Introduction

Schizophrenia is a severe and complex mental disorder marked by delusions, hallucinations, and cognitive impairments. Functional magnetic resonance imaging (fMRI) has helped identify the underlying neural mechanisms of these core illness features1,2. Viewed broadly, prior studies have indicated that schizophrenia is not caused by impairments of a small number of brain regions but by abnormal functional connections among widely distributed brain regions3,4,5. Much of our understanding of connectivity deficits in schizophrenia has been acquired from studies of static or average functional connectivity (FC) among brain regions. Because the functional organization of brain circuits is highly temporally dynamic, involving shifts among discrete brain states, investigating the dynamic changes may provide novel complementary information about abnormal brain neurophysiology in schizophrenia6.

Measures of dynamic functional connectivity (dFC) characterize the dynamic temporal patterns of brain connectivity over time. This is important, as it is the dynamic changes in activity states that are important for adaptive cognitive, affective, and behavioral processes. Dynamic connectivity studies have found that individuals with schizophrenia spend less time in a strongly and widely connected state (typified by strong, largescale, positive connectivity), correspondingly stay more in a weakly connected state (characterized by generally low strength of functional connectivity), and switched less frequently among discrete states in comparison with healthy controls (HCs)7,8,9,10. Patients have been reported to show increased dynamic variability within and between sensorimotor, visual, attention, and thalamic networks, and decreased intra-network and inter-network dynamic reconfigurations of default mode and fronto-parietal networks11. Another study reported increased dynamics of cross-network interactions among default mode network (DMN), salience network (SN), and central executive network (CEN) in schizophrenia12. These temporal and functional neuroimaging findings highlight the novel insights that can be obtained by investigating abnormal brain functional dynamics in schizophrenia.

Dynamic topology properties, which could describe quantitative topologies of brain regions in the context of complex whole brain systems based on graph theory13, have further provided useful information about the temporal dynamics of brain region interaction at a whole brain scale. One study using this type of network analysis reported an increment of dynamic nodal efficiency in the left frontal, right medial parietal, and bilateral subcortical areas in patients with schizophrenia14. In terms of whole-brain dynamic topological characteristics, reduced temporal fluctuation of clustering coefficient and global efficiency with impaired interaction among DMN subsystems have been reported in schizophrenia15,16, suggesting an important role of DMN in the pathophysiology of schizophrenia. Emerging research also indicates that dynamic connectivity alterations may be relevant for clinical symptoms in schizophrenia8,14,17,18.

Most studies investigating dFC have studied long-term ill and treated patients. This may be a limitation because of the potential confounding effects of antipsychotic medication and long-term illness on the brain8,9,10,11,12,14,15,17,19, which have been found to contribute to the dynamic changes observed over the progression of schizophrenia20,21. We are aware of only one study of acutely ill untreated patients, which recruited a relatively small group of first-episode, drug-naive patients to explore altered dFC strength in the mirror neuron system18. Larger studies of untreated first-episode patients are needed to evaluate dFC during acute psychotic illness without confounding the effects of drug treatment22.

We recruited acutely ill, never-treated, first-episode patients with schizophrenia and examined brain dynamic functional activity patterns relative to HCs. We hypothesized that: (1) patients would more often stay in a weakly connected state and (2) there would be an abnormal dynamic topology of regional brain properties in schizophrenia. Based on previous resting-state network meta-analysis4,23, we predicted alterations in DMN11,12,16, ventral attention network (VAN) engaged in processing of salience24, visual network (VN) associated with visual processing25, and sensorimotor network (SMN) involved in sensory and auditory perception26. We also predicted that altered temporal properties would correlate with patients’ clinical symptom severity. In addition, to evaluate dFC features for understanding illness biology at the level of individual patients, we conducted a supplementary machine learning study to identify features that most consistently were useful for identifying schizophrenia patients.

