
==== Front
Hum Brain Mapp
Hum Brain Mapp
10.1002/(ISSN)1097-0193
HBM
Human Brain Mapping
1065-9471
1097-0193
John Wiley & Sons, Inc. Hoboken, USA

10.1002/hbm.70030
HBM70030
Research Article
Research Article
Changes in the structure of spontaneous speech predict the disruption of hierarchical brain organization in first‐episode psychosis
He et al.
He Rui https://orcid.org/0000-0001-6999-1920
1 rui.he@upf.edu

Alonso‐Sánchez Maria Francisca 2
Sepulcre Jorge 3
Palaniyappan Lena 4 5 6
Hinzen Wolfram 1 7
1 Department of Translation and Language Sciences Universitat Pompeu Fabra Barcelona Spain
2 CIDCL, Escuela de Fonoaudiología Universidad de Valparaíso Valparaíso Chile
3 Department of Radiology and Biomedical Imaging, Yale PET Center, Yale School of Medicine Yale University New Haven Connecticut USA
4 Douglas Mental Health University Institute, Department of Psychiatry McGill University Montreal Quebec Canada
5 Department of Medical Biophysics, Schulich School of Medicine and Dentistry Western University London Ontario Canada
6 Robarts Research Institute, Schulich School of Medicine and Dentistry Western University London Ontario Canada
7 Intitut Català de Recerca i Estudis Avançats (ICREA) Barcelona Spain
* Correspondence
Rui He, Department of Translation and Language Sciences, Universitat Pompeu Fabra, Carrer Roc Boronat, 138, Barcelona 08018, Spain.
Email: rui.he@upf.edu

20 9 2024
10 2024
45 14 10.1002/hbm.v45.14 e7003029 4 2024
01 2 2024
04 9 2024
© 2024 The Author(s). Human Brain Mapping published by Wiley Periodicals LLC.
https://creativecommons.org/licenses/by/4.0/ This is an open access article under the terms of the http://creativecommons.org/licenses/by/4.0/ License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited.

Abstract

Psychosis implicates changes across a broad range of cognitive functions. These functions are cortically organized in the form of a hierarchy ranging from primary sensorimotor (unimodal) to higher‐order association cortices, which involve functions such as language (transmodal). Language has long been documented as undergoing structural changes in psychosis. We hypothesized that these changes as revealed in spontaneous speech patterns may act as readouts of alterations in the configuration of this unimodal‐to‐transmodal axis of cortical organization in psychosis. Results from 29 patients with first‐episodic psychosis (FEP) and 29 controls scanned with 7 T resting‐state fMRI confirmed a compression of the cortical hierarchy in FEP, which affected metrics of the hierarchical distance between the sensorimotor and default mode networks, and of the hierarchical organization within the semantic network. These organizational changes were predicted by graphs representing semantic and syntactic associations between meaningful units in speech produced during picture descriptions. These findings unite psychosis, language, and the cortical hierarchy in a single conceptual scheme, which helps to situate language within the neurocognition of psychosis and opens the clinical prospect for mental dysfunction to become computationally measurable in spontaneous speech.

Psychosis involves alterations in the entire cognitive architecture, as organized hierarchically in the cortex from lower‐order sensation to higher‐order cognitive processes like language. Structural changes in spontaneous language have long been documented in psychosis, but have not been related to neurofunctional changes. This work documents alterations in the cortical hierarchy in the first‐episodic psychosis and demonstrates their relation to altered language structures as manifested in spontaneous speech during picture descriptions. Neurofunctional changes affected both the hierarchical distances between sensorimotor and default mode networks, and the hierarchical organization of the semantic network. Moreover, these changes were predicted by graphs representing semantic and syntactic associations between meaningful units. Our findings show the potential of computational measurements of mental disorders via spontaneous speech.

functional gradient
natural language processing
psychosis
semantic network
spontaneous speech
European Research Council 10.13039/100010663 ERC‐2023‐SyG 101118756 Guangdong Provincial Department of Science and Technology 10.13039/501100007162 112175605105 China Scholarship Council 10.13039/501100004543 202108390062 Agencia Nacional de Investigación y Desarrollo 10.13039/501100020884 74200048 Canadian Institutes of Health Research Foundation GrantFDN 154296 Canada First Excellence Research FundInnovation fund for Academic Medical Organization of Southwest OntarioBucke Family FundThe Chrysalis Foundation 10.13039/100002734 The Children Hospital FoundationThe Arcangelo Rea Family Foundation source-schema-version-number2.0
cover-dateOctober 2024
details-of-publishers-convertorConverter:WILEY_ML3GV2_TO_JATSPMC version:6.4.8 mode:remove_FC converted:20.09.2024
He, R. , Alonso‐Sánchez, M. F. , Sepulcre, J. , Palaniyappan, L. , & Hinzen, W. (2024). Changes in the structure of spontaneous speech predict the disruption of hierarchical brain organization in first‐episode psychosis. Human Brain Mapping, 45 (14 ), e70030. 10.1002/hbm.70030
==== Body
pmc Practitioner Points

Neurofunctional changes in the first‐episodic psychosis affected both the hierarchical distances between sensorimotor and default mode networks, and the hierarchical organization of the semantic network.

These changes were associated with graphs representing semantic and syntactic associations between meaningful units.

1 INTRODUCTION

Cognitive impairment is a well‐established feature of schizophrenia, which is seen as early as premorbid stages and first‐episode psychosis (FEP) and persists in the chronic phase of the illness (Gebreegziabhere et al., 2022). No single cognitive domain appears to be fully spared in this illness, suggesting changes in the entire cognitive architecture from lower to higher‐order functions (Gold et al., 2009; McCutcheon et al., 2023). Bottom‐level sensorimotor disturbances, affecting a range of primary cortical functions such as audition and vision (Butler et al., 2008; Dale et al., 2016; Javitt, 2009; Javitt & Sweet, 2015), may propagate upstream when integrative higher‐order functions are employed (Dong et al., 2021). The implementational architecture for such integration has recently been studied with a nonlinear decomposition technique, diffusion mapping, which represents the very high‐dimensional functional connectivity (FC) patterns across the cortex through a small number of continuous low‐dimensional “gradients” (Huntenburg et al., 2018). Such gradients are considered to represent functional continuity and its variability across fundamental axes of organization of the human brain (Sydnor et al., 2021).

The first or principal gradient (G1) captures a cortical hierarchy ranging from functionally specialized sensorimotor cortex to domain‐general association cortex, specifically including regions of the default mode network (DMN). Along this hierarchy, information is processed in increasingly abstract multimodal, heteromodal, and transmodal representational formats, forming a substrate for integrated higher‐order cognition including language (Mesulam, 1998). The second gradient (G2) captures the functional separation of different primary sensory cortices, with visual cortex at one end and somatosensory cortex at the other. Gradients as proxies of hierarchical information streams can be further characterized with stepwise FC (SFC), a graph‐based method to detect both direct and indirect functional links emanating from specific regions in the brain. Cortex‐wise SFC, starting from sensory sources and culminating in the cortical hub of default mode regions (Sepulcre et al., 2012), has already offered insight into the “bottom‐up” dysregulation of the hierarchy in schizophrenia, complementing gradient analysis (Dong et al., 2021).

The DMN is conceived as a dynamic “sense‐making” network, in which extrinsic incoming information is integrated with prior intrinsic information, so as to construct rich situation models with narrative meaning, which unfolds over time (Yeshurun et al., 2021). This conception suggests that the particular hierarchy captured by G1 may functionally relate to meaning‐making as well, that is, the building of higher‐order meaning from perception, in which language plays a role. We hypothesized that a disrupted topological architecture of the cortex in psychosis (Dong et al., 2021) would be reflected in structural language changes, which have also long been observed in psychosis (Corcoran et al., 2020; Corona Hernández et al., 2023).

Specifically, speech production necessarily involves retrieving concepts internally from semantic memory. Out of context, such concepts associate with each other at certain semantic distances, with cat being relatively more similar semantically to kitten than to house. These distances can be computationally quantified through cosine similarities between word embeddings retrieved from large neural language models (LMs) (Baroni, 2013). The meaning of language, however, depends on relating words grammatically, whereby they become parts of phrases (e.g., a noun phrase [NP]), and these phrases become parts of other phrases (e.g., a verb phrase [VP]), giving rise to a syntactic hierarchy. At this level, language can encode full interpretations of perceptual experience, as in the episodic thought It was the cat, which will typically encode reference to a specific event in space and time and form part of a narrative. Both lexical‐semantic associations and syntax can be studied through graphs, using advanced natural processing tools. We reasoned that a disruption in cortical gradients could be reflected in the graph‐theoretic structure of meaning in language, during a simple speech task such as describing a picture. While numerous studies have used semantic distances between words (Alonso‐Sánchez, Ford, et al., 2022; Elvevåg et al., 2007; He et al., 2024; Voppel et al., 2023) as well as syntactic complexity (Ciampelli et al., 2023; Schneider et al., 2023) to quantify language changes in psychosis, relations between these and neurofunctional changes have barely been explored (but see Alonso‐Sánchez et al., 2023; Alonso‐Sánchez, Limongi, et al., 2022).

