Predicting Clinical Improvement in Early Psychosis Using Circuit-Based Resting-State Functional Magnetic Resonance Imaging

Abstract

Background and Hypothesis

Identifying biomarkers that predict treatment response in early psychosis (EP) is a priority for psychiatry research. Previous work suggests that resting-state connectivity biomarkers may have promise as predictive measures, although prior results vary considerably in direction and magnitude. Here, we evaluated the relationship between intrinsic functional connectivity of the attention, default mode, and salience resting-state networks and 12-month clinical improvement in EP.

Study Design

Fifty-eight individuals with EP (less than 2 years from illness onset, 35 males, average age 20 years) had baseline and follow-up clinical data and were included in the final sample. Of these, 30 EPs showed greater than 20% improvement in Brief Psychiatric Rating Scale (BPRS) total score at follow-up and were classified as “Improvers.”

Study Results

The overall logistic regression predicting Improver status was significant (χ² = 23.66, Nagelkerke’s R² = 0.45, P < .001, with 85% concordance). Significant individual predictors of Improver status included higher default mode within-network connectivity, higher attention-default mode between-network connectivity, and higher attention-salience between-network connectivity. Including baseline BPRS as a predictor increased model significance and concordance to 92%, and the model was not significantly influenced by the dose of antipsychotic medication (chlorpromazine equivalents). Linear regression models predicting percent change in BPRS were also significant.

Conclusions

Overall, these results suggest that resting-state functional magnetic resonance imaging connectivity may serve as a useful biomarker of clinical outcomes in recent-onset psychosis.

Introduction

Psychosis may present in a variety of psychiatric illnesses, including schizophrenia spectrum disorders (SZ), bipolar disorder, and major depressive disorder. Response to antipsychotic treatment across these disorders is highly heterogeneous. Indeed, in first-episode patients with SZ, up to 30% of individuals will not respond to first-line antipsychotic treatment.^1–3 These individuals may then be treated with clozapine, of which up to 60% may continue to not respond.⁴ The remaining “ultra” treatment-resistant patients are then prescribed alternative forms of treatment, such as electroconvulsive therapy or combinations of antipsychotics.^5–8 Given that reducing the duration of untreated psychosis is associated with improved outcomes,^9,10 the ability to predict which individuals will be treatment resistant and require alternative approaches to treatment as early as possible would have a major impact on patient care.

Unfortunately, no tools yet exist that accurately predict clinical response in early psychosis (EP), and trial-and-error remains the standard approach to care for individuals presenting for treatment. Furthermore, the neuronal mechanisms that confer treatment response and resistance are poorly understood. For this reason, in recent years researchers have devoted increased attention toward identifying prognostic biomarkers, with resting-state functional magnetic resonance imaging (fMRI) being one of the more commonly studied modalities due to its noninvasive nature and ease of administration. As reviewed by Mehta et al,¹¹ previously reported predictors of treatment response include striatal, default mode network, and salience network connectivity, although the nature of these relationships has varied considerably in direction and magnitude across studies.

In 2011, an intrinsic functional connectivity-based model was proposed in which dysfunction between and within 3 core networks (the attention, default mode, and salience networks) could account for many of the psychopathological features across various psychotic disorders based on studies reporting numerous structural and functional abnormalities in these networks in these illnesses.¹² To our knowledge, however, it is unknown whether intrinsic functional connectivity between these networks can predict treatment response in psychosis. Recently, by performing a Neurosynth-based (neurosynth.org) meta-analysis, Goldstein-Piekarski et al¹³ were able to identify a core set of regions of interest (ROIs) that effectively captured intrinsic states of functional connectivity between these networks. Here, we evaluated the ability of within- and between-network connectivity of the attention, salience, and default networks, as defined using the ROIs in Goldstein-Piekarski et al,¹³ to predict clinical improvement in EP. We hypothesized a significant relationship between the intrinsic connectivity of these networks and reduction in symptoms in recent-onset illness, tested using both a binary outcome (>20% improvement in total Brief Psychiatric Rating Scale (BPRS) score) as well as a linear outcome (change in total BPRS score from baseline to follow-up).