Results

dFC state characteristics

There were 95 never-treated first-episode schizophrenia and 100 matched HCs scanned and included in analyses (Table 1). Our first analysis was conducted to empirically define identifiable discrete brain states to analyze state change metrics. Two recurring brain functional states were defined by cluster analysis based on the similarity of the dFC correlation matrices. Their centroids are shown in Fig. 1 based on a window size of 22 times of repetition (TR). State 1 was characterized by weak connectivity both within and between networks, with negative couplings between DMN and other networks and between fronto-parietal task control network (FPN) and SMN, VN, cingulo-opercular task control network (CON), and auditory network (AN). State 2 had stronger and mostly positive connectivity in general, including strong and positive dFCs within and between SMN and VN.

Table 1 Demographic and clinical characteristics of patients with schizophrenia and healthy controls.
Fig. 1: Cluster centroids of reoccurring dynamic functional connectivity (FC) patterns in the main analysis.
figure 1

a Cluster centroids for each state across all participants. b Cluster centroids for each state for each group. The value of each cell in the FC matrix is the Pearson correlation coefficient between two brain regions. The color bar shows the strength of FC between two nodes (warm color, positive FC; cool color, negative FC). DMN default mode network, VN visual network, SMN_hand/mouth sensorimotor hand and mouth networks, FPN fronto-parietal task control network, SN salience network, CON cingulo-opercular task control network, AN auditory network, DAN dorsal attention network, VAN ventral attention network, MN memory retrieval network, SZs schizophrenia patients, HCs healthy controls.

The frequency of occurrence of the hypoconnected state 1 (71.3%) was higher than that of hyperconnected state 2 (28.7%) across all participants (Fig. 1a). The fractional time of brain function in state 1 was significantly higher in patients relative to controls after false discovery rate (FDR) correction (patients 75.6 ± 25.1%, HCs 67.2 ± 27.0%, FDR-P = 0.017). Accordingly, patients spent less time in state 2 of high connectivity compared with controls (patients 24.4 ± 25.1%, HC 32.8 ± 27.0%, FDR-P = 0.017). We also found that patients demonstrated longer mean dwell time in state 1 (patients 61.5 ± 52.7, HCs 41.0 ± 35.0, FDR-P = 0.002) and reduced mean dwell time in state 2 relative to HC (patients 12.2 ± 10.2, HCs 16.6 ± 15.6, FDR-P = 0.017). The number of transitions between states was lower in schizophrenia patients than HCs (patients 5.2 ± 4.3, HCs 6.2 ± 3.7, FDR-P = 0.044) (Fig. 2a). No significant correlations between these global state characteristics and clinical ratings were observed in patients. In addition to the primary analysis using 22TR window sizes, we did a secondary analysis with 30TR window sizes to show that our findings were not window length dependent. This ancillary validation yielded similar findings (details in Supplementary materials and Figs. S2, S3).

Fig. 2: Altered dynamic properties between groups and correlations between dynamics and symptoms in patients.
figure 2

a Patients with schizophrenia (SZs) showed altered fractional time and mean dwell time in both states, and decreased transitions between states relative to healthy controls (HCs). b The coefficient of variation (CV) of eigenvector centrality (EC) of the right medial prefrontal cortex (mPFC) was decreased in SZs than HCs and was negatively correlated with psychosis symptom severity in patients. * Indicates false discovery rate corrected P value < 0.05.

Dynamic topological properties

Relative to controls, patients had lower coefficients of variation (CVs) of nodal efficiency and eigenvector centrality (EC) in the right medial prefrontal cortex (mPFC, part of DMN) and higher CVs of nodal EC in the left middle cingulate gyrus, left postcentral gyrus, right inferior temporal gyrus, bilateral paracentral lobules, and bilateral middle temporal gyrus, most of which are in SMN (Table 2). The CV of nodal EC in right mPFC in patients was negatively correlated with the total ratings (r = −0.29, P = 0.005) and general symptom ratings (r = −0.21, P = 0.048) from the positive and negative syndrome scale (PANSS) (Fig. 2b). There were no significant differences in temporal dynamics of global efficiency and cluster coefficient between groups. Most results were replicated using a larger (30 TR) sliding window (details in supplementary materials and Table S2).