We predicted alterations in meaning structure should bear on large‐scale cortical gradients and SFC in psychosis. This is because language not only plays a crucial role in the creation of meaning but is also inherently integrative of cognition as whole: any utterance we make requires integration with perception, fine motor control, semantic memory, attention, and social cognition. Empirical support for this expectation comes from the fact that interindividual differences in the geodesic distance of sensorimotor landmarks to the language and DMNs covary with activations in these networks during language tasks (Wang et al., 2022). Furthermore, greater separation between the DMN and sensorimotor network is related to semantic cognition, especially our ability to retrieve links between concepts from semantic memory (Shao et al., 2022). In this context, the semantic brain network, which includes essential left‐hemispheric default mode and limbic regions (Binder et al., 2009), is of particular interest, given its reported associations with language disruptions in psychosis (Matsumoto et al., 2023; Pintos et al., 2022). The results of our study provide, for the first time, speech graph alterations related to aberrations in the large‐scale cortical gradients with specific involvement of the semantic network in psychosis. The graphs reveal a semantic space altered both in terms of its lexical‐conceptual semantic associations between individual units of meaning and of the syntactic structures that enable meaning at the sentential and narrative level. This finding contributes to our understanding of the functional significance of neurocognitive changes in psychosis and highlights spontaneous speech as a readout of these changes and a potential source for their clinical measurement.

2 RESULTS

2.1 Altered cortical gradients and SFC

Figure 1a depicts the workflow of the fMRI analysis. Then, 29 untreated first‐episode psychosis (FEP) patients, matched to 29 healthy controls (HC) on age, gender, and education, participated in this study (see Table 1). Subjects underwent resting‐state fMRI scans on a Siemens 7 T Plus (Erlangen, Germany) and were asked to describe three pictures from the Thematic Apperception Test (1 min for each picture) immediately prior to the scans. FC matrices were extracted from 1000 regions‐of‐interest (ROIs) using the Schaefer 1000‐parcellation (Schaefer et al., 2018) for each subject and group. Affinity matrices were computed from these, where affinity values between any pair of regions are the cosine similarities between the vectors representing these regions in terms of their FC to all other cortical regions. Diffusion embedding was then applied to these affinity matrices to extract 10 group‐level gradients from these affinity matrices, resulting in 10 gradient values (i.e., affinities) assigned to each parcel (Vos de Wael et al., 2020). Group templates for all gradients with scree plots showing the variance explained by each gradient, as well as those for SFC analyses, are available in Figures S12–S24. Subject‐level gradient templates were aligned to their corresponding group templates. Gradient values of ROIs were compared between HC and FEP using surface‐based linear models (SLM).

FIGURE 1 Workflow of this study. To determine the cortical hierarchy, we carried out gradient analysis in the whole cortex and within the semantic subnetwork, with gradient dispersion derived from it, and stepwise functional connectivity (SFC) to better interpret gradient results. The illustration graph for SFC is adapted from Sepulcre et al. (2012) For linguistic organization of spontaneous speech, we computed formal syntactic measures and lexical‐conceptual semantic graph‐theoretical measures, the latter of which capture shortest paths, clustering coefficients, and the balance between them (small world organization). The visualization of these graph‐theoretical measures is adapted from Rubinov and Sporns (2010).

TABLE 1 Descriptive statistics of the dataset.

	HC	FEP	Test	p	Effect size	
Num	29	29	/	/		
Age	21.5 (3.6)	21.4 (2.2)	Welch's T test	.861	−0.046	
Sex	24.1%	27.6%	χ 2 test	.764	0.039	
Education level	71.4%	51.7%	χ 2 test	.127	0.198	
Parental SES	58.0%	45.0%	χ 2 test	.139	0.196	
P1: Delusions	/	5 (2)	/	/	/	
P2: Conceptual disorganization	/	3 (3)	/	/	/	
P3: Hallucinatory behavior	/	5 (1)	/	/	/	
N1: Blunted affect	/	2 (3)	/	/	/	
N4: Passive/apathetic social withdrawal	/	3 (4)	/	/	/	
N6: Lack of spontaneity and flow of conversation	/	1 (2)	/	/	/	
G5: Mannerisms and posturing	/	1 (2)	/	/	/	
G9: Unusual thought content	/	4 (2)	/	/	/	
TLI Impoverishment of thought	0.15 (0.26)	0.52 (0.60)	Welch's T test	.005	0.796	
TLI Disorganization in thought	0.18 (0.26)	1.09 (1.12)	Welch's T test	<.001	1.120	
Total TLI score	0.30 (0.30)	1.51 (1.30)	Welch's T test	<.001	−1.226	
Note: Age was indicated by mean ± standard deviation with effect size indicated by Cohen's d. Sex: proportion of female subjects. Education level: proportion of subjects receiving formal education of over 12 years. Parental social economic state (SES): proportion of subjects with parental SES less than 3. TLI: thought and language index, indicated by mean ± standard deviation. Effect sizes for group comparison on sex and education were indicated by contingency coefficient. Clinical scores for FEP were indicated by median (IQR, interquartile range). Statistical tests within this table were carried out with JASP 0.16.4.

Figure 2 summarizes the gradient findings. Figure 2a shows an overall (global) gradient compression in FEP, with a smaller range and a higher peak, suggesting a decreased functional distance between the endpoints of cortical hierarchies. More specifically, when plotting gradient values of G1 and G2 together, both groups show an expected geometry (see Figure 2b), with sensorimotor networks maximally separated from the DMN in the first gradient (G1, vertical axis), and the visual network (VN) and somatomotor network (SMN) maximally separated from each other in the second gradient (G2, horizontal axis). However, in G1, the functional distance between the DMN on top and the VN at the bottom decreases in FEP, while that of the DMN and the SMN increases. Moreover, in G2, the dispersion between VN and SMN is maintained, while, between them, the DMN is less distant from the SMN and more distant from the VN. Group templates of all 10 cortical gradients can be found in Figures S12–S15. In line with this pattern, Figure 2c shows that for G1 the gradient values of ROIs in VN significantly increased in FEP while those in SMN significantly decreased. For G2, gradient values of ROIs in the two extremes, VN and SMN, were unchanged, while gradient values in DMN significantly decreased in FEP.

FIGURE 2 Summary of changes in cortical functional hierarchy. (a) Histogram with Global distribution of all 10 cortical gradients, with a density line estimated by kernel density estimator (KDE) and arrows showing changes in both the peaks and bilateral ends of the density lines in first‐episode psychosis (FEP) relative to health controls (HC). Red: HC; gray: FEP. (b) Scatterplots of the first (y‐axis) and the second (x‐axis) gradient in HC (left) and FEP (right). The color coding refers to Yeo's 7 networks: blue: visual (Vis); orange: Somatomotor (SomMot); green: dorsal attention (DosAttn); red: salience/ventral attention (SalVentAttn); purple: Limbic (Limbic); brown: frontoparietal control (Cont); pink: default mode (Default). Arrows indicate how the three vertices of the triangular figure (visual, somatomotor, and default mode networks) change in FEP. (c) T values from the surface‐based linear model (SLM) model plotted on the cortical surface, showing regions in FEP that differ from HC on the first (top) and second (bottom) gradients. Only T values with statistical significance (q < 0.05) are plotted. Blue indicates higher gradient values in FEP while yellow indicates lower values. The denser the color is, the larger the T values. (d) Group comparisons of gradient dispersion between (a) VN and DMN, (b) SMN and DMN, and (c) VN and SMN. Significance levels: ***.001, **.01, *.05. (e–g) T values out from the SLM model showing the effect of being FEP on the seed‐based step‐wise functional connectivity (SFC) degrees at each of the six steps, for (e) primary visual seed (V1), (f) primary somatosensory seed (S1), and (g) primary auditory seed (A1). Only T values with statistical significance (q < 0.05) are plotted. Blue indicates higher gradient values in FEP while yellow indicates lower values. The denser the color is, the larger the T values.

For further confirmation of this pattern, we directly quantified the relative distance between networks using refined gradient dispersion from Bethlehem et al. (2020). Dispersion between VN and DMN, as well as dispersion between SMN and DMN, was defined as the difference between the averaged first gradient values across ROIs within each network. Dispersion between SMN and VN was defined as the difference between the averaged second gradient values across ROIs within each network. As shown in Figure 2d, there was significantly lower VN‐DMN dispersion in FEP (suggesting lesser functional separation between VN and DMN), while dispersion between SMN and DMN was higher in FEP (suggesting decreased affinity).

Seed‐based SFC maps were calculated following Sepulcre et al. (2012), using bilateral primary visual (V1), auditory (A1) and somatosensory (S1) ROIs as seeds. These maps show the evolution of counts of functional links between a ROI and the remainder of cortical regions as the stepwise distance (number of links) between them is increased. Group‐level cortical SFC maps are available in Figures S20–S22. Figure 2e–g show significant increases (blue) and decreases (yellow), in FEP relative to controls, as link‐steps increase step by step (Sepulcre et al., 2012). Specifically, in FEP, there were less links in FEP propagating from V1 to DMN regions and remaining regions of association cortex (Figure 2e), while, to a lesser extent, there were more links between S1 and DMN (Figure 2f), as well as between A1 and DMN (Figure 2g). These results confirm at the level of SFC, an alteration, more pronounced in the case of the visual seed, in the emergence of (or progression to) a trans‐modal architecture relevant for higher‐order functions such as language. Overall, the observation is that of decreased stepwise links between V1 and DMN along with a reduced dispersion between them, and an increase in stepwise connections between the somatosensory cortex and the DMN, in FEP.

