Dynamic and progressive changes in thalamic functional connectivity over the first five years of psychosis

Abstract

The early stage of psychosis (ESP) is a critical period where effective intervention has the most favorable impact on outcomes. Thalamic connectivity abnormalities have been consistently found in psychosis, and are associated with clinical symptoms and cognitive deficits. However, most studies consider ESP patients as a homogeneous population and fail to take duration of illness into account. In this study, we aimed to capture the progression of thalamic connectivity changes over the first five years of psychosis. Resting-state functional MRI scans were collected from 156 ESP patients (44 with longitudinal data) and 82 healthy controls (24 with longitudinal data). We first performed a case-control analysis comparing thalamic connectivity with 13 networks in the cortex and cerebellum. Next, we modelled the shape (flat, linear, curvilinear) of thalamic connectivity trajectories by comparing flexible non-linear versus linear models. We then tested the significance of duration of illness and diagnosis in trajectories that changed over time. Connectivity changed over the ESP period between the thalamus and default mode network (DMN) and fronto-parietal network (FPN) nodes in both the cortex and cerebellum. Three models followed a curvilinear trajectory (early increase followed by a subsequent decrease), while thalamo-cerebellar FPN connectivity followed a linear trajectory of steady reductions over time, indicating different rates of change. Finally, diagnosis significantly predicted thalamic connectivity. Thalamo-cortical and thalamo-cerebellar connectivity change in dynamic fashion during the ESP period. Better understanding of these changes may provide insights into the compensatory and progressive changes in functional connectivity in early stages of illness.

1. Introduction

Psychotic disorders are associated with significant functional impairments and brain abnormalities involving structural and functional connectivity (1–5). These psychosis-related brain and behavioral abnormalities have been observed throughout different stages of illness, from the prodrome through early psychosis into chronic illness. However, neurocognitive and structural neuroimaging evidence suggests that disease-related neuroprogressive changes in the brain do not progress uniformly across stages of illness (6–10). Thus, different modalities appear to follow different and potentially non-linear neuroprogressive patterns. Understanding these patterns may provide deeper insights into the mechanisms of progression of psychotic disorders.

The early stage of psychosis (ESP) is considered to be a critical period during which effective intervention may help slow functional decline and promote favorable outcomes (11–13). Therefore, there is an urgent need to accurately characterize the progression of brain abnormalities, and understand the neurophysiological changes throughout the ESP period. However, studies typically treat patients in the early stage of illness as a homogeneous population, categorically comparing ESP patients with healthy, chronic psychosis, or clinical high risk (CHR) cohorts. Such an approach fails to take advantage of the opportunity to examine the dynamic mechanisms unfolding during this critical period, because complex neurobiological changes are not captured when patients with varied, albeit brief, duration of illness are collapsed into one cohort.

Thalamic functional connectivity (FC) is of particular interest in this field as it is consistently abnormal throughout the course of psychosis, and associated with impaired cognition and symptom severity (14–18). Thalamic connectivity is elevated with the motor and somatosensory cortices, but reduced with the prefrontal cortex, striatum, and cerebellum (15, 19–24). Studies focusing on cerebellar circuits have also suggested that affected cerebellar regions are in cognitive-related networks such as the frontoparietal network (FPN) and default mode network (DMN) (25). Thus, abnormalities in cortico-thalamo-cerebellar circuits are thought to underlie a range of symptoms in schizophrenia (26). Resting state connectivity studies have demonstrated that both cerebellum and cerebrum can be parcellated into a low dimension number of networks linked to specific functions (27, 28). A limitation of some thalamic studies in psychosis is that they measure connectivity to specific networks in cerebrum while treating the cerebellum as a homogenous entity. It is unknown whether there are abnormalities to specific cerebellar networks in ESP patients.

Diagnostic specificity contributes an additional level of complexity. Psychotic disorders can present as affective psychosis (e.g. bipolar disorder [BD] and major depressive disorder [MDD] with psychotic features) or non-affective psychosis (e.g. schizophrenia spectrum disorders [SSD]). During the ESP, diagnosis is fluid due to the clinical and neurophysiological overlap between affective and non-affective psychosis (29). However, cognitive studies suggest that affective and non-affective psychosis may follow different disease trajectories (30, 31). Neuroimaging studies, including thalamic FC, suggest a more graded pattern, where deficits are more severe in SSD patients and less pronounced in affective psychosis patients (19, 32, 33).

