Whole-brain functional connectivity predicts ultra-high risk for psychosis status and level of functioning

Abstract

Resting-state functional magnetic resonance imaging (rs-fMRI) has offered insights into the neural mechanisms underlying psychosis, particularly when associated with clinically relevant features. 102 individuals at ultra-high risk for psychosis (UHR) and 105 matched healthy controls (HC) aged 18–40 underwent clinical and cognitive assessments and rs-fMRI at baseline. Using a recently developed prediction-based extension of the network-based statistics (NBS-predict), incorporating nested cross-validation, we tested the predictive power of functional connectivity estimated from rs-fMRI data, investigating diagnostic classification and prediction of level of functioning, estimated IQ, and UHR-symptoms. Hyper-connectivity predicted group with a classification accuracy of 0.58, p = 0.043, and hypo-connectivity predicted group with a classification accuracy of 0.59, p = 0.018. Hyper-connectivity in UHR-individuals was observed primarily in interhemispheric and cortico-thalamic connections, within networks that predicted poorer levels of functioning across groups. Hypo-connectivity in UHR-individuals was observed mainly in thalamic connections with posterior cingulate cortex, frontal medial, and precuneus, within networks that predicted higher level of functioning across groups. Post hoc analyses identified a significant groupwise interaction effect on the association between functional connectivity and level of functioning (ρ = 0.34, p < 0.001), with main nodes in the frontal medial regions connected across hemispheres. Within-group, no connections predicted level of functioning or UHR-symptoms. Whole-brain functional connectivity predicted UHR-status in hyper- and hypo-connected networks, with thalamus as a central integrative hub across networks. Connections that predicted level of functioning across groups were equivalent to the connections predicting UHR-status, hence capturing a neural correlate to a key clinical component of the UHR-status.

Introduction

The ultra high-risk state for psychosis (UHR) is established through clinical screening interviews that identify individuals at elevated risk for developing psychosis. Although the majority of UHR individuals do not transition into frank psychosis, they frequently experience persistent long-term impairments in level of functioning and cognition^1,2. Identifying biomarkers associated with such clinically relevant outcomes during the prodrome may inform early intervention strategies³ and complement clinical predictions². Resting-state functional magnetic resonance imaging (rs-fMRI) has emerged as a tool for elucidating the neural mechanisms underlying the onset of psychosis^4,5. Alterations of neural circuits among UHR-individuals may indicate transition risk⁵, and reflect mechanisms underlying the equally important clinical features of persistent low functioning and symptomatic distress in UHR-individuals who do not transition to psychosis^1,6.

Previous studies have documented altered resting-state connectivity in UHR-individuals, particularly within the frontoparietal network⁷, and within distinct patterns of thalamo-cortical connectivity, which may reflect neurobiological features associated with UHR-status compared to healthy controls (HC)⁸. However, diverse methodologies and modest populations of previous studies, the majority with sample sizes below 50 UHR-individuals^4,9–14, call for further studies using gold standard methodology, such as whole brain prediction studies employing cross-validation in larger samples.

Graph models are a common framework for studying functional connectivity, and network-based statistics (NBS)¹⁵ is a well-known tool for minimizing the challenges handling mass univariate analyses. More recently, a prediction-based extension has been developed, NBS-Predict, combining machine learning with graph theory in a cross-validation structure, providing an applicable tool to identify generalizable neuroimaging-based biomarkers¹⁶. Another comparable method to NBS-Predict is Connectome-based Predictive Modeling (CPM) [refs: Finn et al. (2015), Shen et al. (2017)]. An advantage of NBS-Predict is that it only uses the largest connected component identified in the suprathreshold connections, removing potentially spurious connections, whereas CPM uses all suprathreshold connections as features. Serin et al. demonstrated the application of NBS-Predict on rs-fMRI data in patients with schizophrenia and healthy controls, showing good classification accuracy of 90%. The study identified a hypo-connected subnetworks localized in frontotemporal-, visual-, motor-, as well as in subcortical regions in patients with schizophrenia. Moreover, NBS-Predict revealed a large-scale sub-network associated with general intelligence. To our knowledge, no other studies have applied NBS-predict on UHR-individuals.

Aims

Primarily, we investigate the predictive power of whole-brain functional connectivity to classify a well-characterized sample of UHR-individuals from matched HC using NBS-predict. Based on previous studies^17,18, we expected UHR-individuals to present with dysconnectivity in fronto-parietal and thalamo-cortical networks.

Secondly, we examine whether functional connectivity can predict level of functioning and estimated general IQ in all participants across groups.

