Neurobiological substrates of the positive formal thought disorder in schizophrenia revealed by seed connectome-based predictive modeling

Abstract

Formal thought disorder (FTD) is a core symptom cluster of schizophrenia, but its neurobiological substrates remain poorly understood. Here we collected resting-state fMRI data from 276 subjects at seven sites and employed machine-learning to investigate the neurobiological correlates of FTD along positive and negative symptom dimensions in schizophrenia. Three a priori, meta-analytically defined FTD-related brain regions were used as seeds to generate whole-brain resting-state functional connectivity (rsFC) maps, which were then compared between schizophrenia patients and controls. A repeated cross-validation procedure was realized within the patient group to identify clusters whose rsFC patterns to the seeds were repeatedly observed as significantly associated with specific FTD dimensions. These repeatedly identified clusters (i.e., robust clusters) were functionally characterized and the rsFC patterns were used for predictive modeling to investigate predictive capacities for individual FTD dimensional-scores. Compared with controls, differential rsFC was found in patients in fronto-temporo-thalamic regions. Our cross-validation procedure revealed significant clusters only when assessing the seed-to-whole-brain rsFC patterns associated with positive-FTD. RsFC patterns of three fronto-temporal clusters, associated with higher-order cognitive processes (e.g., executive functions), specifically predicted individual positive-FTD scores (p = 0.005), but not other positive symptoms, and the PANSS general psychopathology subscale (p > 0.05). The prediction of positive-FTD was moreover generalized to an independent dataset (p = 0.013). Our study has identified neurobiological correlates of positive FTD in schizophrenia in a network associated with higher-order cognitive functions, suggesting a dysexecutive contribution to FTD in schizophrenia. We regard our findings as robust, as they allow a prediction of individual-level symptom severity.

Keywords: Formal thought disorder; Machine learning; Neuroimaging.

Fig. 1

Seed regions used for functional connectivity map construction and illustration of the overall feature selection and multivariable predictive modeling procedure. A) We used three clusters as seeds that showed consistently aberrant activation associated with FTD in our previous activation likelihood estimation based meta-analysis (Wensing, 2017). MNI coordinates: left superior temporal gyrus (STG; −54, −28, 4); ventral posterior middle temporal gyrus (vpMTG; −46, −50, 22); and dorsal posterior MTG (dpMTG; −56, −56, 12). B) Flowchart illustrates the identification of rsFC patterns that were robustly associated with FTD specific dimensions, and the use of the identified rsFC as features for predictive modeling. GLM, general linear model; RVM, relevance vector machine; rsFC, resting-state functional connectivity; FTD, formal thought disorder.

Fig. 2

Differential resting-state functional connectivity (rsFC) with the seed regions in schizophrenia patients compared to healthy subjects. A) For the left STG seed, clusters showing decreased (blue) rsFC for schizophrenia patients were located in the left fusiform gyrus and left middle temporal gyrus, while one cluster in the right medio-superior frontal gyrus showed increased rsFC (red). B) For the left dpMTG seed, we identified one cluster indicating increased rsFC in the bilateral dorsal thalamus of schizophrenia patients, which is displayed on both axial sections and a three-dimensional thalamic surface. The above significant clusters were identified at a cluster-level with family-wise error corrected p < 0.05, cluster-forming threshold of p < 0.001 at voxel level. L, left; R, right; MFG, middle frontal gyrus; vpMTG, ventral posterior middle temporal gyrus; STG, superior temporal gyrus; dpMTG, dorsal posterior middle temporal gyrus. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 3

Cross-validation based feature selection and prediction of individual positive formal thought disorder (FTD) scores from resting-state functional connectivity (rsFC). A) For each seed region and each 10-fold cross-validation, clusters whose rsFC to the seeds that were significantly associated with positive-FTD were accumulated. Mean rsFC values extracted within these clusters were used as features. Frequency of the clusters selected in the overall general linear models conducted on the 500 training samples is color-coded from blue to red. The three seed regions are displayed on a transparent brain. Blue lines indicate negative associations while the red line indicates positive association. Partial correlation coefficients (adjusted for age, gender, and head motion parameters) between the identified significant rsFC patterns and the positive FTD scores, derived from the repeated 10-fold cross-validation, were shown in three boxplots separately for the three seeds (red line depicts the median, green diamond depicts the mean, whiskers represent the 5th and 95th percentiles); B) Scatter shows the Pearson’s correlation between the actual (confound-adjusted) positive-FTD scores and their predictions in 10-fold cross-validation within the main sample. The predicted data were the average predictions for each subject across the repeated 10-fold cross-validation runs. The regression line in blue was fitted with the suspected “outlier” (i.e., the circled blue point), while the dark yellow line was fitted after excluding this “outlier”; C) Brain clusters with rsFC to the seeds that were identified as significantly associated with positive FTD scores in leave-one-site-out cross-validation within the main sample. Scatter plots show the correlation between the actual positive-FTD scores and their predictions. The data points were colored differently per site. The regression line in blue was fitted with the suspected “outlier” (i.e., the circled blue point), while the dark yellow line was fitted after excluding this “outlier”; D) Our RVM model, that trained within the main sample based on the rsFC patterns extracted from the three most robustly identified brain clusters, was applied to the independent sample for validation. Scatter plot shows the correlation between the actual positive-FTD scores and their predictions in the independent sample. Shaded areas represent 95% confidence intervals. L, left; R, right; FC, functional connectivity; AT, abstract thinking; FTD, formal thought disorder; STG, superior temporal gyrus; dpMTG, dorsal posterior middle temporal gyrus; vpMTG, ventral posterior middle temporal gyrus; MFG, middle frontal gyrus; SMG, supramarginal gyrus; ITG, inferior temporal gyrus; IPL, inferior parietal lobule. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 4

Functional characterization of the identified brain clusters. Bar charts indicating A) functional profiles of the clusters showing differential resting-state functional connectivity (rsFC) with the two seeds: superior temporal gyrus and dorsal posterior middle temporal gyrus between schizophrenia patients and healthy controls, and B) functional characterization of the brain clusters whose rsFC to the seeds were identified as robustly associated with the severity of positive FTD. Quantitative “forward inference” and “reverse inference” experiments were used to determine the functional profile of each cluster. Regions with significant functional assignments (false discovery rate corrected p < 0.05) are presented. The significant activation likelihood ratios for each cluster with respect to a given domain or paradigm (forward inference) and the significant probability of a domain's or paradigm's occurrence given activation in a cluster (reverse inference) are depicted separately. The “P (Activation I Domain/Paradigm)” refers to the activation likelihood (in the forward reference) in a significant region given a particular label (i.e., a behavioral domain or a task paradigm). The “P (Domain I Activation)” refers to the probability of behavioral domains given activation in a particular brain region, which tests for a regions’ functional profile based on reverse inference. Similarly, the “P (Paradigm I Activation)” represents the probability of paradigm classes given activation in this particular region. FTD, formal thought disorder; LH, left hemisphere; RH, right hemisphere.