In this study, we present a novel approach to examine connectivity trajectories through a retrospective hybrid pseudo-longitudinal dataset, analyzing cross-sectional data from patients at different durations of illness along with a subset of available longitudinal data (n = 44). Although a longitudinal design with frequent sampling of time-points (>2) is optimal to best model the shape and key inflection points of thalamic FC during critical periods in ESP, practically it is difficult and resource-intensive to collect follow-up data from the same individuals more than twice with sufficient sample size. To our knowledge, there are no published thalamic connectivity studies with >2 time-points in ESP literature. By examining both longitudinal and cross-sectional effects in our hybrid approach, we can overcome the challenges of longitudinal studies (only 2 time-points; underpowered studies; attrition bias), and test the assumption in pseudo-longitudinal studies (34, 35) that cohorts at each time-point are representative. This novel approach enabled us to fully utilize the available data collected to study trajectories over six time-points over the ESP period.

We first investigated whether thalamic connectivity with 13 cortical and cerebellar RSNs was different between ESP patients and controls. Next, for each thalamo-RSN connection, we explored the shape of the trajectory in patients during the first five years of illness. Data collected from ESP patients were categorized into six time-points based on duration of illness in years. We specifically examined whether connectivity would remain stable (flat trajectories) or follow a linear (constant rate of change) or curvilinear trajectory as psychosis progressed. Third, for thalamus-RSN connections that changed over time (linear/curvilinear trajectories), we tested if duration of illness (DOI) was a significant predictor of connectivity. Finally, we examined diagnosis specificity (affective/non-affective psychosis) among four thalamic connectivity trajectories that changed over time.

2. Methods

2.1. Participants

A total of 238 participants (ESP=156, healthy controls (HCs) = 82) were included in this study. Patients were recruited from McLean Hospital OnTrack, the first episode psychosis clinic (36) (supplement S1). Each patient’s scan was grouped according to duration of illness (DOI) in years calculated by subtracting participant’s age at the onset of psychosis from the age when scan data was collected (i.e. Time = 1 if scan occurred one year after onset of psychosis). Longitudinal data from 44 patients with repeat scans (n = 109 scans total; supplement S2c) were available. Exclusion criteria for patients included DOI over six years (supplement S1). Age-matched HCs were recruited from the community. This study was approved by the McLean Hospital Institutional Review Board. All participants provided written informed consent. Medication information is presented in supplement S3.

2.2. MRI data acquisition and pre-processing

Imaging data acquisition and pre-processing are described in detail in supplement S4. Briefly, neuroimaging data were acquired using a Siemens 3T Tim Trio scanner with a 12-channel phased-array head coil. Participants were instructed to remain still, and stay awake for the duration of the scan. 2 runs of functional data (6.2 mins each, TR = 3000ms) were acquired per participant. A T1-weighted image was also collected for co-registration. Imaging data pre-processing and connectivity analyses were performed using the CONN toolbox (37). Functional data underwent realignment, slice-timing correction, ART-based outlier detection, direct segmentation (including white matter, gray matter and CSF), co-registration of structural and functional data, MNI normalization, and functional smoothing. ART-based functional outlier detection settings were set at conservative settings (95^th percentile in normative sample) with subject-motion threshold set at 0.5mm. The number of volumes that exceeded this motion threshold for each scan were determined to be potential outliers and summed as invalid volumes count.

2.3. ROIs and First-level analysis

An ROI-to-ROI analysis was conducted in the CONN toolbox between the thalamus (FSL Harvard-Oxford atlas) and the seven-network cortical parcellation identified by Yeo et al. (28) as well as the corresponding cerebellar networks identified by Buckner et al. (27) (supplement S5). As the regions corresponding to the visual network in the cerebellum are not represented in the cerebellum (27), this resulted in a total of 13 thalamo-cerebellar/thalamo-cortical functional connections. Functional connectivity maps for each participant were computed by measuring the bivariate correlation coefficients of the BOLD time series between each pair of seed and target ROIs through a haemodynamic response factor (hrf)-weighted general linear model (GLM).

2.4. Second-Level Analysis: Case-Control comparisons

Second-level analysis was performed in the CONN toolbox to assess differences between patient/control groups at the baseline scan. The functional connectivity maps were entered in a GLM with group as the between-subject variable, and controlled for gender, age and number of invalid scan volumes. For each of the 13 connections, GLM results were thresholded at a significance level of 0.05 using the false discovery rate (FDR) approach to correct for multiple target ROI comparisons.