In planned post hoc, we examine (a) if functional connectivity can predict level of functioning and estimated general IQ separately within the two groups; and predict symptom level in UHR-individuals, and (b) if there is a groupwise interaction effect when comparing UHR-individuals to HC on the association between functional connectivity and level of functioning.

Methods

Data were acquired from a randomized clinical trial examining the effect of cognitive remediation compared to treatment as usual in UHR-individuals (The FOCUS-trial)¹⁹. Recruitments occurred at psychiatric facilities in Copenhagen, Denmark, from April 2014 to December 2017. For a balanced sample size we included 50 HC examined in parallel (the PECANS-II study)²⁰. All participants provided written informed consent, and the trial protocols were approved by the Committee on Health Research Ethics of the Capital Region, Denmark (H-6-2013-015, H-3-2013-149).

Participants

The baseline sample consisted of 102 UHR-individuals and 105 HC aged 18-40 years. UHR-individuals met at least one of the three UHR-criteria according to the Comprehensive Assessment of At-Risk Mental States²¹: attenuated psychotic symptoms, brief limited psychotic episodes, or state-and-trait vulnerability (a first-degree relative with psychotic disorder, or a diagnosis of schizotypal personality disorder). Exclusion criteria included a previous psychotic episode of ≥ 1-week’s duration, psychiatric symptoms explained by a physical illness with psychotropic effect or acute intoxication, or a diagnosis of a serious developmental disorder (e.g., Asperger’s syndrome or IQ < 70). Current or lifetime prescription of any psychotropic medicine was accepted, except current treatment with methylphenidate.

HC were matched to the UHR-individuals on age, sex, and parental socioeconomic status, recruited through internet and community-based advertising, and had no current or previous psychiatric diagnoses, substance abuse or dependency, or first-degree relative with psychotic disorder. Detailed inclusion- and exclusion criteria are described elsewhere¹⁹. This study includes cross-sectional data on rs-fMRI not previously analyzed (see flow chart in Supplementary Fig. S1).

Assessments

Image acquisition and processing

Imaging was performed on a Philips Achieva 3.0 Tesla MRI scanner, using a 32-channel Invivo head coil. For anatomical reference, a whole-brain T1-weighted structural image was acquired at the beginning of each scanning session (TR = 10 ms, TE = 4.6 ms, flip angle = 8°, and voxel size = 0.79 × 0.79 × 0.80 mm³). For functional MRI, a T2-weighted EPI sequence was obtained (TR = 2 s, echo time = 25 ms, flip angle 75°). Matrix size = 128 × 128 × 38 and field of view = 230 × 230 × 128 mm³, resulting in a voxel-size of 1.8 × 1.8 × 3.4 mm³. A total of 300 volumes were acquired over 10 min. Subjects were instructed to close their eyes and let their mind wander.

fMRI processing

Image processing was performed using SPM12 and Matlab2023b, including motion correction, co-registration, segmentation, and normalization to MNI space. Resting-state postprocessing consisted of an initial high-pass temporal filter (0.005 Hz) to all voxel time-courses and to the motion time-courses and nuisance correction to remove gradual drifts in the signal. Nuisances regression of all voxels using CompCor^22,23 included the twelve motion parameters (i.e., absolute and differential motion) and the CompCor components explaining at least 30% of the variance for white matter and CSF voxels. To identify the white matter and CSF voxels that entered CompCor, each subject’s white matter and CSF tissue probability maps were thresholded at a high tissue probability (i.e., 99.8%) to avoid partial volume effects. CompCor nuisance components were identified by performing a Principal Component Analysis (PCA) within the white matter and CSF voxels and selecting those PCA component time-courses that accounted for 30% of the variance within each of these tissue types. After nuisance regression, functional images were spatially-smoothed (6-mm FWHM) and temporally-filtered (bandpass 0.01–0.1 Hz).

Connectivity matrices

Nodes of the whole-brain functional connectivity matrices were delineated using two recent homotopic functional parcellations, which both showed a high degree of replicability and correspondence with cytoarchitecture^24,25. For the subcortical parcellation, the 3 T, 54 parcel dimensionality was used²¹, and for the cortical parcellation, we used the 200 parcel dimensionality²⁵. The average time-courses within each of the 254 parcels were extracted. The time-courses were scrubbed to remove any timepoints with frame-wise displacement >0.3 mm. The connectivity matrices were generated by computing the pairwise correlation between the average parcel time-courses (i.e., the average time-course across all voxels within a parcel). Subsequently, Fisher’s z-transformation of Pearson’s R correlation coefficients was calculated. This resulted in undirected connectivity matrices with 254 nodes and up to: n*(n-1)/2 = 254 * 253/2 = 32,131 connections. Each participant’s connectivity matrix was input to the subsequent data analyses.