2.2 Compressed gradient within the semantic network

The same analysis pipeline as above was therefore applied within the semantic network, defined as comprising the left default and limbic regions following Binder et al. (2009) specifically: (1) the frontal–parietal heteromodal association cortex to process heteromodal and multimodal information, mainly comprising the dorsomedial and dorsolateral prefrontal cortex (dmPFC, dlPFC), middle frontal gyrus, inferior frontal gyrus (IFG), the angular gyrus (AG), the adjacent supramarginal gyrus (SMG), and the lateral temporal lobe; and (2) the self‐related medial limbic cortex, mainly comprising the ventromedial temporal lobe, ventromedial and orbital prefrontal cortex, and the posterior cingulate gyrus (PCC) and adjacent ventral precuneus (Prec) (Binder et al., 2009). Group templates of all 10 semantic gradients can be found in Figures S16–S19 and Tables S3 and S4. As shown in Figure 3a, we observed a principal semantic gradient ranging from lateral heteromodal association cortex (IFG, dlPFC, SMG/AG, and posterior middle temporal gyrus [pMTG]) (as seen in bright colors), to medial limbic regions (PCC, Prec) (as seen in dark colors). In FEP, this hierarchy was generally retained. However, as seen in Figure 3b,c, the high‐end points of this hierarchy in the anterior PCC and medial temporal lobe lowered in FEP, while the low‐end points (AG, SMG, pMTG, and dlPFC) ascended. This, together with a significantly lower within‐network dispersion in FEP (Figure 3e), suggests a compressed semantic network, in which the overall hierarchy is reduced in FEP. Figure 3d further illustrates that the distribution of this gradient also became more centralized in FEP, changing from a left‐skewed trimodal distribution to a bell‐shaped unimodal distribution.

FIGURE 3 Summary of changes in the functional hierarchy within the semantic network. (a) Spatial map of the first gradient of the semantic network on cortical surface. (b) Scatterplots of the first (y‐axis) and the third (x‐axis) gradient in health controls (HC, left) and first‐episode psychosis (FEP, right). The color refers to two cortices in the semantic network as described in the main text: orange: heteromodal association cortex; green: medial limbic cortex. (c) Histogram with global distribution of the first semantic network gradient with density line and arrows showing changes in the peaks and bilateral ends of the density line from HC to FEP. Red: HC; gray: FEP. (d) T values out from the surface‐based linear model (SLM) model showing the effect of being FEP on the first gradient of the semantic network. Only T values with statistical significance (q < 0.05) are plotted. Blue indicates higher gradient values in FEP while yellow indicates lower values. The denser the color is, the larger the T values. (e) Group comparisons of the within‐network dispersion of the first semantic network gradient. (f, g) T values out from the SLM model showing the effect of being FEP on the seed‐based step‐wise functional connectivity (SFC) degrees at each of the five steps, for (f) seeds for the frontal heteromodal association cortex (IFG and pMTG) and (g) seeds for the parietal heteromodal association cortex (AG and SMG). Only T values with statistical significance (q < 0.05) are plotted. Blue indicates higher gradient values in FEP while yellow indicates lower values. The denser the color is, the larger the T values.

Again, we followed up on this gradient analysis with a SFC analysis, choosing as seeds four ROIs at the “low” end of the hierarchy as depicted in Figure 3b: the frontal seeds IFG and dlPFC, on the one hand, and the parietal ones, SMG and AG, on the other. Group‐level SFC maps within the semantic network are available in Figures S23 and S24. Figure 3f,g illustrates increased links between the two ends of the principal gradient in FEP. Specifically, SMG and AG had more links to the whole semantic network in FEP. Moreover, the IFG and dlPFC had more connections to medial limbic regions but fewer links to these two ROIs themselves.

2.3 Behavioral linguistic results

Several findings of changes in semantic cognition in psychosis (Alonso‐Sánchez, Ford, et al., 2022; Corona Hernández et al., 2023; de Boer et al., 2023; Elvevåg et al., 2007; Iter et al., 2018; Mota et al., 2017; Voppel et al., 2023) motivate the prediction of a relation between semantic structure changes in spontaneous speech and the changes in gradients observed above. We constructed semantic graphs to represent the semantic structure of discourse produced during picture descriptions at two levels, based on the associations among units of meaning, which were either content words (woman, farm, horse, etc.) or utterances. Word meanings were represented through embeddings derived from the FastText (FT) model pre‐trained on English data (Grave et al., 2018) (cc.en.300.bin). We computed binarized sparsified semantic similarity graphs using the normalized cosine similarity scores between every pair of two embeddings, from which graph‐theoretical measures were extracted showing centrality, clustering, and the balance between these two in the small‐worldness coefficient, as specified in Table 2. Following a similar procedure, we embedded every utterance using a sentence‐transformers (ST) model (Reimers & Gurevych, 2019) (https://huggingface.co/ST/all-MiniLM-L6-v2), and constructed similar semantic graphs with utterances as the nodes. As shown in Figure 4b, despite a similar quantity of lexical categories in both groups, both semantic graphs (of content words and of utterances), became more “concentrated” in FEP, as indicated by higher closeness centrality and global efficiency. The balance between graph‐wise segregation of semantic units (centrality) and integration of closely related units (clustering), as indicated by the small‐worldness coefficient (sigma, σ), was significantly lower in FEP.

TABLE 2 Features extracted from language graphs.

Domains	Names	Definition	Linguistic indication	
Semantics	Ent_Num	Number of lexical categories	The number of lexical categories	
	Utt_Num	Number of utterances	The number of utterances	
Measures	CC	Closeness centrality averaged across nodes, defined as the reciprocal of average shortest path distance from a node to all reachable nodes of it	Centrality of the semantic graph	
	GE	Global efficiency, averaged multiplicative inverse of shortest path distance between all node pairs	Smoothness and ease of information flow in the semantic graph	
	Clustering	Averaged clustering coefficient, defined as the fraction of possible triangles through a node that exists	Strength of semantic connections, as stronger connections tend to form more clusters	
	Sigma	Small‐worldness coefficient, normalized clustering coefficient divided by normalized path length	Harmonious balance between widely spread semantic connections and tightly knit, conceptually related clusters	
Syntax

	Leaf_num	Number of leaves (nodes without descendants)	The number of words	
Production quantity	Nodes	Total number of nodes	The total number of syntactic productions in the network	
	Phrase fraction	Number of phrases (nodes with descendants) divided by number of leaves	The ratio of generated phrases to the number of words from which the phrases originated	
Hierarchical complexity	Depth	Maximum number of edges that need to be traversed from source (sentence) node to the farthest target (token) node	The maximum hierarchical complexity of embedded syntactic structures	
	DepthMean	The averaged number of edges that need to be traversed to reach every target node from the source node	The mean hierarchical complexity of embedded syntactic structures	
	DepthApEn	Approximate entropy of the number of edges that need to be traversed to reach every target node from the source node	Predictability of the syntactic depth of the next word based on previous words. Higher entropy indicates similar pattern of syntactic depth growth during sentence development more similar.	
NP‐related measures	NP_count	Number of NP nodes	The number of produced noun phrases	
	NP_nest	Number of NPs with another NP(s) embedded in it (nested NPs) divided by the number of NPs	The ratio of nested NPs out of all NPs	
	NP_length	Averaged length of NPs.	The averaged number of words per NP	
VP‐related measures	VP_count	Number of VP nodes	The number of produced noun phrases	
	VP_nest	Number of VPs with another VP(s) embedded in it (nested VPs) divided by the number of VPs	The ratio of nested VPs out of all VPs	
	VP_length	Averaged length of VPs.	The averaged number of words per VP	

FIGURE 4 Linguistic changes and their relations to gradient dispersion. Significance levels: ***.001, **.01, *.05. (a, b) Line plots indicating the semantic and syntactic changes in FEP, respectively. Asterisks on the right denotes the levels of statistical significance. Each error bar featured a dot representing the z value, while the length of the bar represented the 95% confidence interval. The dash lines anchored z values of zero. Error bars with the dot falling left to the line signified an increase in FEP while falling right to the line signified a decrease in FEP. (c) Heatmaps indicating the predictive effect of linguistic measures on gradient dispersion between VN and DMN, SMN and DMN, VN and SMN, and the intrinsic dispersion of semantic network (SemN_G1) via the dispersion of its first gradient. Red cells signified positive predictive effects while blue cells signified negative predictive effects. Only cells where significant predictive effects were observed (corrected for false discovery rate with q < 0.05) were colored. Moreover, we also marked those cells with q in the interval of [0.5, 0.1) with the number sign (#). FT for FastText‐based measures and ST for Sentence transformers. NP: noun phrases; VP: verb phrases; ApEn: approximate entropy; VN: visual network; SMN: somatomotor network; DMN: default mode network; SemN: semantic network; GE: global efficiency; CC: closeness centrality.

At the sentence level, meaning arises through syntactic structure. We therefore tested how lexical concepts were organized into meaningful sentences based on formal syntactic rules rather than their semantic associations with other lexical concepts (as in the above graphs). Syntactic structures geometrically take a hierarchical form typically represented as trees, within which words are organized into phrases (e.g., NPs), which become parts of other phrases (e.g., VPs), until the top node (the S node representing the sentence itself) is reached. The measures extracted from syntactic trees are specified in Table 2.