Methods

Sample

Baseline neuroimaging data were available for 127 participants with EP (“EPs”; 86 with SZ (including schizophrenia, schizoaffective, and schizophreniform disorder), 26 with type I bipolar disorder (BD) with psychotic features, 2 with delusional disorder, 6 with major depressive disorder with psychotic features, and with 7 psychosis-not-otherwise-specified). Eighty-seven healthy controls (HCs) were included to calculate standardized connectivity z scores. EPs were recruited as outpatients from the University of California, Davis (UCD) Early Diagnosis and Preventive Treatment (of Psychosis) (EDAPT) research clinic (http://earlypsychosis.ucdavis.edu). Treatment in the clinic follows a coordinated specialty care (CSC) for EP model delivered by an interdisciplinary treatment team. Treatment includes detailed clinical assessments using gold-standard structured clinical interviews and medical evaluations, targeted pharmacological treatments including low-dose atypical antipsychotic treatment, individual and family-based psychosocial education and support, cognitive behavioral therapy for psychosis, and support for education and employment. The Structured Clinical Interview for DSM-IV-TR (SCID)¹⁴ was used for the diagnosis of psychopathology. Diagnoses were confirmed by a group of trained clinicians during case conferences. All participants reported psychosis onset within 2 years of the date of informed consent. Individuals were excluded for a diagnosis of major medical or neurological illness, head trauma, substance abuse in the previous 3 months (as well as a positive urinalysis on the day of scanning), Weschler Abbreviated Scale of Intelligence-2 score (WASI-2)¹⁵ score <70, and magnetic resonance imaging (MRI) exclusion criteria (eg, claustrophobia, metal in the body). All participants provided written informed consent and were compensated for participation. The UCD Institutional Review Board approved the study. The medication regimen (type and chlorpromazine [CPZ]-equivalent dose) was assessed by clinical records at baseline and follow-up. Symptoms were assessed using the 24-point BPRS¹⁶ rescaled to the lowest score of zero (ie, score of 24 = score of 0). At baseline, all participants had BPRS scores ≥5 to ensure sufficient resolution to detect a 20% improvement in score at follow-up.

MRI Acquisition

T1-weighted MPRAGE structural images were acquired for realignment and normalization during preprocessing. Structural imaging parameters were 2530 ms TR (repetition time), 3.5 ms TE (echo time), flip-angle 7°, 256 mm² FOV (field of view), 1 mm isotropic voxels.

Resting-state functional images were acquired on a 3T MR scanner (Siemens Magnetom TimTrio) using a standard quadrature head coil. Images were acquired with the following parameters: 6 m scan time, 2000 ms TR, 28 ms TE, 220 mm² FOV, 3.4 × 3.4 × 4.0 mm voxel size, 33 slices, interleaved, anterior to posterior phase encoding, flip angle 75°, 180 volumes. Subjects were instructed to rest with eyes open while observing a fixation cross.

fMRI Preprocessing

fMRI data were preprocessed using SPM12 (Wellcome Department of Imaging Neuroscience). Briefly, images were realigned and then normalized to the Montreal Neurological Institute (MNI) template using a rigid-body transformation followed by nonlinear warping to individual normalized, segmented T1 images. Data were not smoothed in accordance with CONN guidelines for ROI-to-ROI analysis.¹⁷