Table 2 Case-control differences in coefficients of variation (CV) of nodal efficiency and eigenvector centrality.

Machine learning analysis

To verify the utility of dFC features for case identification at the individual level, F values were used to select features from temporal properties to include in the machine learning analysis, and a linear support vector classifier (SVC) was applied for individual classification. We gradually increased the number of features to find a balance between establishing an efficient and robust classifier while avoiding including too many features which could lead to overfitting (the number of features included varied from 1% to 100% of all temporal properties, with one percent increments). The SVC classifier showed 72.3% accuracy (area under the curve = 0.81) when 3% features were included, with 70.5% sensitivity and 74.0% specificity. The features selected every time across leave-one-out cross-validation (LOOCV) are shown in Table 3. The performance rate of classification was significantly higher than a randomly permuted group (P = 0.001). Almost all features included also showed significant between-group differences in the primary group comparisons (Table 3). The accuracy of the classification model is not high enough for clinical practice but is useful in showing that the modeled abnormalities together reflected commonly present brain features at the individual patient level.

Table 3 Features selected for classification model and the weights of these features in case-control classification.

Lastly, in a subgroup of patients with follow-up data, the regression model with temporal properties contributing to the classification at baseline predicted a decline of PANSS total ratings over the course of treatment (R2 = 0.43, P < 0.01) at 6-week follow-up. Also, we observed a high accuracy (81.0%) of pretreatment data for differentiating short-term medication treatment responders/non-responders (details in supplementary materials and Tables S4S9).

Discussion

Our findings document the temporal and regional pattern of altered dynamic functional brain connectivity present in acutely ill, first-episode, never-treated patients with schizophrenia. Schizophrenia patients tended to remain in a weakly connected state (state 1) and spend less time in a more highly connected state (state 2). These findings suggest that previously observed static FC deficits in schizophrenia27 may reflect a reduced rate of shifting into the actively connected brain state as was observed in the present study.

Topological analyses found that schizophrenia patients demonstrated decreased functional dynamics in mPFC that were related to illness severity, and increased dynamics in temporal lobe and sensorimotor regions. Additionally, a machine learning analysis indicated that the observed pattern was commonly observed in individual schizophrenia patients. Importantly, these observations were present in acutely ill early-course patients prior to medication treatment, so that drug effects and course of illness effects did not confound the evaluation of illness-related brain features during acute psychosis.

dFC states

The cluster analysis of dFC features identified two reoccurring dFC states representing discrete brain states of participants during scanning28. Compared with controls, patients with schizophrenia displayed more occurrences and dwelled longer in state 1 characterized by negative connectivities and relatively weak positive connectivities. The rate of transition between the states was also reduced in patients. These findings indicate that patients with schizophrenia demonstrate an atypical temporal distribution of functional brain states, with patients spending less time in a highly connected functional state and longer time in a weakly connected state.

Analysis of mean dwell time in states showed that patients with schizophrenia not only stayed in a weak state most of the time but also more quickly transitioned back away from the strongly connected state when it was engaged. Schizophrenia has been conceptualized as a disorder of decreased connectivity that is believed to be related to symptoms and cognitive impairment27,29,30. Our results indicate that the repeatedly observed reduction in average inter-region FC in schizophrenia27 is in part a result of more frequently staying in an “idling” state (state 1) rather than heightened connectivity in task-active networks8. Moreover, lower rates of switching into the more densely connected state, and faster exiting from that state when it was achieved, suggest resistance to shifting into more energy-demanding highly connected states required by task-active networks as previously reported9 that could contribute to the alterations of behavior and cognition associated with schizophrenia.