We first extracted the total number of nodes in the tree, as well as the ratio of generated phrases to the number of words from which the phrases originated, as indicators for production quantity. Both indicators were significantly higher in FEP than HCs. Then, we defined the syntactic depth of each word as the number of edges in the syntactic tree that need to be crossed from this word to the sentence node. From this, we obtain a time series of syntactic depth values from the first to the last word of a sentence. Both the maximum of such arrays and the averaged word‐level depth increased in FEP. The approximate entropy, which indicates the predictability of the next syntactic depth from the previous ones, however, decreased in FEP. The number and averaged length of NPs, as well as the ratio of nested NP, also significantly decreased in FEP. On the other hand, the number of VPs significantly increased, though with a significant decrease in their average length. In short, with more but shorter predicative structures produced (VPs), FEP filled the argument positions of these predicate structures with fewer, shorter, and simpler NPs.

2.4 Linguistic metrics predict cortical hierarchies

Generalized linear models (GLMs) were applied to predict gradient dispersion from semantic graph measures. As shown in Figure 4c, the dispersion between VN and DMN was negatively predicted by both FT‐ and sentence transformer (ST)‐based small‐worldness coefficients (sigma, σ), and positively by ST‐based graph measures of centrality, efficiency, and clustering. In turn, as seen in Figure 4c, the number of words and nodes, the maximum and mean syntactic depth, and VP measures significantly predicted the dispersion of the semantic brain subnetwork, even after correcting for multiple comparisons, in addition to two noteworthy tendencies (q < 0.1): phrase fraction (z = 1.933, q = 0.053) and averaged length of VP (z = 1.809, q = 0.070).

3 DISCUSSION

This study investigated changes in the hierarchical organization of cortical functions in FEP through gradients of FC and how these relate to the semantic organization of spontaneous speech at lexical and structural levels. First, the results confirm a reorganization of cortical hierarchies in FEP. This includes both the first or principal gradient (G1), which represents the functional spectrum ranging from direct perception and action toward the integration and abstraction of information in a multi‐ to transmodal format, and the second gradient (G2), which separates two primary sensorimotor cortices at its two ends, namely the VN and SMN, respectively (Huntenburg et al., 2018). Specifically, the functional distance in G1 between VN and the DMN was compressed in FEP, while that between SMN and DMN was decompressed. In G2, the DMN moved from its position in the middle of the neurotypical G2 hierarchy toward a greater functional distance from the VN and a lesser such distance to the SMN. Together, these large‐scale shifts in network hierarchies depict a disturbance of the balance between sensorimotor and higher‐order association cortices in schizophrenia, which were previously documented by Dong et al. (2021). The shifts are consistent with the SFC analysis, which also demonstrates an uncommon clustering of sensorimotor cortices and the DMN in FEP in terms of the direct and indirect links between them. In particular, along with a decreased dispersion between the VN and the DMN in G1, stepwise connections between V1 and DMN decreased as well. In contrast, stepwise connections increased along with an increased dispersion between SMN and DMN in G1. Overall, the SFC pattern is that there is greater clustering of SMN and DMN and weaker clustering between VN and DMN as well as VN and SMN regions.

Our key question was how such tightening and loosening of the integration between hierarchically ordered cortices is reflected in language, viewed as a higher‐order cognitive function that depends on the integration of multiple cognitive systems. The graphs of both word‐level and utterance‐level meaning (Figure 4b) demonstrate a pattern of significantly higher centrality, weaker small‐worldness, and a tendency for greater clustering. We capture this conceptually as a “shrinking” or “narrowing” semantic space, consistent with recent findings, which found a lesser average semantic distance (higher cosine semantic similarity) between consecutive words in FEP (Alonso‐Sánchez, Ford, et al., 2022; Voppel et al., 2021). Higher centrality in these graphs suggests excessive connections between semantically distant units, while higher clustering indicates the formation of more semantic subdomains. Put differently, the flow of meaning in the speech produced by FEP develops over more distributed topic structures, with more clustered semantic boundaries (“cliquishness”) within a shrinking semantic space.

These semantic graph properties predicted the separation between primary sensory and higher‐order cognition networks in the principal cortical gradient. Since this gradient reflects the dispersion between peripheral‐perceptual and higher‐order cortices, this relation may indeed reflect the specific task demand in a picture description: converting perceptual (visual) information into language, translating percepts into higher‐order meaning as carried by language. Some may argue that these are preverbal and nonlinguistic processes, but it is not realistic to elicit ecologically valid language in isolation without invoking perceptual functions. In fact, the involvement of regions other than core language areas in thought and language disturbances is a well‐recognized phenomenon, even when clinical rating of language dysfunction is the focus in psychosis (Palaniyappan, 2022; Palaniyappan et al., 2023).

Related work already suggests that greater physical distances between DMN regions and sensorimotor cortices may be essential for cognitive processes drawing on semantic memory (Shao et al., 2022; Wang et al., 2022). Going beyond these previous studies, the present one shows a relation between gradients and the semantic organization of spontaneous speech as assessed with neural LMs. In turn, Smallwood et al. (2021) developed a topographical perspective on the functionality of the DMN, noting that, compared to other unimodal areas; the primary auditory cortex is relatively proximal to key DMN and language processing regions such as the angular, inferior frontal, and middle temporal gyri. They speculate that “this proximity to the auditory system allows these regions of the DMN to capitalize on the capacity for language processes to organize cognitive function, perhaps through the vehicle of inner speech.” This intriguing suggestion is pertinent in the context of psychosis, where auditory verbal hallucinations, among other symptoms, reflect a disruption of a (ordinarily silent) process of inner speech. All patients scoring for hallucinatory behavior reported hearing voices when questioned. Associations between hearing voices and visual–verbal hallucinations (seeing a person who spoke) was only observed in a few ones, with no one experiencing visual hallucinations alone. Increased links between the auditory and somatosensory seeds and DMN in our study are, in this respect, an intriguing lead to follow‐up on in future studies, which might target a group of voice hearers specifically.

In the semantic brain network, we identified a new principal gradient separating the multi‐ and heteromodal association cortex in lateral frontotemporal and parietal areas, from medial limbic regions, specifically the PCC and the parahippocampal gyrus. The hierarchy captured by this gradient thus ranges from the frontotemporal subdivision of the semantic network, functioning as information storage, retrieval, and update, to medial limbic regions, which are related to the encoding of episodic information including emotions and social interactions (Binder et al., 2009; Ritchey & Cooper, 2020). This latter information is critically related to the internal self and its narrative (Binder et al., 2009), potentially indicating that the semantic gradient relates to the process of integrating perceptual and conceptual semantic information from the external world to episodic memory and the self. Remarkably, this hierarchy of semantic processing, based on resting‐state data, aligns with a recent semantic tiling of the cortex derived from fMRI using natural speech stimuli, where a semantic gradient spanned from perceptual and quantitative descriptions, as received from the external world, to human social interaction reflecting aspects of the internal self (Huth et al., 2016). The compression of this hierarchy in FEP may indicate a disturbance of these distinct levels of semantic processing and their relationship, possibly resulting in a decreased ability to differentiate the external world and the internal self.

Building higher‐order meaning from perception is inherently subject to structure‐building, over and above merely retrieving lexical‐semantic concepts (Hinzen & Sheehan, 2015). In this respect, our behavioral linguistic analysis showed another surprising and previously undocumented pattern (Figure 4a), which predicted the gradient dispersion of the semantic network: FEP exhibited more nodes in higher syntactic trees, and higher fraction scores of phrasal nodes out of word nodes, with the number of words as a covariate. This finding contrasts most previous studies (Ehlen et al., 2023), although Silva et al. (2022) showed, in an overlapping sample, that while syntactic complexity dropped over time, it was higher at the FEP state. One possible explanation for the contrary findings in most prior studies, apart from illness stage, is that these might have been strongly biased by word count, whereby the schizophrenia group tended to produce shorter sentences probably with less syntactic complexity (Alqahtani et al., 2022; de Boer et al., 2020; Schneider et al., 2023; Silva et al., 2022). This motivated our inclusion of the number of words in the GLM model, though the FEP group generated broadly similar quantity of speech, as measured with words, utterances, and lexical categories, likely thanks to the control of speech duration within 1 min. We suggest that the key to interpreting the result of greater syntactic depth in FEP is our finding of decreased approximate entropy of syntactic depth for each word in an utterance in FEP, in a context where the mean utterance length remained similar. Lower approximate entropy of syntactic depth suggests higher predictability of the syntactic depth of the next word given the previous words. The increased syntactic depth observed in FEP may thus have been achieved via a low‐level process of simple word addition, rather than a genuine increase in hierarchical syntactic complexity. This would align with previous findings of less clausal embedding in psychosis (Çokal et al., 2019; de Boer et al., 2020; DeLisi, 2001; Schneider et al., 2023), a classical indicator of weakening hierarchical syntactic complexity.

Apart from higher syntactic tress and lower approximate entropy, the FEP group employed a greater number of VPs to organize fewer, simpler, and shorter NPs. The intrinsic dispersion of the semantic brain network was predicted by the quantity of syntactic outputs, the hierarchical syntactic complexity, and predicate structures primarily associated with the VP, but not by argument structures as linked to the NP. This asymmetry between VPs and NPs matters, as they play different roles in the genesis of utterance‐level meaning: VPs tend to encode new information, while NPs pick out objects/referents of which this new information is then predicated. New information generated about a referent can provide episodic details. Hence, a shift in the distribution of VPs and NPs may relate to shifts in episodic thinking processes mediated by semantic network regions.