Clinical and cognitive assessments

The level of UHR-symptoms was assessed using the CAARMS composite score, using the positive symptoms subscale of CAARMS²⁶ to obtain intensity and frequency of attenuated psychotic symptoms (APS). Level of functioning was assessed using the Social and Occupational Function Assessment Scale (SOFAS)²⁷.

Diagnostic assessment were conducted using The Structured Clinical Interview for DSM-IV Axis I Disorders (SCID-I) and part of the Structured Clinical Interview for DSM-IV Axis II Disorders (SCID-II)^28,29 by experienced psychologists and medical doctors with comprehensive training in the assessment instruments. Inter-rater reliability was excellent (ICC ratings from. 97 to. 99)³⁰.

Current total IQ was estimated using the third version of the Danish Weschler Adult Intelligence Scale (WAIS-III)³¹, aggregating the Similarities and the Block Design subtest³².

Statistical analyses

Descriptive data analyses were performed using Statistical Package for the Social Sciences (SPSS) version 25.0 (Armonk, NY: IBM) and reported as percentages, means, and standard deviations. Group-wise analyses (UHR-individuals vs HC) were conducted using Chi-square tests and ANOVA.

Functional connectivity analyses

To address mass-univariate testing limitations and to make individual level predictions, we applied the Prediction-based extension of the Network-based Statistics (NBS-predict)¹⁶. The NBS-Predict¹⁶ toolbox is available from https://github.com/eminSerin/NBS-Predict. We applied the default 10-fold cross-validation (CV) procedure with 10 repeats as implemented in the NBS-Predict toolbox. In each fold, feature selection was performed using a univariate test in the training set, retaining connections with a significance level <0.01. Next, the largest connected component of the whole-brain network was identified. A connected component comprises all nodes and suprathreshold (p < 0.01) connections that are connected (directly or indirectly). The largest connected component is the one with the most nodes. Only connections that were part of the largest component were used as features in the prediction models. By only assessing the largest connected component, spurious connections that may stem from noise are not considered. The connection weights (i.e., importance) were based on how often they were selected across folds and the predictive performance in the test set of the folds in which they were included. For the predictions, a feature weight threshold of zero was consistently applied as per default, i.e., including all significant connections within the largest whole-brain component. Hyperparameter optimization was performed using Grid Search with five steps on the training data using nested 5-fold CV. The significance levels of the predictive performances were estimated using permutation tests with 1000 permutations. In all analyses, we tested all the models available in NBS-Predict: For classification, Logistic Regression (LR), Support Vector Machine Classification (SVMC), and Linear Discriminant Analysis (LDA). The performance was evaluated using accuracy. For the regression analyses, Linear Regression (LR) and Support Vector Machine Regression (SVMR). The performance was evaluated using Pearson’s correlation between the predicted values and the dependent (true) variable. To remove potential confounding effects, we included age, sex, framewise displacement, and percent scrubbed as covariates in all the analyses. The contrast for the prediction of the group interaction was calculated by multiplying the variable of interest (i.e., level of functioning or IQ) with the group variable mapped as {−1, 1}.

Sensitivity analyses tested potential confounding effects of antipsychotic medication by four repeated analyses, comparing HC with subsamples of (a) antipsychotic-naïve UHR-individuals; (b) antipsychotic-medicated UHR-individuals; (c) UHR-individuals who were antipsychotic free at baseline; and (d) using chlorpromazine equivalent dose³³ (CPZ-EQ) as covariate in the full sample of UHR-individuals. Additionally, we tested if functional connectivity could differentiate between the subsamples of antipsychotic-naïve UHR-individuals versus antipsychotic medicated UHR-individuals.

All network figures are visualized using BrainNet Viewer³⁴.

Results

Demographic characteristics of participants at baseline are reported in Table 1. Of the UHR individuals, 3% were included on the trait/state criteria alone. The remaining UHR individuals fulfilled either the APS criteria alone (76%) or APS in combination with BLIPS or trait/state (21%). Only limited (68%) and self-reported data on illness duration were available, where illness duration was defined as the estimated time since the first psychiatric contact due to APS. Apart from three outliers, the illness duration appears to be mainly limited to the first year before study enrollment.

Table 1.

Sociodemographic and clinical data.