Connectivity Analysis

Connectivity analysis was performed using the CONN v.21 toolbox.¹⁸ Functional connectivity was analyzed between a set of 5 mm radius spherical ROIs centered on MNI coordinates that constitute the attention, default mode, and salience networks as determined by Neurosynth (supplementary table 1) as previously determined.¹³ Amygdala ROIs were defined using the AAL atlas, also consistent with Goldstein-Piekarski et al.¹³ Within- and between-network connectivity was calculated by averaging connectivity (Fisher-transformed correlation coefficient absolute values) within and between the ROIs associated with each network (eg, averaging all connectivity values between the salience network ROIs to calculate salience within-network connectivity). Connectivity values were converted to absolute values to facilitate interpretability, as beta values and odds ratios from connectivity-based features would be unable to differentiate between negative connectivity and lower connectivity (ie, it would be unclear whether a negative beta was driven by lower positive connectivity or greater negative connectivity). Prior to analysis, preprocessed fMRI images were scrubbed for movement and other artifacts using the ArtRepair toolbox implementation in CONN. Scrubbing thresholds were global-signal z value >3 and interscan subject motion >0.5 mm, corresponding to the “conservative” settings in CONN (95th percentiles in normative data). Individuals with >50% of scans scrubbed were excluded from the analysis. Vectors constituting the 6 rigid-body movement parameters (x, y, z, roll, pitch, yaw) as well as individual signals from white matter and cerebrospinal fluid (using Freesurfer-derived masks for each participant) were included as first-level nuisance regressors prior to calculating connectivity.

HC vs EP and Improver vs Non-Improver Comparisons

Exploratory comparisons between HC and EP as well as Improver vs Non-Improver connectivity values were performed using t tests with significance set to P < .05, with P < .10 considered trend level.

Missing Data Analysis

Baseline BPRS and connectivity values for EP individuals with and without follow-up data were compared using 2-sample t tests to determine if the sample was biased based on the presence of follow-up data.

Regression

Logistic and linear regression analyses were performed in SPSS v.29 (IBM) using the 6 connectivity values (within-attention, within-default mode, within-salience, attention-default mode, attention-salience, and default mode-salience) as predictors. Time to follow-up, duration of illness, and baseline BPRS score were also included as predictors in secondary models. Finally, to determine if connectivity variables explained significant variance above baseline BPRS scores, stepwise regression models were performed with baseline BPRS as an initial feature and entry criterion (probability) for connectivity features set to P < .05. Change in likelihood ratio was used to calculate probabilities for entry for stepwise logistic regression.

Predictors were converted to z scores prior to regression modeling so that they were all on the same scale. Connectivity values were normalized to the entire starting sample, including HCs. The threshold for statistical significance of the overall model and individual predictors was set to P < .05. Logistic regression models were also assessed based on accuracy, sensitivity, and specificity (setting the classification cutoff to 0.5) as well as receiver operating characteristic area under the curve (ROC AUC) across the full range of classification cutoffs.

Results

Demographic and Clinical Information

Of the starting sample of 127 EPs and 87 HCs, 6 EP individuals and 1 HC were excluded for having >50% of their scans scrubbed, and 63 EPs of the remaining sample did not have follow-up clinical data. Demographic and clinical characteristics of the remaining sample are provided in table 1. No significant differences were observed in age or sex between HCs and EPs. BPRS scores improved by 3% on average (rescaled from the lowest score of 0) from baseline to follow-up. Thirty of the 58 EP participants showed greater than 20% improvement in total BPRS score from baseline to follow-up and were classified as Improvers, based on a prior working group criterion used to distinguish treatment responders from nonresponders.¹⁹ Improvers were more likely to be male and had higher baseline BPRS scores than Non-Improvers.

Table 1.