3.1 Limitations

Limitations first include that our dataset was relatively small, and we conducted several steps of analyses. Although we controlled for false discovery, type 2 errors cannot be fully ruled out. Furthermore, speech data were only collected in the context of a picture description task, which could potentially account for the observed correlation between semantic graph features and the dispersion of the first cortical gradient, rather than the dispersion of the semantic network gradient. Although it is a common practice to elicit spontaneous speech in such task, subsequent investigations should aim to validate our findings by employing speech elicited under various conditions on a larger dataset. In addition, it is worth noting that our research focused on the cortical structure and did not incorporate subcortical pathways, which could be vital in understanding psychosis (Yang et al., 2016).

4 CONCLUSIONS

We demonstrated alterations of a unimodal‐to‐transmodal axis of cortical organization in first‐episode psychosis and of a functional gradient in the semantic network. Both of these relate to the disintegration of meaning‐making speech structure in patients. This confirms the role of spontaneous speech as one crucial and largely freely available behavioral readout of neurocognitive functioning as a whole. How we speak necessarily reflects how numerous cognitive processes are integrated at the level of the meaning we produce in response to a prompt or task: it reflects what we perceive and how we make conceptual sense of it, how we organize a discourse, dip into our semantic memory, and generate episodic information from concepts we remember. Psychosis involves large‐scale shifts in the organization of neural cognitive substrates reflected in the process of generating meaning in language.

5 MATERIALS AND METHODS

5.1 Dataset

Analyses were performed on 29 untreated FEP patients recruited for this study, matched to 29 HCs on age, gender, and education. This is a subsample with fMRI data derived from original study reported in Alonso‐Sánchez, Limongi, et al. (2022). The severity of symptoms in the patient group was assessed with the 8‐item Positive and Negative Syndrome Scale (PANSS) version. The eight items are: delusions (PANSS8P1), conceptual disorganization (PANSS8P2), hallucinatory behavior (PANSS8P3), blunted affect (PANSS8N1), passive/apathetic social withdrawal (PANSS8N4), lack of spontaneity and flow of conversation (PANSS8N6), mannerisms and posturing (PANSS8G5), and unusual thought content (PANSS8G9). Subjects underwent a resting‐state fMRI scan and were asked to describe three pictures from the Thematic Apperception Test (1 min for each picture) immediately prior to the scans. Demographic data with clinical scores are shown in Table 1. The whole workflow, including brain analyses and linguistic analyses, is shown in Figure 1. The patient group was recruited in the Prevention and Early Intervention Program for Psychosis in London, Ontario. Three psychiatrists made the clinical assessment for the diagnostic consensus, and the criteria were based on the Statistical Manual of Mental Disorders, 5th Edition (DSM‐5). At the time of testing, 50% of the patients were fully antipsychotic naïve, while the other 50% had an average antipsychotic exposure at the time of scan as 2.8 days (SD 3.7) while average the duration of untreated psychosis was 9.4 (SD 13.5) weeks.

5.2 Resting‐state fMRI acquisition and preprocessing

The fMRI data were acquired at the Centre for Functional and Metabolic Mapping (CFMM) at the University of Western Ontario on a Siemens 7 T Plus (Erlangen, Germany). A total of 360 whole‐brain functional images were collected using a multi‐band EPI acquisition sequence with 20 ms of echo time, 1000 ms of repetition time, a flip angle of 30°, in 63 slices with a multi‐band factor of 3, iPat of 3, and an isotropic resolution of 2 mm. The T1‐weighted MP2RAGE anatomical volume was acquired at a 750 μm isotropic resolution (TE/TR = 2.83/6000 ms). Images were preprocessed using fMRIPrep 21.0.1 (Esteban, Blair, et al., 2018; Esteban, Markiewicz, et al., 2018) based on Nipype 1.6.1 (K. Gorgolewski et al., 2011; K. J. Gorgolewski et al., 2018). Details on how fMRIPrep preprocessed the anatomical and functional data can be found in the first section of the supplementary information (SI‐1). Additional processing steps were applied to the preprocessed images before extracting time series data, including regressing out the effect of head‐motion (rotational and translational on the three coordinate axes) and confound variables from cerebrospinal fluid and white matter, temporal high‐pass filtering of 0.009, and removing the first volume. Details can be found in SI‐1.

5.3 Cortical functional hierarchy with gradient dispersion and SFC

For the whole‐brain gradient analysis, we extracted functional time‐series from 1000 ROIs using Schaefer 1000‐parcels (Schaefer et al., 2018) (Figure S1) organized into seven Yeo networks (Yeo et al., 2011): VN, SMN, dorsal attention (DAN), ventral attention (VAN), limbic (LN), frontoparietal control (FPN), and DMN. We computed a 1000 × 1000 FC matrix using z‐transformed correlation coefficients of the time series. A standardized pipeline using the BrainSpace toolbox (Vos de Wael et al., 2020) was employed to map the FC profiles to low‐dimensional manifolds, and here we give a brief summary of the pipeline. We averaged the subject‐level FC matrices for each group (HC and FEP) to compute a group‐level template. The group templates were sparsified with all negative values, and 90% of the smallest elements per row were zeroed out to keep only positive, strong, and less noisy correlations. Then, for each two ROIs, we computed the normalized cosine similarity scores between their FC profiles for affinity matrices, and applied diffusion mapping to reduce the dimensionality to extract 10 gradients. We also carried out this pipeline on every subject's FC matrix. We aligned the subject‐level gradient maps to the group‐level gradient solution with generalized Procrustes rotation for more robustness. The first two gradients, G1 and G2, were analyzed due to their more established relationship with cognitive functions (Mckeown et al., 2020), with G1 reflecting the extent of separation between unimodal and transmodal cortex (i.e., from VN and SMN to DMN), and G2 differentiating SMN and VN (Shao et al., 2022).

Furthermore, we calculated the SFC based on the pipelines from Sepulcre et al. (2012). The same FC matrices used for the gradient analysis were used to derive the statistical significance of correlation coefficients (SI‐3.1), which were corrected with false discovery rate (FDR) for multiple comparisons (Benjamini & Hochberg, 1995), and binarized based on the corrected FDR q‐values using the threshold of 0.001. The SFC degrees derived from such sparsified FC matrices indicate the count of all paths of a particular exact length, which connect a given seed region to all other areas (Sepulcre et al., 2012). Following Sepulcre et al. (2012), we used bilateral seeds of primary visual (V1), auditory (A1), and somatosensory (S1) areas (SI‐3.2, Table S1 and Figures S2 and S3). While the gradient analyses shed light on connectivity segregation, SFC could serve as a window to the integration, which, from a small‐wordless perspective, is not comprehended by investigating segregation only (Suo et al., 2018). SFC analysis starting in these unimodal areas has been shown to expand toward multimodal areas as the number of steps (i.e., the length of paths) grows, thus offering further insight into the first and second gradients. Similar to gradient analysis, we first computed the group‐level SFC degrees from the group‐level FC templates. SFC maps at every step were normalized using a min‐max scaler into the range between 0 and 1. Second, for each subject, we computed the subject‐level SFC degrees, normalized the map using the same scaler, and multiplied the subject‐level SFC maps with the corresponding group‐level SFC map at each step, which, as suggested by Hong et al. (2019), increases the robustness of SFC maps. Spatial correlation across consecutive steps in group‐level SFC maps, in both HC and FEP, converged at the sixth step (r > .999, Figures S4 and S5). We considered six steps to be sufficient for depicting connectivity profiles in our data.

5.4 Functional hierarchy, gradient dispersion, and SFC in the semantic network

Within the semantic network, we first computed the 10 gradients from FC matrices, analyzed the first gradient, and then defined the dispersion of the first gradient of the semantic network as the sum squared Euclidean distance of every node to its centroid. We manually matched Schaefer's ROIs in the semantic network with these three subdivisions of the semantic network as grouped by Binder et al. (2009), as in Table S4. Based on the gradient results (Table S3), where the hierarchy started from the heteromodal cortex and the multimodal and heteromodal association cortex to medial limbic regions, we chose two ROIs, respectively, for IFG and dlPFC as heteromodal seeds, and two ROIs for SMG and AG as multimodal and heteromodal association seeds. The coordinates are available in Table S2 and Figure S6. SFC maps in the semantic network converged at the fifth step (Figures S7 and S8).

5.5 Graph‐based analysis of semantic graphs in spontaneous speech

Semantic graphs at two levels, one of lexical categories and another of utterances, were constructed through analogous procedures. Tokenization and part‐of‐speech tagging were carried out with spaCy (en_core_web_sm, 3.4.2) (Montani et al., 2022). We computed the normalized cosine similarity scores between every pair of two embeddings, either lexical categories or utterances, for an affinity matrix binarized by proportional thresholding. Consistent with previous studies, we did not pick a single threshold in this case, as we did for gradient and SFC, but selected the lowest threshold from 0.05 to 0.8 with intervals of 0.05, which returns a sparsified matrix with the average degree (the degree of a node is the number of connections linked to the node) over all nodes larger than two multiplies the e‐base logarithm of number of nodes (2logN) (Zhang et al., 2011). This procedure assures that the thresholded network exhibits small‐world properties (Watts & Strogatz, 1998) with as few edges as possible. Notably, we did not require the small‐worldness coefficients to be larger than 1.1 as done in some previous studies, as this is a value derived from specific brain graph data and did not apply to the semantic graph. In the next step, we moved from lexical categories to utterances as the basic units of narratives. The mean thresholds for lexical category graphs were 0.275 for the control group and 0.297 for FES, whereas the thresholds for utterance graphs were 0.448 for controls and 0.453 for FES. Group differences in the thresholds for both lexical category graphs (p = .170) and utterance graphs (p = .558) were not statistically significant, as indicated by GLMs with a Gaussian distribution and an identity link function. Age, education level, and sex were included as covariates in the GLMs. Semantic graphs were constructed for each narrative, instead of each utterance, because there were too few nodes in a considerable number of utterances to construct a meaningful graph (e.g., I see a man and a woman. There will be only three lexical categories see, man, and woman remained after removing stopwords, whereas at least four nodes are required for constructing a small‐world graph (Maslov & Sneppen, 2002).) All semantic measures were extracted from each of the three picture descriptions and averaged across them. Examples of the semantic graphs are provided in Figure S10 (lexical categories) and Figure S11 (utterances).