2.5. Shape of the connectivity trajectory over first six years of psychosis

The Fisher-transformed bivariate correlation coefficients of all 13 connections were extracted for ESP patients, including all baseline and follow-up visits within the first six years since age of onset of psychosis (Time = 0 to 5/6). To visualize the shape of each connectivity trajectory, generalized additive models (GAMs) were compared against linear models. GAMs provide a qualitative visual method to plot whether connectivity remained flat, or increased/decreased over time, and were used to select a subset of connectivity variables for further analysis. GAMs were ran using the mgcv package in R (38) where each thalamo-RSN connection was the outcome, time (smoothing function applied) was the predictor of interest, controlling for age and invalid scan volume counts. All scans were included in the GAM models, and each scan was treated as independent even if they were longitudinal data. Based on graphical examination of GAM model results (supplement S6), we selected four thalamo-RSN connections that showed non-flat trajectories (implying change over time). These four variables had (i) non-flat trajectories; (ii) p-value of smoothed time variable < 0.1 or deviance explained > 5%. Model fit (linear vs curvilinear) was assessed using four model evaluation parameters (Akaike information criterion, generalized cross-validation, R², and the χ² test).

2.6. Linear Mixed-Effects Models

The p-value estimates produced by GAMs should be interpreted with caution as they assume independence in the dataset, and do not take into account the correlational structure of longitudinal data. Linear mixed effects models (LME) are more appropriate in handling the semi-longitudinal nature of the dataset, as well as intra-subject variability. Therefore, after the shape of a trajectory was determined (linear vs. curvilinear), random-intercept LME models (StataIC v15.1) were used to quantitatively assess whether Time (DOI) is a significant predictor of connectivity. In the LME models, FC was the outcome and time was a predictor of FC with age, gender, and invalid scan volume count as co-variates. In the curvilinear models, quadratic terms of Time were added. Cross-sectional and longitudinal effects were examined in supplement S7.

To examine diagnosis specificity among those connectivity trajectories that changed over time, diagnosis (Affective/Non-affective psychosis) was added in the LME models as a predictor. Medication effects were also assessed by including medication as a predictor.

3. Results

3.1. Demographics at Baseline

Supplement S8 presents demographics data at their baseline scan. Demographics were similar except for age, where controls were, on average, 1.5 years older than patients. Controls also had fewer scan volumes excluded during pre-processing. Additional cohort characteristics are provided in supplements S3 and S9.

3.2. Cross-sectional Case-Control comparison

Overall, we observed a trend towards increased thalamo-cortical connectivity across multiple networks but selective network-specific reductions in thalamo-cerebellar connectivity in patients when compared with controls (Fig 2; supplement S10). Patients had significantly higher thalamo-cortical connectivity in the somatomotor network (p-FDR < 0.001), visual network (p-FDR < 0.001), dorsal attention network (p-FDR < 0.001), and the limbic network (p-FDR = 0.00935). Conversely, patients had significantly lower thalamo-cerebellar connectivity in the DMN (p-FDR = 0.0144) and FPN (p-FDR = 0.0144) related regions in the cerebellum.

Fig 2:

Results summarizing case-control analysis of thalamic connectivity by cortical/cerebellar networks. Thalamus connectivity with (A) four cortical networks was significantly elevated in ESP patients compared to controls; (B) two cerebellar networks was significantly decreased in ESP patients compared to controls. Multiple comparison correction (13 connections) was performed with the false discovery rate. Results have undergone Fisher’s z-transformation (Fz) and are expressed as mean ± standard error mean (s.e.m).

Significance: *p < 0.05; **p < 0.01; ***p < 0.001.

3.3. Dynamic connectivity over first five years since psychosis onset

3.3.1. Qualitative assessment of shape of trajectory (GAMs)

Four out of the 13 potential thalamo-cortical/cerebellar FC had a trajectory that changed as psychosis progressed (Fig 3), while the remaining 9 were stable (flat trajectory) (supplement S6).