Variable, mean (S.D.)/percent (N)	UHR (N = 102)	HC (N = 105)	Group comparison significance, effect size
Age	23.8 (4.3)	23.4 (4.1)	p = 0.469, F = 0.53
Sex
Female	53.9% (55)	56.2 % (59)	p = 0.743
Parental SES, % (high/medium/low)	52.9%/36.3%/10.8%	63.8%/34.0%/2.1%	p = 0.156
Clinical
Functioning (SOFAS)	55.0 (11.1)	88.0 (5.2)	p < 0.001, F = 434.48
UHR symptoms (CAARMS)	50.8 (14.6)	–	-
ESTIMATED IQ (FIQ)	103.2 (12.1)	112.4 (12.2)	p < 0.001, F = 29.66
DSM diagnosis
Affective MDD (63), DD (3), BD (6)	71% (72)	–
Anxiety PD (23), SP (19), AP (13), SpP (9), OCD (5), GA (5)	51% (52)	–
Personality disorders SPD (36), BPD (12), PPD (10)	36% (37)	–
Other SoD (1), BDD (1), ED (6), PTSD (4), H (1)	13% (13)	–
Medication
Antipsychotic-naive	66% (67)	–	-
Antipsychotics current	31% (32)	–
Antidepressants current	25% (25)	–	-
Mood stabilizers current	6% (6)	–	-
Benzodiazepines current	8% (8)	–	-
Substance use %
Nicotine			p < 0.001
Never/once/monthly/weekly/daily use	45%/6%/4%/4%/41%	39%/10%/43%/5%/2%
Alcohol			p = 0.001
Never/once/monthly/weekly/daily use	14%/14%/37%/32%/3%	40%/6%/29%/23%/2%
Cannabis			p < 0.001
Never/once/monthly/weekly/daily use	72%/13%/6%/6%/3%	54%/38%/3%/5%/0%
fMRI
Framewise displacement	0.160 (0.122)	0.138 (0.077)	p = 0.117, F = 2.47
Framewise displacement scrubbed	0.145 (0.089)	0.135 (0.076)	p = 0.410, F = 0.68
Percentage scrubbed	0.175 (0.184)	0.159 (0.166)	p = 0.514, F = 0.43

Significant effect of group is marked in bold.

A Agoraphobia, AP antipsychotics, BD bipolar disorder, BDD body dysmorphic disorder, BPD borderline personality disorder, CAARMS comprehensive assessment of at-risk mental states, DD dystymic disorder, ED eating disorder, FIQ estimated total IQ, GA generalized anxiety, H hypochondriasis, HC healthy controls, MDD major depressive disorder, N number, OCD obsessive-compulsive disorder, PD panic disorder, PPD paranoid personality disorder, PTSD Post traumatic stress disorder, SD standard deviation, SoD somatoform disorder, SOFAS social and occupational functioning assessment scale, SP social phobia, SPD Schizotypical personality disorder, SpP specific phobia, UHR individuals at ultra-high risk for psychosis.

There were no differences between UHR individuals and HC on age, sex, and parental socioeconomic status (Table 1). Significant differences between UHR individuals and HC with large effect sizes on level of social- and occupational functioning (p < 0.001, F = 800.57), and estimated IQ (p < 0.001, F = 29.66). UHR-individuals reported significantly higher tobacco use compared to HC, particularly daily use within the last three months (UHR 41% vs. HC 2%, p < 0.001). Alcohol consumption was also elevated among UHR-individuals during this period, with higher frequencies observed for weekly (UHR 32% vs. HC 23%) and monthly use (UHR 37% vs. HC 29%), while fewer UHR-individuals reported abstaining from alcohol (UHR 14% vs. HC 40%, p = 0.001). Five met criteria for current alcohol abuse or dependency. Regarding cannabis use, three UHR-individuals reported daily use in the past three months (p < 0.001), with one meeting diagnostic criteria for cannabis dependence. In contrast, recreational cannabis use appeared more prevalent among HC, as indicated by a lower proportion reporting no use in the past three months (UHR 72% vs. HC 54%). There were no differences on sociodemographic and clinical variables of age, sex, parental socioeconomic status, level of functioning, and APS when comparing antipsychotic-naïve to antipsychotic medicated UHR-individuals. However, antipsychotic-medicated UHR-individuals had a higher estimated IQ of 106 (F = 5.127, p = 0.026) when compared to antipsychotic-naïve UHR-individuals with an estimated IQ of 101.

Diagnostic classification

NBS-Predict achieved a diagnostic classification accuracy of 0.584 (95% CI: 0.574–0.594, p = 0.043) and 0.588 (95% CI: 0.574–0.601, p = 0.018), respectively, when assessing hyper- and hypo-connectivity in the UHR individuals compared to HC. The logistic regression classifier was the best-performing algorithm in both cases, although the three algorithms had similar performance (Table 2).