5.6 Graph‐based analysis of syntactic trees in spontaneous speech

Syntactic trees were derived from a constituency parser (benepar_en3) together with spaCy (Kitaev & Klein, 2018; Montani et al., 2022). Punctuations were kept for parsing the trees but removed from the trees for subsequent analyses. The original trees from the parser were in Greibach normal form, which were converted to Chomsky normal form, allowing a parent node to have at most two children nodes, using the nltk package (3.7) (Bird et al., 2009). Trees were represented as directed acyclic graphs starting from the sentence node on top, through phrasal nodes, to the part‐of‐speech tags of every token forming the leaves of the tree at the bottom. Syntactic measures were extracted at the level of utterance. We first averaged the syntactic measures across utterance for each picture description, then averaged across picture descriptions for each subject. Examples of the syntactic trees are provided in Figure S12.

5.7 Statistical tests

The SLM from the BrainStat toolbox (Larivière et al., 2023) was employed to investigate the fixed effect of group on functional gradients and SFC degrees, with age, sex, and education level as covariates. p‐Values from SLM were corrected for multiple comparisons with FDR. Corrected p values were reported as q values. Mann–Whitney U tests were carried out to compare the four gradient dispersion scores between HC and FEP, also corrected with FDR.

For language measures, we first correlated them with the quantity of speech production (lexical category and utterance counts) using partial Spearman's ranked correlation with diagnosis as a covariate. Semantic graph measures were correlated with the number of lexical categories or utterances. Syntactic measures were correlated with the number of words. All results were domain‐wise corrected for multiple comparisons using FDR. Specific classification of features into the domains can be found in SI‐7. Then, we applied GLMs to explore the group differences of these language measures with age, sex, and education level as covariates. To regress out the effect of simply the quantity of speech production, the number of outputs was included as an additional covariate in those cases where correlations were significant. All semantic measures were compared with the number of lexical categories or utterances as the additional covariate (see Figure S25). All syntactic measures, except for Phrase_fraction, were compared with the number of words as the additional covariate (see Figure S26). We fit most of the data using Gaussian distribution with Identity link function, except for the number of words, where we used Tweedie distribution with log link function, and the numbers of entities and utterances, where Gamma distribution with log link function was used. Deviance goodness‐of‐fit tests indicated all GLMs fit the data well (p > .05).

To explore the relationship between functional hierarchy and different language measures, we employed GLM models with Tweedie distribution and log link function to predict the four gradient dispersion scores as descriptors of the functional hierarchy.

All statistical tests were carried out with statsmodels (0.13.5) and pingouin (0.5.3). Statistical significance was acknowledged when the p‐value (or q‐value when corrected) was less than .05.

AUTHOR CONTRIBUTIONS

Rui He: Conceptualization, methodology, software, validation, formal analysis, writing—original draft, writing—review and editing, visualization. Maria Francisca Alonso‐Sánchez: Conceptualization, resources, data curation, writing—review and editing. Jorge Sepulcre: Methodology, software, writing—review and editing. Lena Palaniyappan: Conceptualization, resources, investigation, project administration (clinical samples), writing—review and editing, supervision, funding acquisition. Wolfram Hinzen: Conceptualization, methodology, writing—original draft, writing—review and editing, supervision, funding acquisition.

CONFLICT OF INTEREST STATEMENT

LP reports personal fees for serving as chief editor from the Canadian Medical Association Journals, speaker/consultant fee from Janssen Canada and Otsuka Canada, SPMM Course Limited, UK, Canadian Psychiatric Association; book royalties from Oxford University Press; investigator‐initiated educational grants from Janssen Canada, Sunovion and Otsuka Canada outside the submitted work. All other authors report no relevant conflicts.

Supporting information

DATA S1. Supporting Information.

ACKNOWLEDGMENTS

The authors appreciate all the participants and their families for the time and effort to contribute to this study. The authors also acknowledge Michael Mackinley, Jenny Chan, and Sabrina Ford of Western University for their support in preparing the dataset. This research was supported by European Research Council (ERC‐2023‐SyG, 101118756 to WH), China Scholarship Council (grant 202108390062 to RH), the Department of Science and Technology of Guangdong Province (grant 112175605105 to WH and RH), and the National Agency for Research and Development (ANID), Scholarship Program, Becas Chile 2019, Postdoctoral Fellow (MA) (to MFAS). The data acquisition for this study was funded by Canadian Institutes of Health Research Foundation Grant (FDN 154296) to LP and was supported by the Canada First Excellence Research Fund to BrainSCAN, Western University (Imaging Core); Innovation fund for Academic Medical Organization of Southwest Ontario; Bucke Family Fund, The Chrysalis Foundation, The Children Hospital Foundation, and The Arcangelo Rea Family Foundation (London, Ontario). Compute Canada Resources (Application No. 1530) were used in the storage and analysis of imaging data. LP acknowledges research support from the Canada First Research Excellence Fund, awarded to the Healthy Brains, Healthy Lives initiative at McGill University (New Investigator Supplement); Monique H. Bourgeois Chair in Developmental Disorders and Graham Boeckh Foundation (Douglas Research Centre, McGill University) and a salary award from the Fonds de recherche du Quebec‐Sante ´ (FRQS).

DATA AVAILABILITY STATEMENT

All codes for analysis were developed by the authors in Python (3.9.12) using free and open‐source packages. All scripts will be available at https://github.com/RuiHe1999/FEP_gradient_sem_syn.
==== Refs
REFERENCES