Demographic and Clinical Information for Participants, Excluding Individuals Without Follow-up Data and Those Who Were Censored for Having >50% of Their fMRI Scans Scrubbed

	HC Mean or Count (SD)	EP Mean or Count (SD)	HC vs EP t or χ2 (P)	Improvers Mean or Count (SD)	Non-Improvers Mean or Count (SD)	Improvers vs Non-Improvers t or χ2 (P)
N BD/DD/MDD/PNOS/SZ	—	12/1/2/2/41	—	3/1/1/0/25	9/0/1/2/16	7.92 (.095)
Age (y)	20.13 (3.70)	20.04 (3.76)	0.15 (.89)	20.80 (3.73)	19.22 (3.68)	1.62 (.11)
Biological sex (N), male/female	50/36	35/23	0.07 (.79)	22/8	13/15	4.38 (.036)
Duration of illness (d)	—	298.85 (164.08)	—	301.22 (169.92)	296.22 (160.52)	0.11 (.91)
Days to follow-up	—	340.8 (91.9)	—	328.53 (79.08)	353.96 (103.71)	−1.05 (.30)
Baseline (N), medicated/unmedicateda	—	48/10	—	7/23	3/25	1.62 (.20)
Follow-up (N), medicated/unmedicateda	—	45/13	—	9/21	4/24	2.06 (.15)
Baseline antipsychotic CPZ-equivalent dose	—	261.54 (239.60)	—	287.99 (190.91)	237.20 (278.79)	0.73 (.47)
Follow-up antipsychotic CPZ-equivalent dose	—	253.19 (207.32)	—	273.50 (211.44)	235.41 (206.49)	0.61 (.55)
Baseline BPRS total scoreb	—	19.62 (11.57)	—	24.87 (11.55)	14.00 (8.73)	4.02 (<.001)
Follow-up BPRS total scoreb	—	14.52 (8.41)	—	11.73 (8.29)	17.50 (7.59)	−2.76 (.008)
N Improvers/Non-Improvers	—	30/28	—	30	28	—

Note: BD, bipolar disorder; BPRS, Brief Psychiatric Rating Scale; CPZ, chlorpromazine; DD, delusional disorder; EP, early psychosis; fMRI, functional magnetic resonance imaging; HC, healthy control; MDD, major depressive disorder; PNOS, psychosis-not-otherwise-specified; SZ, schizophrenia spectrum disorder. Duration of illness data was missing for 1 Non-Improver.

^aWith antipsychotics.

^bRescaled to a lowest score of zero.

HC vs EP fMRI Data Comparison

fMRI quality control data (%frames scrubbed and mean framewise displacement) for HCs vs EPs included in analyses are presented in table 2. EPs had significantly more frames scrubbed and greater framewise displacement vs HCs. 9% and 14% of scans were scrubbed on average for HCs and EPs, respectively. fMRI quality control data were not associated with baseline BPRS or percent change in BPRS scores.

Table 2.

Comparison Between fMRI Scanning Quality Control Measurements (%Frames Scrubbed and Framewise Displacement) as well as Network Connectivity Values Between Healthy Controls (HCs) and Participants With Early Psychosis (EPs)

	HC Mean (SD)	EP Mean (SD)	t (P)
%Frames scrubbed	8.84 (6.76)	13.86 (9.90)	−3.37 (<.001)
Mean framewise displacement (mm)	0.16 (0.05)	0.19 (0.05)	−3.67 (<.001)
Attention within-network connectivity	0.33 (0.08)	0.32 (0.06)	0.43 (.67)
Default mode within-network connectivity	0.48 (0.16)	0.52 (0.18)	−1.25 (.21)
Salience within-network connectivity	0.38 (0.09)	0.38 (0.09)	0.22 (.83)
Attention-default mode network connectivity	0.21 (0.05)	0.22 (0.05)	−1.07 (.29)
Attention-salience network connectivity	0.18 (0.04)	0.19 (0.04)	−1.12 (.27)
Default mode-salience network connectivity	0.22 (0.07)	0.22 (0.07)	−0.32 (.75)

Note: fMRI, functional magnetic resonance imaging.

Connectivity values for HCs and EPs as well as t score comparisons between groups are also presented in table 2. No significant differences were observed in connectivity between HCs and EPs.

Missing Data Analysis

Clinical and fMRI connectivity data for individuals with and without follow-up data are provided in supplementary table 2, as well as the results of t tests comparing these groups. Baseline BPRS and fMRI connectivity values did not significantly differ between individuals with and without follow-up data.