Fig 3:

Results summarizing the 4 thalamic connectivity trajectories that changed over the ESP period. For each thalamus-network connection, data is presented in three ways: actual (average connectivity at each time-point ± s.e.m [black]), linear (linear model of the data [red]), and GAM (generalized additive model of the data [blue with confidence interval in grey]). χ² test was used to assess model fit between Linear and GAM. The GAM was significantly better at modelling data for all but the thalamus-cerebellar(FPN) connection. Note: Y-axes have different scales

Three connections (Thalamo-cortical DMN [Thal-corDMN] Fig 3A; Thalamo-cortical FPN [Thal-corFPN Fig 3B]; Thalamo-cerebellar DMN [Thal-cerDMN] Fig 3C) followed a curvilinear trajectory where connectivity increased initially, appeared to peak at different time-points, and subsequently decreased. A fourth, Thalamo-cerebellar FPN [Thal-cerFPN] (Fig 3D) connectivity, followed a linear trajectory as the GAM did not model data better than a linear model (p = 0.163; supplement S11); there was steady reduction in connectivity in this network. Follow-up analyses (supplements S12 and S13) suggest that thalamic FC trajectories were relatively flat in controls, and distinct from the trajectories of patients modelled by DOI.

3.3.2. Duration of illness as a predictor of thalamic connectivity

For all four models (thal-corDMN, thal-corFPN, thal-cerDMN, thal-cerFPN), cross-sectional and longitudinal effects were not significantly different in our models (supplement S7) and thus, were combined in the final mixed effects models summarized in Table 1.

Table 1:

Results of the four thalamic connectivity trajectory models that changed over time (thal-corDMN thal-cerDMN corFPN cerDMN)

Outcome	Predictor	β	p	95% CI
				Lower	Upper
Thal-corFPN	Time	0.0412	0.112	−0.00956	0.0919
	Time2	−0.0102	0.041	−0.0199	−0.00043
Thal-corDMN	Time	0.0596	0.019	0.00976	0.109
	Time2	−0.0116	0.018	−0.0212	−0.00195
Thal-cerFPN	Time	−0.0160	0.054	−0.0323	0.000297
Thal-cerDMN	Time	0.06	0.01	0.0141	0.106
	Time2	−0.0147	0.001	−0.0235	−0.00586

The quadratic term of time was a significant predictor for all models with a curvilinear trajectory, while time was a trending predictor for the linear model (p = 0.054). That is, during the ESP period, thalamic connectivity with the FPN and DMN networks in both the cortex and cerebellum followed trajectories that changed over time. These trajectories had different shapes (linear vs curvilinear), indicating network-specificity and plasticity of thalamo-RSN FC.

Medication effects (CPZE dosage and mood stabilizers) were included in the models as co-variates and did not have a significant impact on the observed results (supplement S14).

3.3.3. Diagnostic specificity of thalamic connectivity trajectories

Among four non-flat thalamo-cortical/cerebellar network trajectories, diagnosis had a significant effect in three and was not a significant predictor in the thal-corFPN connectivity (p > 0.5) [Fig 4B].

Fig 4:

Results summarizing the effects of diagnosis specificity on thalamic connectivity trajectories that changed over the ESP period. Patients are diagnosed with affective psychosis (green) or non-affective psychosis (purple). Results are presented as actual data (average connectivity at each time-point ± s.e.m) with the modelled trajectory (linear/quadratic) in bold. P-values displayed test the significance of diagnosis as a predictor in each modelled trajectory, where diagnosis specificity has a significant effect on thal-corDMN and thal-cerFPN connectivity, a trending effect on thal-cerDMN connectivity. Note: Y-axes have different scales

For thal-corDMN connectivity, non-affective psychosis had higher connectivity compared to affective psychosis (β = 0.0799, p < 0.001) [Fig 4A]. A clear pattern emerged where non-affective psychosis patients show a strongly non-linear trajectory whereas the affective psychosis patient trajectory was largely flat.

Conversely, non-affective psychosis had significantly reduced connectivity for thal-cerFPN (β = −0.0538, p = 0.028) [Fig 4D], and a trending reduction in connectivity for thal-cerDMN (β = −0.0392, p = 0.068) [Fig 4C].

3.3.4. Duration of illness as a predictor of thalamic connectivity in longitudinal subset

To further confirm that cross-sectional and longitudinal effects were not significantly different (supplement S7), we performed a longitudinal analysis in a subset of patients with repeated measures (N = 44, 109 scans in total; supplement S2).

Compared to the full dataset, the effect estimates (i.e., beta values) of DOI in the longitudinal subset cohort on the four thalamic connectivity trajectories were similar but with reduced significance due to smaller numbers (supplement S15).