Alonso‐Sánchez, M. F. , Ford, S. D. , MacKinley, M. , Silva, A. , Limongi, R. , & Palaniyappan, L. (2022). Progressive changes in descriptive discourse in first episode schizophrenia: A longitudinal computational semantics study. Schizophrenia, 8 (1 ), 1. 10.1038/s41537-022-00246-8 35132080
Alonso‐Sánchez, M. F. , Hinzen, W. , He, R. , Gati, J. , & Palaniyappan, L. (2023). Unexpected utterances and the control of meaning in untreated first‐episode psychosis: An ultra‐high field dynamic causal modeling MRI study of the semantic network. ResearchGate. Retrieved from https://www.researchgate.net/publication/373045786
Alonso‐Sánchez, M. F. , Limongi, R. , Gati, J. , & Palaniyappan, L. (2022). Language network self‐inhibition and semantic similarity in first‐episode schizophrenia: A computational‐linguistic and effective connectivity approach. Schizophrenia Research, 259 , 97–103. 10.1016/j.schres.2022.04.007 35568676
Alqahtani, A. , Kayi, E. S. , Hamidian, S. , Compton, M. , & Diab, M. (2022). A quantitative and qualitative analysis of schizophrenia language. Proceedings of the 13th International Workshop on Health Text Mining and Information Analysis (LOUHI), 173–183. 10.18653/v1/2022.louhi-1.20
Baroni, M. (2013). Composition in distributional semantics. Language and Linguistics Compass, 7 (10 ), 511–522. 10.1111/lnc3.12050
Benjamini, Y. , & Hochberg, Y. (1995). Controlling the false discovery rate: A practical and powerful approach to multiple testing. Journal of the Royal Statistical Society: Series B (Methodological), 57 (1 ), 289–300. 10.1111/j.2517-6161.1995.tb02031.x
Bethlehem, R. A. I. , Paquola, C. , Seidlitz, J. , Ronan, L. , Bernhardt, B. , Consortium, C.‐C. , & Tsvetanov, K. A. (2020). Dispersion of functional gradients across the adult lifespan. NeuroImage, 222 , 117299. 10.1016/j.neuroimage.2020.117299 32828920
Binder, J. R. , Desai, R. H. , Graves, W. W. , & Conant, L. L. (2009). Where is the semantic system? A critical review and meta‐analysis of 120 functional neuroimaging studies. Cerebral Cortex, 19 (12 ), 2767–2796. 10.1093/cercor/bhp055 19329570
Bird, S. , Klein, E. , & Loper, E. (2009). Natural language processing with Python (1st ed.). O'Reilly Media.
Butler, P. D. , Silverstein, S. M. , & Dakin, S. C. (2008). Visual perception and its impairment in schizophrenia. Biological Psychiatry, 64 (1 ), 40–47. 10.1016/j.biopsych.2008.03.023 18549875
Ciampelli, S. , de Boer, J. N. , Voppel, A. E. , Corona Hernandez, H. , Brederoo, S. G. , van Dellen, E. , Mota, N. B. , & Sommer, I. E. C. (2023). Syntactic network analysis in schizophrenia‐spectrum disorders. Schizophrenia Bulletin, 49 (suppl 2 ), S172–S182. 10.1093/schbul/sbac194 36946532
Çokal, D. , Zimmerer, V. , Turkington, D. , Ferrier, N. , Varley, R. , Watson, S. , & Hinzen, W. (2019). Disturbing the rhythm of thought: Speech pausing patterns in schizophrenia, with and without formal thought disorder. PLoS One, 14 , e0217404. 10.1371/journal.pone.0217404 31150442
Corcoran, C. M. , Mittal, V. A. , Bearden, C. E. , Gur, R. , Hitczenko, K. , Bilgrami, Z. , Savic, A. , Cecchi, G. A. , & Wolff, P. (2020). Language as a biomarker for psychosis: A natural language processing approach. Schizophrenia Research, 226 , 158–166. 10.1016/j.schres.2020.04.032 32499162
Corona Hernández, H. , Corcoran, C. , Achim, A. M. , de Boer, J. N. , Boerma, T. , Brederoo, S. G. , Cecchi, G. A. , Ciampelli, S. , Elvevåg, B. , Fusaroli, R. , Giordano, S. , Hauglid, M. , van Hessen, A. , Hinzen, W. , Homan, P. , de Kloet, S. F. , Koops, S. , Kuperberg, G. R. , Maheshwari, K. , … Palaniyappan, L. (2023). Natural language processing markers for psychosis and other psychiatric disorders: Emerging themes and research agenda from a cross‐linguistic workshop. Schizophrenia Bulletin, 49 (suppl 2 ), S86–S92. 10.1093/schbul/sbac215 36946526
Dale, C. L. , Brown, E. G. , Fisher, M. , Herman, A. B. , Dowling, A. F. , Hinkley, L. B. , Subramaniam, K. , Nagarajan, S. S. , & Vinogradov, S. (2016). Auditory cortical plasticity drives training‐induced cognitive changes in schizophrenia. Schizophrenia Bulletin, 42 (1 ), 220–228. 10.1093/schbul/sbv087 26152668
de Boer, J. N. , Brederoo, S. G. , Voppel, A. E. , & Sommer, I. E. C. (2020). Anomalies in language as a biomarker for schizophrenia. Current Opinion in Psychiatry, 33 (3 ), 212–218. 10.1097/YCO.0000000000000595 32049766
de Boer, J. N. , Voppel, A. E. , Brederoo, S. G. , Schnack, H. G. , Truong, K. P. , Wijnen, F. N. K. , & Sommer, I. E. C. (2023). Acoustic speech markers for schizophrenia‐spectrum disorders: A diagnostic and symptom‐recognition tool. Psychological Medicine, 53 (4 ), 1302–1312. 10.1017/S0033291721002804 34344490
DeLisi, L. E. (2001). Speech disorder in schizophrenia: Review of the literature and exploration of its relation to the uniquely human capacity for language. Schizophrenia Bulletin, 27 (3 ), 481–496. 10.1093/oxfordjournals.schbul.a006889 11596849
Dong, D. , Yao, D. , Wang, Y. , Hong, S.‐J. , Genon, S. , Xin, F. , Jung, K. , He, H. , Chang, X. , Duan, M. , Bernhardt, B. C. , Margulies, D. S. , Sepulcre, J. , Eickhoff, S. B. , & Luo, C. (2021). Compressed sensorimotor‐to‐transmodal hierarchical organization in schizophrenia. Psychological Medicine, 1–14 , 771–784. 10.1017/S0033291721002129
Ehlen, F. , Montag, C. , Leopold, K. , & Heinz, A. (2023). Linguistic findings in persons with schizophrenia—A review of the current literature. Frontiers in Psychology, 14 , 1287706. 10.3389/fpsyg.2023.1287706 38078276
Elvevåg, B. , Foltz, P. W. , Weinberger, D. R. , & Goldberg, T. E. (2007). Quantifying incoherence in speech: An automated methodology and novel application to schizophrenia. Schizophrenia Research, 93 (1–3 ), 304–316. 10.1016/j.schres.2007.03.001 17433866
Esteban, O. , Blair, R. , Markiewicz, C. J. , Berleant, S. L. , Moodie, C. , Ma, F. , Isik, A. I. , Erramuzpe, A. , Kent, M. , James, D. , Goncalves, M. , DuPre, E. , Sitek, K. R. , Gomez, D. E. P. , Lurie, D. J. , Ye, Z. , Poldrack, R. A. , & Gorgolewski, K. J. (2018). fMRIPrep. Software. 10.5281/zenodo.852659
Esteban, O. , Markiewicz, C. , Blair, R. W. , Moodie, C. , Isik, A. I. , Erramuzpe Aliaga, A. , Kent, J. , Goncalves, M. , DuPre, E. , Snyder, M. , Oya, H. , Ghosh, S. , Wright, J. , Durnez, J. , Poldrack, R. , & Gorgolewski, K. J. (2018). fMRIPrep: A robust preprocessing pipeline for functional MRI. Nature Methods, 16 , 111–116. 10.1038/s41592-018-0235-4 30532080
Gebreegziabhere, Y. , Habatmu, K. , Mihretu, A. , Cella, M. , & Alem, A. (2022). Cognitive impairment in people with schizophrenia: An umbrella review. European Archives of Psychiatry and Clinical Neuroscience, 272 (7 ), 1139–1155. 10.1007/s00406-022-01416-6 35633394
Gold, J. M. , Hahn, B. , Strauss, G. P. , & Waltz, J. A. (2009). Turning it upside down: Areas of preserved cognitive function in schizophrenia. Neuropsychology Review, 19 (3 ), 294–311. 10.1007/s11065-009-9098-x 19452280
Gorgolewski, K. , Burns, C. D. , Madison, C. , Clark, D. , Halchenko, Y. O. , Waskom, M. L. , & Ghosh, S. (2011). Nipype: A flexible, lightweight and extensible neuroimaging data processing framework in Python. Frontiers in Neuroinformatics, 5 , 13. 10.3389/fninf.2011.00013 21897815
Gorgolewski, K. J. , Esteban, O. , Markiewicz, C. J. , Ziegler, E. , Ellis, D. G. , Notter, M. P. , Jarecka, D. , Johnson, H. , Burns, C. , Manhães‐Savio, A. , Hamalainen, C. , Yvernault, B. , Salo, T. , Jordan, K. , Goncalves, M. , Waskom, M. , Clark, D. , Wong, J. , Loney, F. , … Ghosh, S. (2018). Nipype. Software. 10.5281/zenodo.596855
Grave, E. , Bojanowski, P. , Gupta, P. , Joulin, A. , & Mikolov, T. (2018). Learning word vectors for 157 languages. Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018). Language Resources and Evaluation Conference. Retrieved from https://infoscience.epfl.ch/record/253313
He, R. , Palominos, C. , Zhang, H. , Alonso‐Sánchez, M. F. , Palaniyappan, L. , & Hinzen, W. (2024). Navigating the semantic space: Unraveling the structure of meaning in psychosis using different computational language models. Psychiatry Research, 333 , 115752. 10.1016/j.psychres.2024.115752 38280291
Hinzen, W. , & Sheehan, M. (2015). The philosophy of universal grammar. Oxford University Press.
Hong, S.‐J. , Vos de Wael, R. , Bethlehem, R. A. I. , Lariviere, S. , Paquola, C. , Valk, S. L. , Milham, M. P. , Di Martino, A. , Margulies, D. S. , Smallwood, J. , & Bernhardt, B. C. (2019). Atypical functional connectome hierarchy in autism. Nature Communications, 10 (1 ), 1. 10.1038/s41467-019-08944-1
Huntenburg, J. M. , Bazin, P.‐L. , & Margulies, D. S. (2018). Large‐scale gradients in human cortical organization. Trends in Cognitive Sciences, 22 (1 ), 21–31. 10.1016/j.tics.2017.11.002 29203085