Table 2.

Accuracy of diagnostic classifiers.

	Hyper-connectivity (UHR > HC) Accuracy (95% CI), significance	Hypo-connectivity (UHR < HC) Accuracy (95% CI), significance
Logistic regression	0.584 (0.574–0.594), p = 0.043	0.588 (0.574–0.601), p = 0.018
Support vector machine classifier	0.577 (0.565–0.589), p = 0.005	0.586 (0.576–0.597), p = 0.005
Linear discriminant analysis	0.574 (0.561–0.587), p = 0.023	0.569 (0.549–0.589), p = 0.045

Tne table shows the results from the three different classification algorithms available in NBS-predict.

CI confidence interval, HC healthy controls, UHR individuals at ultra-high risk for psychosis.

The single node with the most connections (52 edges) was left prefrontal cortex 2. However, when merging the thalamic subregions, thalamus presented overall as the region with the most cortical and subcortical connections (510 edges).

Applying a conservative feature weight threshold of 1 to visualize the largest subnetwork comprising the most relevant nodes and edges consistently selected across the CV folds and repeats, we identified:

1. A sub-network characterized by higher functional connectivity in UHR-individuals compared to HC, see Fig. 1a. This hyper-connectivity in UHR individuals was observed mainly in interhemispheric and thalamo-cortical connections between the thalamic subsegments (135 edges), connecting thalamus across hemispheres with the somato-motor network (57 edges), the temporo-parietal network (42 edges), the dorsal attention network (25 edges) including the temporo-occidental, postcentral, and the precentral regions; and connecting thalamus with nodes in the default mode network (14 edges) including the temporal region (for details see Supplementary Table S5). The thalamic subsegments involved were particularly the dorso-anterior, ventro-anterior, and ventro-posterior segments, as shown in Fig. 2c, d.

2. A sub-network characterized by lower functional connectivity in UHR-individuals compared to HC, see Fig. 1b. This hypo-connectivity in UHR-individuals was observed mainly in thalamo-cortical networks, connecting thalamus with posterior cingulate cortex and precuneus within the control network (28 edges); and connecting thalamus with frontal medial nodes within the salience ventral attention networks (22 edges) (for details see Supplementary Table S6). The thalamic subsegments involved were particularly the dorso-anterior, ventro-anterior, and ventro-posterior segments, visualized in Fig. 2c, d.

Fig. 1. The largest network components of the group difference and correlations to the functional level.

The figure illustrates in the top row a the largest hyper-connected network component at threshold = 1 in UHR-individuals compared to HC, and b the largest hypo-connected network component at threshold = 1 in UHR-individuals compared to HC. In the bottom row, c the network with negative correlation between functional level and connectivity, all participants, and d the network with positive correlation between functional level and connectivity, all participants. The nodes are colored according to the network they are part of, defined by the color bar. The visualizations are made using BrainNet Viewer.

Fig. 2. Thalamic subsegments and connections.

The figure illustrates the specific subsegments of the thalamus, which are a hyper-connected with cortical regions and b hypo-connected with cortical regions. c Illustrates thalamus location, and d the subsegments of thalamus as visualized using BrainNet Viewer.

Groupwise associations to clinical and cognitive measures

Level of functioning

Functional connectivity predicted the level of functioning in all participants across both groups, with a Pearson’s correlation coefficient of ρ = 0.27 (95% CI: 0.25–0.28, p = 0.003) between predicted and actual level of functioning. Support vector machine regression was found to yield the highest predictive performance.

Applying a conservative feature weight threshold of 1 to visualize the largest subnetwork comprising the most relevant nodes and edges across CV folds and repeats, we identified:

1. A subnetwork positively correlated with the level of functioning. The subnetwork consisted mainly of nodes in thalamo-cortical networks, with 48 edges connecting the thalamic subsegments to the control (18 edges) and salient ventral attention (23 edges) networks. The thalamic subsegments involved were particularly the dorso-anterior, ventro-anterior, and antero-posterior segments, see Fig. 1d (Supplementary Table S7 for details).

2. A subnetwork negatively correlated with the level of functioning. The subnetwork comprised mainly nodes in interhemispheric and thalamo-cortical networks, connecting the thalamic subsegments (86 edges) across hemispheres within the somato-motor (26 edges), tempo-parietal (33 edges), dorsal attention (21 edges), and default (7 edges) networks. Figure 1c shows the thalamic subsegments involved, particularly the dorso-anterior, ventro-anterior, and ventro-posterior segments (Supplementary Table S8 for details).