Dynamic nodal topology

Decreased nodal dynamics in mPFC

The mPFC node of the DMN exhibited reduced time-varying nodal efficiency and EC. As a highly centralized hub region, the mPFC integrates information from abundant cortical and subcortical areas and assembles updated information transferred to multiple brain regions60 and such window sizes have typically been used in dFC studies43,61. During the whole scanning period, there are 169 dFC matrices per person when there were 190 brain scanning volumes and 22TR window sizes and the step was 1TR (details in Supplementary methods).

dFC state analysis

To identify discrete reoccurring patterns of dFC states across subjects, we performed k-means clustering with Scikit-Learn (version 0.23.1) on Python (version 3.8.3) on a series of 264 × 264 FC matrices for all participants. The similarity between each windowed FC matrix was estimated using the Euclidean distance. To identify the common and robust discrete categorical data organization, the k value was varied from two to ten to search the number of such groups of events in the data before clustering. We used the “elbow criteria” to find the inflection of the sum of the squared errors and applied the maximum silhouette coefficient together to determine the optimal k value. Both methods which aim to identify the number of types of overall data organization (here, types of brain states) indicated the best k to be two in the present study (Fig. S1), indicating that there were two types of functional brain states in the data. In primary analyses, we used the cluster centroids of all participants to represent the two reoccurring dFC states (Fig. 1a). For visualization of dFC states in the two groups respectively, we also calculated the group-specific cluster centroids (Fig. 1b). To examine temporal properties of the two dFC states, we assessed three different state characteristics: (1) fractional time, the proportion of time window belonging to each state; (2) mean dwell time, the average length of consecutive time windows spent in each state; and (3) number of transitions, the number of switches between states over time.

Dynamic topological analysis

The analysis of dynamic functional topology can characterize how all brain regions work together and delineate anatomic and functional sources of aberrant dFC14,36. To avoid a different number of edges caused by weak spurious FCs and reduce the computational complexity in the assessment of topological architecture, the top K% (K = 5–20 with an interval of 5) strength of the positive FCs for each 264 × 264 FC matrix were kept to obtain a series of sparse weighted matrices. To represent global and regional characteristics of topological networks, global efficiency, clustering coefficient, nodal efficiency, and nodal EC were computed for each sliding window for each person at each sparsity using the Brain Connectivity Toolbox (nitrc.org/projects/bct, version 2019-03-03). A detailed interpretation of these topological characteristics is presented in Table S1. The area under the curve (AUC) measures across the range of sparsity thresholds were obtained to ensure the robustness of the four topological characters. The CV of the AUC of all topological metrics was calculated across all time windows to characterize temporal variability for all subjects. High CVs reflect high fluctuation rates of topological metrics, and low CV values reflect more stable states. To assess the consistency and validity of the dynamic analysis at different window sizes, we used another sliding window parameter (30TR window width and 2TR step length) to repeat the above analyses and validate the results. The results of findings from the longer window did not meaningfully differ from those of the primary analysis. The results reported in the main text are under 22TR window sizes, and the results of 30TR were reported in the supplementary materials.

Statistical analysis

All neuroimaging statistical analysis was performed in MATLAB R2017b. Permutation testing iterated 10,000 times was conducted to test for group differences due to the non-normal distribution of state characteristics and dynamic topological metrics (details in supplementary methods). Significance was set at P < 0.05 with FDR correction for multiple comparisons. In patients, we performed Spearman partial correlations to explore relations of dynamic properties (state characteristics and CVs of topological metrics showing intergroup differences) with acute illness severity (PANSS ratings), treating demographic measures (age, sex, and education years) as covariates. Nominal significance was set at P < 0.05 for these correlation analyses conducted for heuristic and descriptive purposes.

Classification based on dynamic properties

We next identified the brain features that most robustly and consistently distinguished individual schizophrenia patients from HCs using a linear SVC. The feature selection by F score was performed instead of directly using the previous features with significant between-group differences to avoid data leakage from the training set to the test set and reduce the risk of overfitting. After that, SVC was trained by the selected features. LOOCV was performed to evaluate model generalization. Permutation testing was conducted to test the performance of the classifier.

In an additional exploratory analysis with a subsample of 21 patients followed clinically after 6 weeks of antipsychotic treatment (Table S3), we determined whether the dynamic properties contributing to the classification model could predict clinical treatment response with support vector regression. The prediction of short-term treatment response procedures is described in detail in supplementary materials.