4. Discussion

4.1. Summary of results

Our study provides a proof-of-concept demonstration that a hybrid pseudo-longitudinal design can model trajectories over multiple time-points. To our knowledge, our study is the first to investigate the course of thalamic connectivity with both cortical and cerebellar networks in ESP, and to model the longitudinal trajectories of each network. Cross-sectionally we observed significant differences between ESP patients and controls, and showed that thalamo-cerebellar abnormalities were also restricted to specific networks. FC changed over time in the DMN and FPN thalamo-cortical and thalamo-cerebellar networks. Three of these showed a non-linear pattern of FC elevations in early years of illness followed by reduction, whereas thal-cerFPN showed a linear reduction in FC. SSD was associated with a pronounced non-linear pattern of FC elevation followed by reduction in thal-corDMN, whereas affective psychosis showed no change in FC in this network over time. We did not uncover any evidence for antipsychotic medication-induced changes.

4.2. Case-control differences among all resting state networks

Our findings are consistent with previous findings showing increased thalamic connectivity in motor and sensory cortices and decreased connectivity with the cerebellum in psychotic disorders. We extend the existing literature showing that thalamo-cerebellar FC abnormalities were specific to the DMN and FPN-associated regions of the cerebellum, implicating that these abnormalities were more linked to cognitive-related, rather than movement-related, functions. We did not find a reduction in thalamo-prefrontal cortex connectivity observed in other studies (15), possibly because of functional parcellation methodological differences, i.e. the most similar network (corFPN) also includes parietal regions. It is also possible that as the thal-corFPN FC trajectory was dynamically altered over time, collapsing the time dimension in a cross-sectional analysis led to an averaging of variance in opposite directions.

4.3. Patterns of change of thalamic connectivity over the ESP period

Critically, our results support the notion that thalamo-RSN FC abnormalities do not progress uniformly across the early stages of illness. Four thalamo-cortical networks (motor, visual, DAN, limbic) showed elevated FC at psychosis onset and remained stable over time (supplement S6). On the other hand, thalamic connectivity with DMN and FPN in both the cerebellum and cortex (thal-cerDMN; thal-corDMN; thal-cerFPN; thal-corFPN) showed linear or curvilinear patterns of changes. The observed stable and dynamic patterns of trajectories might be associated with differential timing and maturation of brain circuits. Different functional networks follow different maturation timeframes - sensory and motor networks are thought to have fully matured by adolescence, while networks that mediate higher cognitive processes are among the last to mature (39). Therefore, at the onset of psychosis, mechanisms contributing to abnormalities in thalamic connectivity with networks that mature first would have occurred earlier during development and already stabilized by psychosis onset, while networks linked to higher cognitive functions (DMN, FPN) could still be undergoing refinement allowing changes to be observed during ESP. In addition to maturation timeframes, the observed changes in DMN and FPN networks are involved in higher cognition. It is possible that connectivity changes observed may, to some extent, underlie cognitive changes or associate with clinical symptom severity.

In the four connections that changed over the ESP period, we also observed different trajectory shapes, highlighting dynamic changes following disease onset. Three (thal-corDMN, thal-corFPN, thal-cerDMN) showed curvilinear trajectories while thal-cerFPN followed a linear trajectory. This suggests that the rate of change of thalamic abnormalities significantly differed between networks. Our results support the inefficient neural system stabilization framework proposed by Palaniyappan et al. (40). The different rates of change observed may reflect that these connections are at different stages of the homeostatic compensation process. The thalamus, along with regions linked to the DMN and FPN (e.g cingulate cortex, prefrontal cortex and parietal regions) are known to be highly-interconnected hubs with high activity levels that are most vulnerable to aberrant connectivity and the resulting homeostatic re-stabilization mechanisms (41). Also, the observed trajectory patterns in DMN and FPN resemble findings from MRS studies which show hyperconnectivity in the early stage (linked to excitotoxicity/excess glutamate) but hypoconnectivity in the chronic stage (linked to reduced glutamatergic signaling) (42). In the present study, there is an initial period of hyperconnectivity in thal-corDMN followed by a linear decrease. This pattern again suggests a compensatory mechanism where the network is hyperconnected in reaction to other processes and then downregulates its FC over time as the patient’s condition stabilizes. Such a process may even be related e.g. to dendritic spine reduction which may occur in the first few years after psychosis onset, leading to decreased connectivity. For the linear thal-cerFPN connection, it is possible that compensatory processes started during the CHR stage, thus only the linear decrease portion was captured within the ESP period. Finally, our results suggest that cerebellar regions linked to key cortical networks also undergo re-stabilization mechanisms, similar to re-stabilization processes in cortical hubs/regions posited in existing literature (40).