Functional connectivity did not significantly predict estimated IQ across (ρ = 0.022, 95% CI: −4.682 * 10⁻⁴−0.044, p = 0.444).

Post hoc analyses

Functional connectivity did not significantly predict level of functioning separately within groups (UHR-individuals: ρ = 0.15, 95% CI: 0.13–0.18, p = 0.166; HC: ρ = 0.17, 95% CI: 0.14–0.20, p = 0.150) when including all UHR. Furthermore, functional connectivity did not significantly predict estimated IQ within groups (UHR-individuals: ρ = 0.11, 95% CI: 0.08–0.14, p = 0.173; HC: ρ = 0.09, 95% CI: 0.06–0.12, p = 0.364). Finally, functional connectivity did not significantly predict APS within UHR-individuals: ρ = 0.15 (95% CI: 0.12–0.19, p = 0.227) between predicted and actual CAARMS scores.

We identified a significant groupwise interaction effect on the association between functional connectivity and functional level (ρ = 0.34, 95% CI: 0.32–0.36, p < 0.001). For visualization purposes, a feature weight threshold was applied to visualize only the connections which most consistently contributed to the prediction, i.e., the threshold did not influence the prediction performance. Three equally sized subnetworks comprised the most important connections, depending on the specific threshold. Specifically, decreasing the threshold minimally, from 1 to 0.95 changed which of the subnetworks that were the largest component, as an additional similarly sized network appeared, comprising nodes in the default mode network. At a threshold of 0.8, the three subnetworks at threshold = 1 were connected and constituted the largest component, and the default mode network constituted the second largest component, visualized in Supplementary Fig. S2.

When applying a threshold of 1, the largest subnetwork comprised 6 nodes and 12 edges. The main nodes, as shown in Fig. 3 were frontal medial regions involved in the salient ventral attention network, which connected across hemispheres and with nodes in left hemisphere to the somato-motor and dorsal attention networks (8 edges), see details in Supplementary Table S9. The second largest component comprised 5 nodes, mainly in the default mode network, and the third largest component comprised 4 nodes located in somato-motor and salience ventral attention networks with cortico-cortical connections across hemispheres.

Fig. 3. The largest network component of the groupwise interaction effect.

The figure illustrates the largest subnetwork that predicted the groupwise interaction effect on the association between connectivity and level of functioning. To the left, the subnetwork, with nodes colored according to the network they are part of. To the right, scatterplots of the groupwise associations of connectivity between the specific nodes and the level of functioning. UHR-individuals are marked as red and healthy controls as blue. The brain visualizations are made using BrainNet Viewer.

The groupwise interaction effect on the association between functional connectivity and estimated IQ was borderline significant (ρ = 0.17, 95% CI: 0.15–0.19, p = 0.057).

Sensitivity analyses

We tested for potential confounding effects of antipsychotic medication by four repeated analyses with subsamples of antipsychotic-naïve, antipsychotic-free, or antipsychotic-medicated UHR-individuals, as well as CPZ-equivalent-adjusted parameters as covariates in the full sample of UHR-individuals. We found that the prediction of group remained significant for hyper-connectivity, when analyzing the subgroups of antipsychotic-naïve and antipsychotic-free at baseline UHR-individuals. For hypo-connectivity, only the analysis of the antipsychotic-free subgroup remained significant. When using CPZ-EQ dose as a covariate to analyze all UHR-individuals in the model, the diagnostic classification was not significant. When examining the groupwise interaction effect on the association between connectivity and level of functioning, the results remained highly significant regardless of UHR-subgroup, see Supplementary Table S10. Connectivity could not classify antipsychotic medicated UHR-individuals from HC, nor from antipsychotic-naïve UHR-individuals (Supplementary Table S11).

We further investigated the potential effect of age³⁵ comparing individuals above and below or equal to 30 years of age. Only 9 UHR individuals were above age 30. We found no differences on functional level (p = 0.113), estimated IQ (p = 0.239), or UHR symptoms (p = 0.538). Additionally, we tested the difference in the average functional whole-brain connectivity between the two age groups and found no statistical differences, see Supplementary Fig. S3 for details.

Discussion

In this first study using NBS-Predict in a clinically well-characterized sample of 102 UHR individuals, we find that functional connectivity can predict individual UHR-status, with altered connections in UHR-individuals equivalent to those predicting level of functioning across groups. Networks with hypo-connectivity in UHR individuals were associated with better functioning, whereas networks with hyper-connectivity in UHR individuals were associated with poorer functioning across groups.