Improver vs Non-Improver fMRI Data Comparison

fMRI quality control data (%frames scrubbed and mean framewise displacement) for Improvers vs Non-Improvers as well as t score comparisons between groups are presented in table 3. No significant differences between Improvers and Non-Improvers were observed in %frames scrubbed or framewise displacement.

Table 3.

Comparison Between fMRI Scanning Quality Control Measurements (%Frames Scrubbed and Framewise Displacement) as well as Connectivity Data for Improvers and Non-Improvers

	Improver Mean (SD)	Non-Improver Mean (SD)	t (P)
%Frames scrubbed	13.50 (10.15)	14.25 (9.79)	−0.28 (.78)
Mean framewise displacement (mm)	0.19 (0.05)	0.20 (0.06)	−0.59 (.56)
Attention within-network connectivity	0.33 (0.06)	0.31 (0.07)	0.84 (.41)
Default mode within-network connectivity	0.56 (0.19)	0.47 (0.17)	1.83 (.073)
Salience within-network connectivity	0.37 (0.09)	0.39 (0.09)	−0.87 (.39)
Attention-default mode network connectivity	0.23 (0.05)	0.20 (0.05)	3.06 (.003)
Attention-salience network connectivity	0.20 (0.04)	0.18 (0.04)	2.53 (.014)
Default mode-salience network connectivity	0.24 (0.07)	0.21 (0.06)	1.84 (.071)

Note: fMRI, functional magnetic resonance imaging.

Connectivity values for Improvers and Non-Improvers are also provided in table 3, with histograms showing frequency distributions for these values in supplementary figure 1. Improvers showed significantly higher attention-default network and attention-salience network connectivity.

Logistic Regression: Full Models

We first tested the hypothesis that within- and between-network connectivity between the ROI-defined attention, salience, and default mode networks (ie, 6 network predictors) would predict Improver status in EP (defined as >20% improvement in total BPRS score at follow-up) using logistic regression. As hypothesized, the overall regression model was significant (χ² = 23.66, Nagelkerke’s R² = 0.45, P < .001). Significant individual predictors included default mode within-network connectivity (B = 0.88, SE = 0.42, odds ratio per unit change = 2.40 [95% CI = 1.05–5.48), P = .037], attention-default mode between-network connectivity (B = 0.96, SE = 0.44, odds ratio per unit change = 2.60 [95% CI = 1.10–6.11], P = .029), and attention-salience between-network connectivity (B = 0.88, SE = 0.39, odds ratio per unit change = 2.42 [95% CI = 1.13–5.20], P = .023). The ROC AUC value was 0.85 (figure 1A) and setting the predictive probability threshold to 0.50 resulted in 78% accuracy with 77% sensitivity and 79% specificity. Neither sex (P = .20), nor duration of illness (P = .83), nor days to follow-up (P = .35) significantly predicted outcome. Furthermore, neither baseline antipsychotic dose (P = .70) nor change in antipsychotic dose (P = .61) significantly predicted outcome. When individuals unmedicated at follow-up were excluded, the model remained significant (χ² = 27.66, Nagelkerke’s R² = 0.61, P < .001).

Fig. 1.

(A) Receiver operating characteristic area under the curve (ROC AUC) for the logistic regression model predicting clinical Improver status in early psychosis using connectivity data as predictors. (B) ROC AUC for the logistic regression model using connectivity data and Brief Psychiatric Rating Scale (BPRS) baseline total score as predictors.

When baseline BPRS score (zero-rescaled z score) was added to the model, it improved fit and explained additional variance (χ² = 38.05, Nagelkerke’s R² = 0.64, P < .001, Δ-2LL = 14.39, P(ΔLL) < .001), as BPRS baseline (z score) was also a significant predictor (B = 1.90, SE = 0.67, odds ratio per unit change = 6.71 [95% CI = 1.80–24.98), P = .005]). The ROC AUC value was 0.92 (figure 1B) and setting the predictive probability threshold to 0.50 resulted in 88% accuracy with 89% sensitivity and 87% specificity.

