as “influencing ongoing cortical activity by modulating cortico-thalamo-cortical transmission” (124, 125). As early as 1984, Francis Crick hypothesized an attentional role for the TRN, suggesting “if the thalamus is the gateway to the cortex, the reticular complex might be described as the guardian of the gateway”(126). Later, lesion studies on the TRN have shown impairments in the cognitive domain of attention (127, 128). Moreover, attention modulates visual processing via a reciprocal relationship between increasing LGN responses and decreasing neuronal TRN responses (30), consistent with Crick’s proposals. Thus, a vast connectivity network and polysynaptic receptors allowing tonic and burst modes of firing (129), regulate how the intrinsic circuitry dynamically modifies cortico-thalamic synchronization. This synchronization in turn is seen as a major functional basis for multiple cognitive functions (130). Unsurprisingly, a loss of this functional synchronization, particularly between cortical, striatal and thalamic circuits has been proposed to underlie OCD (131, 132) and has been documented in studies of resting state and task-based functional connectivity (55, 133, 134). Moreover, numerous studies implicate a broad range of neurocognitive deficits in attention, working memory, inhibition, and executive functions deficits in OCD (135). These have been observed during attention-related task paradigms such as the Trail-Making and Stroop Tasks and measurements of attentional set shifting, reversal learning, and cued tasked switching (136, 137). OCD patients also show hypo-activity in the dPFC as well as other PFC structures such as the OFC during reversal learning and task switching paradigms (51, 62, 138). From these results, Ahmari & Rauch have suggested that OCD subjects performing tasks with changing conditions “display less recruitment of multiple PFC structures” compared to controls. Furthermore, studies using the Tower of London task, a test which assesses goal-directed planning have also found hypo activation and slowed reaction times of the dPFC, suggesting a decrease in “dorsal prefrontal-striatal responsiveness” (138, 139).

Evidence of a loss of contextual modulation of the ascending relay falls firmly within this emerging framework. As we have noted, effective connectivity methods provide particular insight on the “causative” bases of inter-network effects (4, 9). They offer the possibility of identifying task-induced specificity in evoked network deficits. Such evidence is informative about how disease-related traits are manifested in functional brain states that are themselves transient and frequently task specific (91, 140). Such an approach may illuminate understanding of the mechanistic basis for pathology in neuropsychiatric conditions (2). We suggest that the reduced modulation of the ascending relay regardless of attention demand (Figure 7) reflects a basic loss of network tone evoked by sustained attention (22). It is plausible that this deficit reflects a relative loss of cortical innervation in OCD that subsequently impairs CST function (56), and that it provides functional evidence for increasingly clear impairments of thalamic structure in OCD (141). While circuit hyper-activity may be a trait emergent aspect of OCD (142), this may impact state-related network interactions that underlie basic processes such as attention. Thus, these deficits may affect how sensory information propagates to the cortex, where maladaptive inputs impact the fidelity of cortical network function in the disorder. This loss of network fidelity leads to a degradation of internal signals in the OCD brain, results in the emergence of erroneous perceptions (“something is wrong”), and eventually drives the anxieties that are characteristic of the disorder (143) and that drive trait-related hyperactivity.

The current study was partly motivated by prior investigations on the contextual modulation of cortico-thalamic pathways by attention demand (22), and which utilized a similar approach albeit only in healthy adolescents. That study used a winner-take-all approach from Bayesian model selection (26, 41), subsequently using Bayesian Parameter Averaging to estimate connectivity parameters from the likeliest generative model architecture. By comparison, BMA as employed in the current investigation is used to weight connectivity parameter estimates across all models, based on each model’s posterior probability. BMA is suitable for estimating parameters values for large model spaces (88), avoids uncertainties of model selection and mitigates any brittleness associated with parameter estimation (144). These differences in DCM-related methodology, as well as the focus on inter-group comparisons (as opposed to intra-group effects) may account for some differences between the sets of results from the two studies. For instance, a critical difference was that Jagtap and Diwadkar (22) noted differential effects of demand on the contextual modulation of the descending dPFC → Thal pathway (High Demand > Low Demand), suggestive of the role of frontally driven attention gain on the thalamus when demand for attention-resources was high. This effect was not replicated in the current study (see Figure 5), though our results on the ascending relay (significant contextual modulation for both low and high attention demand) in healthy controls (Figures 5, 7) was consistent with observations from the previous work.

Effective connectivity studies are beginning to illuminate understanding about casual brain network interactions in OCD. For example, using a reward learning task, Alves-Pinto et al., showed alterations in intrinsic connectivity and contextual modulation of fronto-striatal circuits, mainly from the ventromedial PFC (vmPFC) to the left and right orbitofrontal cortex (OFC) (145). In another study, Schlösser and colleagues used a Stroop color-word task to elicit aberrant effective connectivity pathways between OCD and HC in the fronto-cingulate system, consistent with an over-active error control system during decision-making observed in OCD (13). Han et al., showed that negative emotional distraction (working memory task) induced significantly reduced contextual modulation in the dorsolateral PFC (DLPFC) to OFC pathway in OCD compared to HC (146). Finally, van Velzen and colleagues have documented altered directionality of the connectivity from the amygdala to the OFC in OCD participants during a stop signal task (12).

Understanding the relationship between behavior and brain network dynamics is perhaps the central challenge in human neuroscience (147) and the thalamo-cortical network appears to play a central role in mediating this relationship (148). Our focus on thalamo-cortical networks in OCD (1) reflects understandable interest in the role of the thalamus in the disorder (149), and (2) acknowledges the value of methods like DCM in revealing task-induced “mechanistic” deficits in brain networks (68). However, several limitations to our intriguing results should be acknowledged. Firstly, task-based fMRI results provide crucial information about demand-induced modulation of network connectivity and its relevance for psychiatric conditions in ways that resting state studies cannot (10). However, understanding specifically how impairments in micro or macro-circuits in the brain explain mental disorders remains a significant open problem in clinical neuroscience (150) and one that limits interpretations from a majority of fMRI studies including our own. Secondly, an absence of longitudinal data limits our ability to comment on the long-term impact of our observed connectivity deficits on the course of OCD symptomatology. Finally, as an exploration on putative mechanisms (see Introduction), we are not in a position to comment on whether treatment (or time) results in an amelioration of our observed connectivity impairments. This brief list of limitations is far from exhaustive, but serves as a road map for potential future investigations in the clinical neuroscience of the disorder.

In conclusion, large scale meta-analyses of network function are challenging given that network interactions are contextually tuned by task characteristics (151). However, we suggest that a growing corpus of studies on the effective (dys)connectivity of brain networks in OCD can contribute to understanding of the task conditions under which network dysfunction is evoked in the illness. Moreover, mechanism-based markers of dysfunction can also serve as therapeutic targets for long term intervention.