Previously, we have reported aberrant structural connectivity from this UHR-sample³⁶, showing clusters of subtle focal white matter alterations, contrasted with the widespread alterations of whole-brain functional connectivity observed here. fMRI is proposed to be more sensitive than structural MRI in detecting subtle dynamic changes in brain connectivity during early stages of mental illness³⁷. Notably, the dynamic alterations in functional connectivity may precede observable structural changes associated with illness progression, providing critical insights into the early pathogenetic processes underlying the development of psychosis³⁸ and guiding timely interventions.

Although the involvement of various neural pathways in psychosis risk is supported by previous studies across diverse methodologies³⁹, our whole-brain findings, as expected, revealed dysconnectivity in fronto-parietal and thalamo-cortical networks of UHR-individuals. This is corroborating a recent study by Anticevic et al.¹⁷ that specifically investigated functional connectivity in a large group of young individuals at clinical high risk for psychosis while using thalamus as seed. Thalamic dysconnectivity appeared particularly pronounced in those who transitioned to psychosis. Thalamo-cortical dysconnectivity has been reported in both early and chronic stages of psychotic disorders using ROI- and seed-based analyses^40,41, and similarly to findings in patients with schizophrenia⁴², the marked role of thalamus as a central hub for integration of information across networks is confirmed in our study using a whole-brain approach⁴³. This central role of thalamus in integrating sensory and cognitive information may also be crucial for the regulation of the interplay between default and task-positive networks⁴⁴, highlighting the importance of the default network in the pathophysiology of psychosis⁴⁵. Our findings align with literature on aberrant thalamo-cortical connectivity across the psychosis spectrum, potentially underlying behavioral impairments⁴⁶.

Importantly, the functional networks predicting UHR-status mirrored those predicting level of functioning across groups. This robust finding, supported by the prediction model, indicates that functional networks serving as early biomarkers of psychosis risk are strongly associated with social and occupational functioning. Specifically, connectivity that supports better functioning across populations is hypo-connected in UHR individuals, whereas connectivity associated with poorer functioning across populations is hyper-connected in UHR-individuals. Our finding underscores that even in the absence of psychotic transition, a low level of functioning is a major hallmark of the UHR-status, highlighting its clinically relevance and neurobiological underpinning⁴⁷.

Interestingly, a longitudinal study by Allen et al.⁴⁸ reported that higher cortical and subcortical functional activation during a verbal fluency task at baseline correlated negatively with functional outcome after 18 months. Similarly, Bernard et al.⁴⁹ found that higher connectivity in the medial motor regions, supplementary motor areas, and thalamus at baseline was associated with symptomatic worsening over 12 months. The association between higher functional activation and poor functional and clinical outcomes may reflect that hyper-connectivity is the result of compensatory activation to maintain performance under the strain of underlying neurological deficits⁴⁸. Whether hyper-connectivity can serve as a predictor of transition cannot be investigated in the current sample, as only limited data on transition is available; only ten UHR-individuals had transitioned to psychosis one year after baseline. However, univariate comparison of the average connectivity in the hyper-connected network indicates numerically higher connectivity in the non-transition group compared to the transition group (Supplementary Fig. S4). In future studies, including more individuals with transition, the analytical framework can be applied to predict transition instead of UHR-status. In our study, thalamo-cortical networks negatively associated with level of functioning across groups are hyper-connected in UHR-individuals, but particularly the group interactions between connectivity and level of functioning revealed a set of cortical connections that may provide such a compensatory role in the UHR group. As displayed by the scatterplots in Fig. 3, more connectivity of these cortico-cortical connections predicted better level of functioning in the UHR group, in line with the interpretation that stronger cortical connectivity among these salience and motor regions compensates for aberrant thalamo-cortical connectivity within the same regions. This result should, however, be interpreted with caution due to the marked ceiling effect of the functional level of HC, potentially obscuring both the correlation with connectivity and contributing to the interaction effect. The multiple interaction effects revealed by selection of different thresholds (see Supplementary Fig. S2) underscores the complexity of such a relationship between brain function and level of functioning in UHR. Interestingly, beyond the largest component reported at threshold 1, potentially important nodes from smaller components, such as homotopic connections between somato-motor nodes, and a marked stable component comprising Default network. Studies have shown that emergent networks from varying thresholds may yield insights into functional networks, exhibiting informative variation in connectivity when predicting behaviour or symptoms⁵⁰. Although we here aim to reduce potential noise from the data⁵¹ by maintaining the strict criterion and report only the largest connected component, NBS-predict appear to enable more layered complexity and comprehensive exploration beyond the scope of our study.