4.4. Diagnostic Specificity

While affective and non-affective psychosis share overlaps in symptomatology and genetics, our results showed diagnostic specificity in different thalamo-RSN connections and neuroprogressive trajectories. FC study by Anticevic et al. found a “graded response” between BD and SZ, where SZ patients showed greater abnormalities, although resolution to discern differences in specific networks was low (19). Cross-sectionally, we observed a “graded response” that non-affective psychosis patients generally had more severe abnormalities than affective patients specifically in thalamo-cortical-cerebellar RSNs (supplement S16). Longitudinally, we found that three thalamo-RSN connections (thal-corDMN, thal-cerDMN, thal-cerFPN) follow distinct trajectories, suggesting that mechanisms associated with these networks could underlie diagnosis-specific elements. For example, abnormalities in thal-corDMN was only observed in non-affective psychosis.

Other networks (cortical visual, motor, limbic and DAN) share similar trajectory patterns irrespective of diagnosis. A qualitative assessment of trajectories (supplement S17) suggests that thalamic connectivity abnormalities with these networks are already present at the onset of psychosis and remain relatively stable over ESP. However, patients diagnosed with non-affective psychosis had more severe abnormalities that were especially obvious in the motor and limbic networks.

4.5. Validation for the hybrid pseudo-longitudinal approach

Results from the longitudinal subset were similar to the full dataset (supplement S15). This supports our statistical approach where we found no statistical differences between cross-sectional and longitudinal effects (supplement S7). Furthermore, demographics were similar between patients with and without longitudinal data, however symptoms were significantly milder in the longitudinal subset (supplement S2a). This brings up the possibility of attrition bias, as patients with severe symptoms may not return for repeated visits. Thus, it is possible that the estimated effects observed in the longitudinal subset are conservative. On the whole, a hybrid pseudo-longitudinal approach is capable of modelling thalamic connectivity trajectories.

4.6. Limitations

Some limitations have to be considered when interpreting our results. Firstly, this study looked at the thalamus as a whole. However, the thalamus is composed of sub-nuclei with specific functions, and it is possible that different RSNs would be linked to different regions in the thalamus (43, 44). Several studies have published thalamic parcellations, however different groups have found different sub-regions and none match our functional parcellation (15, 43, 45). Furthermore, existing published studies also use the whole thalamus as a ROI, allowing for direct comparisons with our study (20, 25, 46, 47). It would be interesting future work to see if our findings are differentially associated with different sub-divisions of the thalamus, but this is outside the scope of the current study. Secondly, it is difficult to draw parallel comparisons between ESP patients and healthy controls in terms of thalamic connectivity trajectories because controls have no proxy for DOI. We have attempted to address this by modelling thalamic connectivity trajectory changes in healthy controls using age as a predictor (supplement S12), and by showing that time was not a significant predictor in a longitudinal subset of controls (n = 24; supplement S13). Nevertheless, age is controlled for in both GAM and mixed models, and thalamic connectivity trajectories are relatively flat in controls, indicating that the trajectories observed in patients are an effect of illness.

4.7. Conclusion

Thalamic connectivity is abnormal in ESP patients, and follows diverse trajectories depending on resting state functional network. We have shown that DOI is a significant predictor of thalamic connectivity, and several thalamo-RSN connections show distinct trajectories (curvilinear/linear) that are associated with diagnosis specificity, highlighting a potential source of heterogeneity in the critical ESP. In addition, the different rate of change of dynamic trajectories could be associated with re-stabilization processes. Our results have clinical implications by providing a potential timeline for intervention.

Fig 1:

Study design and analysis pipeline. (A) Inclusion criteria for participants and breakdown of time-points defined as duration of illness in years. If data was collected multiple times within a year, only one time-point was included for the within-patient trajectory analysis. (B) The analysis pipeline involving 1. a case-control analysis, and 2. modelling thalamic connectivity in ESP patients over the first 5 years of illness. Analysis for (2) include assessing the shape of dynamic trajectories, testing the significance of time as a predictor of thalamic connectivity, and testing the significance of diagnosis specificity.