In a systematic review and meta-analysis published in the journal Molecular Psychiatry, scientists have analyzed recent evidence on the usefulness of single and multimodal neuroimaging techniques in characterizing psychotic disorders.
Psychosis is a distressing disorder associated with structural and functional abnormalities in different parts of the brain. Structural abnormalities associated with psychosis include impaired nerve myelination, increased demyelination, and altered cortical thickness and gray matter volume.
Regarding functional abnormalities, disrupted communications between brain regions and altered strength of neural connections are primarily associated with psychosis.
T1-weighted imaging (T1) and diffusion tensor imaging (DTI) are non-invasive neuroimaging methods to assess structural changes in the brain network. Resting-state functional connectivity (rs-FC) is another non-invasive technique used for functional assessment of the brain network.
In this systematic review and meta-analysis, scientists have compared the effectiveness of different neuroimaging methods in classifying individuals with schizophrenia spectrum disorders from healthy individuals. Moreover, they have evaluated whether statistical, methodological, and demographic information can influence psychosis classification.
Scientists searched various scientific databases to identify studies that applied machine learning classification models to classify individuals with schizophrenia spectrum disorders using neuroimaging data of the whole-brain network. This was the primary inclusion criterion for the meta-analysis. Another inclusion criterion was the presence of internal cross-validation or external dataset validation.
The effectiveness of three neuroimaging methods including T1, DTI, and rs-FC in classifying schizophrenia spectrum disorders was evaluated in the meta-analysis. Studies with internal prediction (within dataset cross-validation) and external prediction (validation in a new dataset) were analyzed separately.
A total of 95 and 93 studies were selected for qualitative and quantitative analyses, respectively. Among these studies, 30 used T1, 9 used DTI, 40 used rs-FC, and 14 used a combination of these methods to classify schizophrenia spectrum disorders.
The study analysis revealed that all tested neuroimaging methods have moderate abilities to predict schizophrenia spectrum disorders. Notably, no significant difference in classification performance was observed between the tested methods. However, considering external datasets, rs-FC was found to outperform T1 in classifying schizophrenia spectrum disorders from health participants.
A high level of heterogeneity was observed across the study findings. Specifically, reported effect sizes within each imaging group appeared asymmetric. This indicates the presence of systematic bias in previous reports of the classification of schizophrenia spectrum disorders.
A separate analysis conducted after removing studies with large heterogeneity indicated similar findings, i.e., no significant differences between methods, except for rs-FC which outperformed DTI when internal datasets were used for classification, and T1 when external datasets were used for classification.
No significant difference in classification performance was observed between single-method (unimodal) and combined-method (multimodal) approaches. Both rs-FC and multimodal approaches showed largely overlapping performance in classifying schizophrenia spectrum disorders.
No significant impact of the statistical, methodological, and demographic information on schizophrenia spectrum disorder classification was observed in the analysis. The analysis of head motion, which is a major confounding factor in neuroimaging analyses, revealed no significant impact on classification performance.
This systematic review and meta-analysis reveal that commonly used structural and functional neuroimaging methods are able to classify schizophrenia spectrum disorders.
The modalities tested exhibited similar performances in terms of classifying psychosis, although in some specific cases, the rs-FC method, which measures functional aspects of the brain network, showed slight advantages in terms of better classification performance compared to structural methods including DTI and T1.
Notably, the study highlights that the multimodal approaches do not have any advantage over unimodal approaches in classifying schizophrenia spectrum disorders and that rs-FC is sufficient to provide the required information for classification.
As mentioned by the scientists, this systematic review included studies with relatively smaller sample sizes than the recommended size. Thus, they could not certainly validate whether multimodal methods provide significant advantages in analyses involving large sample sizes.
In this meta-analysis, only whole-brain network data and rs-FC-based functional connectivity data were included. Thus, the scientists recommend that future systematic reviews should consider evaluating the involvement of specific brain regions as well as the ability of different functional connectivity methods, such as task-based FC and dynamic approaches, in classifying psychosis.