Stepwise Logistic Regression

Next, we used stepwise logistic regression to determine what connectivity features, if any, explained significant variance beyond the baseline BPRS score. A logistic regression model that included only baseline BPRS score as a predictor of Improver status was significant (χ² = 15.31, Nagelkerke’s R² = 0.31, P < .001). Adding attention-default mode between-network connectivity as a feature significantly improved model fit (χ² = 24.36, Nagelkerke’s R² = 0.46, Δ-2LL = 9.06, P(ΔLL) = .003). Adding default mode-salience between-network connectivity as a feature further improved model fit (χ² = 29.13, Nagelkerke’s R² = 0.53, Δ-2LL = 4.76, P(ΔLL) = .029).

Linear Regression: Full Models

Next, we tested the hypothesis that within- and between-network connectivity between the ROI-defined attention, salience, and default mode networks would predict the percent change in total BPRS score. As hypothesized, the linear regression model was also significant (F = 2.93, R² = 0.26, P = .016). Salience within-network connectivity was a significant predictor in the model (B = 31.43, SE = 10.08, t = 3.12, P = .003). Neither sex (P = .67), nor duration of illness (P = .41), nor days to follow-up (P = .39) significantly predicted outcome. Furthermore, neither baseline antipsychotic dose (P = .24) nor change in antipsychotic dose (P = .67) significantly contributed to the model.

When baseline BPRS score was added to the model, model fit was significantly improved (ΔR² = 0.16, ΔF = 13.70, P(ΔF) < .001) as baseline BPRS score (zero-rescaled z score) was a significant predictor (B = −37.37, SE = 10.10, standardized beta = −0.44, t = −3.70, P < .001).

Stepwise Linear Regression

A linear regression model that included only baseline BPRS score as a predictor of percent change in BPRS score was significant (F = 18.46, R² = 0.25, P < .001). Adding attention-default mode between-network connectivity as a feature significantly improved model fit (ΔR² = 0.10, ΔF = 8.11, P(ΔF) = .006). Adding salience within-network connectivity as a feature significantly further improved model fit (ΔR² = 0.06, ΔF = 5.39, P(ΔF) = .024).

Discussion

As hypothesized, intrinsic functional connectivity predicted clinical improvement in this study in EP. Specifically, the full model (that included all 6 connectivity-based predictors) was significant, and within the model, significant individual predictors included within-network connectivity of the default mode network as well as attention-default mode and attention-salience between-network connectivity. Logistic regression model performance was also quite robust, as 85% concordance (ROC AUC) was achieved using connectivity features alone and 92% concordance was achieved when baseline BPRS score was included. Connectivity also predicted linear change in the percent BPRS score over time. Some connectivity features also significantly improved model fits beyond baseline BPRS scores in stepwise regression, suggesting these features make independent contributions to outcomes beyond baseline BPRS scores. The lack of significant differences in total BPRS score and connectivity values between individuals with and without follow-up data suggests these results were not biased by differences in outcome or predictor values as a function of follow-up availability. Overall, these findings provide evidence that resting-state fMRI can be used to as a predictive tool to identify EP individuals most likely to show clinical improvement after 1 year of CSC, and represent an important addition to a rapidly growing literature suggesting that intrinsic connectivity as measured by resting-state fMRI is associated with treatment response in psychosis (reviewed by Chan et al²⁰ and Mehta et al¹¹; more recent examples since these reviews include Anhoj et al,²¹ Mehta et al,²² Nelson et al,^23,24 and Yang et al²⁵). Indeed, our findings that higher attention-default mode and attention-salience network intrinsic functional connectivity predicted conversion and were significantly higher for Improvers are in conceptual agreement with studies suggesting stronger medial prefrontal cortex (a default network component)-whole brain connectivity^26,27 and dorsal attention network-whole brain connectivity²⁸ predicts symptom improvement in psychosis. Our study, furthermore, adds to this prior work by suggesting that salience network connectivity to these networks may also help predict improvement.