In our UHR-sample, functional connectivity could not predict IQ, unlike Serin et al.¹⁶ in patients with schizophrenia. This may reflect a true difference between populations, or a potential nonlinear association with IQ, as demonstrated in Kravariti et al.⁵². We could not test this hypothesis due to the removal of the decision tree algorithm from the NBS-predict toolbox⁵¹, which might capture such nonlinear relationships.

Sensitivity analyses indicated our overall results overall were not influenced by antipsychotic medication. However, the non-significant result when including CPZ-EQ dose as covariate, suggest that incorporating UHR-individuals receiving antipsychotics in dose-sensitive models may attenuate functional connectivity differences compared to HCs. We found no evidence supporting the hypothesis that antipsychotics normalizes functional activity as no differences emerged between medicated and antipsychotic-naïve UHR-individuals. This may be due to reduced power or the low dose antipsychotics prescribed, not bringing about significant differences when examining subgroups. Alternatively, the impact of antipsychotic medication on functional connectivity may be multifaceted, influenced by dosage effects⁵¹ and clinical heterogeneity, as demonstrated in patients with first episode psychosis^53,54. Similarly, Fornito et al. suggest a bidirectional relationship between functional connectivity and medication, with varying connectivity patterns linked to different clinical features and responses to antipsychotic interventions³⁹. Future studies should prioritize larger and more diverse samples to fully investigate subgroup-level effects.

Methodological considerations: Strengths include data from a large, well-characterized sample of UHR individuals. A general limitation of the study is that originally the study was not designed to answer the current research question, but with our large sample size and rigorous data collection and management, the current analyses still contribute with novel insights. Sex or age dimensions have not been investigated separately, beyond their use as covariates in the analyses, which constitutes a limitation to the generalizability of our research.

The study used 10 times repeated 10-fold cross-validation for robust individualized predictions of diagnosis and behavior. Only the largest connected component when using a threshold of zero was used in the prediction, thus avoiding spurious connections end ensuring stable and consistent brain-behavior associations. For visualization purposes, thresholding allows us to display the most important connections for the prediction. A threshold of one, as in the results displayed, indicates that only connections present in every fold across the 10 × 10-fold cross-validation are visualized. This ensures high confidence in the reported network structure and reduces the risk of false connections that may arise from data noise⁵¹. However, the largest component may shift depending on the specific threshold if multiple equally sized subnetworks are contributing to the prediction. We demonstrated this shift by testing different thresholds in the interaction analyses (see Supplementary Fig. S2). Here, we showed that at a threshold of 0.95, the same networks constituted the three largest components, but it shifted which network emerged as the largest connected component, likely because it had a few more connections that were only present in 95 out of the 10 × 10 folds, or with slightly lower predictive performance. Importantly, the tested predictions are independent of the selected threshold, but the interpretation of the results may change depending on the threshold used for visualization. Consequently, it is important to report whether the largest component remains stable across multiple thresholds when using NBS-predict. Limitations of NBS-Predict that could be developed in future work include the implementation of other feature selection approaches; currently, only the p value threshold for feature selection is available, as well as the incorporation of both hyper- and hypo-connections as features in a single model.

Conclusion

Whole brain functional connectivity predicted UHR-status in hyper- and hypo-connected networks, confirming the marked role of thalamus as a central integrative hub across networks. Importantly, the connections that predicted level of functioning across groups were equivalent to the connections predicting UHR-status, hence capturing a neural correlate to a key clinical component of the UHR-status. In future studies, it would be relevant to include longitudinal data on not only transition but also remission, hence unraveling potential protective mechanisms of hyper-connectivity.

Author contributions

K.S.A. and T.D.K. have conceptualized the study, performed the analyses, visualized, and written the original draft. ADB has contributed with methodological resources and processing scripts. M.N., B.Y.G., and B.H.E. have provided funding, project administration, supervision, and resources. L.B.G. has contributed with data curation and project administration. All authors have reviewed and edited the manuscript.

Competing interests

B.H. Ebdrup is part of the Advisory Board of Boehringer Ingelheim, Lundbeck Pharma A/S, and Orion Pharma A/S; and has received lecture fees from Boehringer Ingelheim, Otsuka Pharma Scandinavia AB, and Lundbeck Pharma A/S. All other authors report no biomedical financial interests or potential conflicts of interest to disclose.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the authors used Microsoft Copilot in order to reduce and improve the scientific English language in a few selected sections of the manuscript. After using this tool, the authors have thoroughly reviewed and edited the content and take full responsibility for the content of the published article.

Supplementary information

The online version contains supplementary material available at 10.1038/s41537-025-00685-z.