Why might connectivity within/between the salience, attention, and default networks predict clinical outcomes in EP? One hypothesis is that they may reflect levels of dopaminergic activity that may in turn predict response to antipsychotics (which are primarily D2 receptor antagonists; thus, EP individuals with higher dopaminergic activity may be more likely to respond to antipsychotic treatment). Related to the salience network, eg, a 2019 study showed that dopamine release capacity in the ventral striatum was associated with lower connectivity within this network.²⁹ This finding was interpreted as being possibly indicative of a dopaminergic mechanism for abnormal salience processing in treatment-responsive psychosis. Also supporting the idea that salience network connectivity may be associated with dopaminergic activity, relatively high levels of dopamine transporters are expressed in both its insula and cingulate components.^30,31 For attention-default network connectivity, previous work using positron emission tomography and resting-state fMRI in HCs found that increased dopamine synthesis capacity in the midbrain (including the substantia nigra) predicts higher between-network connectivity.³² It is thus possible that EP individuals with higher dopaminergic tone (and thus more likely to be antipsychotic responsive^33,34) will show higher between-network connectivity than those who do not respond to treatment.

Important limitations of our study were the relatively modest sample size and short scanning time (6 min). Because of the short scan time, we used a fairly lenient quality control (QC) threshold of <50% of scans scrubbed for participant exclusion. Despite the strong performance of the regression models, we thus consider our results preliminary, and our findings require replication in independent datasets. Future studies in larger samples may also use machine learning techniques which may be more powerful tools for predictive analytics than the logit function utilized in the present study.^35–37 Additionally, it may be of interest to determine if combining these data with other modalities (eg, cognitive control task fMRI data that also predicts clinical outcome in EP^35,36,38) can enhance accuracy, sensitivity, and specificity to sufficient levels for imaging-based biomarkers (ideally, at least 80% for all 3 performance metrics). A second limitation was that, because this was a naturalistic study, individuals were not excluded based on their antipsychotic medication status at study entry. We did not, however, observe any relationships between dose and clinical change in EP (this may not be surprising as D2 receptor occupancy may vary considerably from person to person at equivalent doses³³). It is also important to note that all unmedicated EP individuals in this study were undergoing other forms of CSC treatment (eg, counseling, support, and cognitive behavioral therapy). We thus felt it appropriate to include these individuals in the sample. Furthermore, even after excluding EPs who were unmedicated at follow-up, the logistic regression model remained significant. A third limitation of our study was that connectivity features were calculated from absolute values as opposed to positive and negative connectivity-based (ie, correlation and anticorrelation-based) features to facilitate interpretability (see Methods). An additional limitation was that the amygdala ROIs had ~2× the number of voxels compared with the other ROIs used to represent the salience network, which may have biased results. Finally, it was somewhat surprising that different features predicted binary vs linear change in BPRS scores and were significant for full but not stepwise regression models (or vice versa). We thus conservatively interpret our findings as suggesting that most of the network connectivity values examined contributed to some degree toward predicting clinical improvement.

Overall, the results of this study suggest that this small set of a priori ROIs that, taken together, capture the intrinsic functional connectivity of the attention, default mode, and salience networks can significantly predict clinical improvement after 1 year of treatment in recent-onset psychosis. Future studies may seek to validate these findings in larger, independent datasets to further establish the utility of these resting-state fMRI-based biomarkers as prognostic indicators in EP.

Supplementary Material

Supplementary material is available at https://academic.oup.com/schizophreniabulletin/.

Funding

J.S. is funded in part by fellowship K01 MH125096 from the National Institute of Mental Health. C.S.C. is funded in part by National Institute of Mental Health grants R01 MH122139 and R01 MH059883.