Huth, A. G. , de Heer, W. A. , Griffiths, T. L. , Theunissen, F. E. , & Gallant, J. L. (2016). Natural speech reveals the semantic maps that tile human cerebral cortex. Nature, 532 (7600 ), 453–458. 10.1038/nature17637 27121839
Iter, D. , Yoon, J. , & Jurafsky, D. (2018). Automatic detection of incoherent speech for diagnosing schizophrenia. Proceedings of the Fifth Workshop on Computational Linguistics and Clinical Psychology: From Keyboard to Clinic, 136–146. 10.18653/v1/W18-0615
Javitt, D. C. (2009). When doors of perception close: Bottom‐up models of disrupted cognition in schizophrenia. Annual Review of Clinical Psychology, 5 , 249–275. 10.1146/annurev.clinpsy.032408.153502
Javitt, D. C. , & Sweet, R. A. (2015). Auditory dysfunction in schizophrenia: Integrating clinical and basic features. Nature Reviews. Neuroscience, 16 (9 ), 535–550. 10.1038/nrn4002 26289573
Kitaev, N. , & Klein, D. (2018). Constituency parsing with a self‐attentive encoder. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2676–2686. 10.18653/v1/P18-1249
Larivière, S. , Bayrak, Ş. , Vos de Wael, R. , Benkarim, O. , Herholz, P. , Rodriguez‐Cruces, R. , Paquola, C. , Hong, S.‐J. , Misic, B. , Evans, A. C. , Valk, S. L. , & Bernhardt, B. C. (2023). BrainStat: A toolbox for brain‐wide statistics and multimodal feature associations. NeuroImage, 266 , 119807. 10.1016/j.neuroimage.2022.119807 36513290
Maslov, S. , & Sneppen, K. (2002). Specificity and stability in topology of protein networks. Science, 296 (5569 ), 910–913. 10.1126/science.1065103 11988575
Matsumoto, Y. , Nishida, S. , Hayashi, R. , Son, S. , Murakami, A. , Yoshikawa, N. , Ito, H. , Oishi, N. , Masuda, N. , Murai, T. , Friston, K. , Nishimoto, S. , & Takahashi, H. (2023). Disorganization of semantic brain networks in schizophrenia revealed by fMRI. Schizophrenia Bulletin, 49 (2 ), 498–506. 10.1093/schbul/sbac157 36542452
McCutcheon, R. A. , Keefe, R. S. E. , & McGuire, P. K. (2023). Cognitive impairment in schizophrenia: Aetiology, pathophysiology, and treatment. Molecular Psychiatry, 1–17 , 1902–1918. 10.1038/s41380-023-01949-9
Mckeown, B. , Strawson, W. H. , Wang, H.‐T. , Karapanagiotidis, T. , Vos de Wael, R. , Benkarim, O. , Turnbull, A. , Margulies, D. , Jefferies, E. , McCall, C. , Bernhardt, B. , & Smallwood, J. (2020). The relationship between individual variation in macroscale functional gradients and distinct aspects of ongoing thought. NeuroImage, 220 , 117072. 10.1016/j.neuroimage.2020.117072 32585346
Mesulam, M. M. (1998). From sensation to cognition. Brain, 121 (Pt 6 ), 1013–1052. 10.1093/brain/121.6.1013 9648540
Montani, I. , Honnibal, M. , Honnibal, M. , Landeghem, S. V. , Boyd, A. , Peters, H. , McCann, P. O. , Geovedi, J. , O'Regan, J. , Samsonov, M. , Altinok, D. , Orosz, G. , de Kok, D. , Kristiansen, S. L. , Bournhonesque, R. , Miranda, L. , Kannan, M. , Baumgartner, P. , Edward , … Murat . (2022). explosion/spaCy: V3.4.2: Latin and Luganda support, Python 3.11 wheels and more [Computer software]. Zenodo. 10.5281/zenodo.7228125
Mota, N. B. , Copelli, M. , & Ribeiro, S. (2017). Thought disorder measured as random speech structure classifies negative symptoms and schizophrenia diagnosis 6 months in advance. NPJ Schizophrenia, 3 , 18. 10.1038/s41537-017-0019-3 28560264
Palaniyappan, L. (2022). Dissecting the neurobiology of linguistic disorganisation and impoverishment in schizophrenia. Seminars in Cell & Developmental Biology, 129 , 47–60. 10.1016/j.semcdb.2021.08.015 34507903
Palaniyappan, L. , Homan, P. , & Alonso‐Sanchez, M. F. (2023). Language network dysfunction and formal thought disorder in schizophrenia. Schizophrenia Bulletin, 49 (2 ), 486–497. 10.1093/schbul/sbac159 36305160
Pintos, A. S. , Hui, C. L.‐M. , De Deyne, S. , Cheung, C. , Ko, W. T. , Nam, S. Y. , Chan, S. K.‐W. , Chang, W.‐C. , Lee, E. H.‐M. , Lo, A. W.‐F. , Lo, T.‐L. , Elvevåg, B. , & Chen, E. Y.‐H. (2022). A longitudinal study of semantic networks in schizophrenia and other psychotic disorders using the word association task. Schizophrenia Bulletin Open, 3 (1 ), sgac054. 10.1093/schizbullopen/sgac054 39144793
Reimers, N. , & Gurevych, I. (2019). Sentence‐BERT: Sentence Embeddings using Siamese BERT‐networks. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP‐IJCNLP), 3982–3992. 10.18653/v1/D19-1410
Ritchey, M. , & Cooper, R. A. (2020). Deconstructing the posterior medial episodic network. Trends in Cognitive Sciences, 24 (6 ), 451–465. 10.1016/j.tics.2020.03.006 32340798
Rubinov, M. , & Sporns, O. (2010). Complex network measures of brain connectivity: Uses and interpretations. NeuroImage, 52 (3 ), 1059–1069. 10.1016/j.neuroimage.2009.10.003 19819337
Schaefer, A. , Kong, R. , Gordon, E. M. , Laumann, T. O. , Zuo, X.‐N. , Holmes, A. J. , Eickhoff, S. B. , & Yeo, B. T. T. (2018). Local‐global parcellation of the human cerebral cortex from intrinsic functional connectivity MRI. Cerebral Cortex, 28 (9 ), 3095–3114. 10.1093/cercor/bhx179 28981612
Schneider, K. , Leinweber, K. , Jamalabadi, H. , Teutenberg, L. , Brosch, K. , Pfarr, J.‐K. , Thomas‐Odenthal, F. , Usemann, P. , Wroblewski, A. , Straube, B. , Alexander, N. , Nenadić, I. , Jansen, A. , Krug, A. , Dannlowski, U. , Kircher, T. , Nagels, A. , & Stein, F. (2023). Syntactic complexity and diversity of spontaneous speech production in schizophrenia spectrum and major depressive disorders. Schizophrenia, 9 (1 ), 35. 10.1038/s41537-023-00359-8 37248240
Sepulcre, J. , Sabuncu, M. R. , Yeo, T. B. , Liu, H. , & Johnson, K. A. (2012). Stepwise connectivity of the modal cortex reveals the multimodal organization of the human brain. The Journal of Neuroscience, 32 (31 ), 10649–10661. 10.1523/JNEUROSCI.0759-12.2012 22855814
Shao, X. , Mckeown, B. , Karapanagiotidis, T. , Vos de Wael, R. , Margulies, D. S. , Bernhardt, B. , Smallwood, J. , Krieger‐Redwood, K. , & Jefferies, E. (2022). Individual differences in gradients of intrinsic connectivity within the semantic network relate to distinct aspects of semantic cognition. Cortex, 150 , 48–60. 10.1016/j.cortex.2022.01.019 35339787
Silva, A. M. , Limongi, R. , MacKinley, M. , Ford, S. D. , Alonso‐Sánchez, M. F. , & Palaniyappan, L. (2022). Syntactic complexity of spoken language in the diagnosis of schizophrenia: A probabilistic Bayes network model. Schizophrenia Research, 259 (22 ), 88–96. 10.1016/j.schres.2022.06.011 35752547
Smallwood, J. , Bernhardt, B. C. , Leech, R. , Bzdok, D. , Jefferies, E. , & Margulies, D. S. (2021). The default mode network in cognition: A topographical perspective. Nature Reviews Neuroscience, 22 (8 ), 8. 10.1038/s41583-021-00474-4
Suo, X. , Lei, D. , Li, L. , Li, W. , Dai, J. , Wang, S. , He, M. , Zhu, H. , Kemp, G. J. , & Gong, Q. (2018). Psychoradiological patterns of small‐world properties and a systematic review of connectome studies of patients with 6 major psychiatric disorders. Journal of Psychiatry & Neuroscience, 43 (6 ), 416–427. 10.1503/jpn.170214
Sydnor, V. J. , Larsen, B. , Bassett, D. S. , Alexander‐Bloch, A. , Fair, D. A. , Liston, C. , Mackey, A. P. , Milham, M. P. , Pines, A. , Roalf, D. R. , Seidlitz, J. , Xu, T. , Raznahan, A. , & Satterthwaite, T. D. (2021). Neurodevelopment of the association cortices: Patterns, mechanisms, and implications for psychopathology. Neuron, 109 (18 ), 2820–2846. 10.1016/j.neuron.2021.06.016 34270921
Voppel, A. E. , de Boer, J. N. , Brederoo, S. G. , Schnack, H. G. , & Sommer, I. (2021). Quantified language connectedness in schizophrenia‐spectrum disorders. Psychiatry Research, 304 , 114130. 10.1016/j.psychres.2021.114130 34332431
Voppel, A. E. , de Boer, J. N. , Brederoo, S. G. , Schnack, H. G. , & Sommer, I. E. C. (2023). Semantic and acoustic markers in schizophrenia‐spectrum disorders: A combinatory machine learning approach. Schizophrenia Bulletin, 49 (suppl 2 ), S163–S171. 10.1093/schbul/sbac142 36305054
Vos de Wael, R. , Benkarim, O. , Paquola, C. , Lariviere, S. , Royer, J. , Tavakol, S. , Xu, T. , Hong, S.‐J. , Langs, G. , Valk, S. , Misic, B. , Milham, M. , Margulies, D. , Smallwood, J. , & Bernhardt, B. C. (2020). BrainSpace: A toolbox for the analysis of macroscale gradients in neuroimaging and connectomics datasets. Communications Biology, 3 (1 ), 1. 10.1038/s42003-020-0794-7 31925316
Wang, X. , Krieger‐Redwood, K. , Zhang, M. , Cui, Z. , Wang, X. , Karapanagiotidis, T. , Du, Y. , Leech, R. , Bernhardt, B. C. , Margulies, D. S. , Smallwood, J. , & Jefferies, E. (2022). Physical distance to sensory‐motor landmarks predicts language function. Cerebral Cortex, 33 , bhac344. 10.1093/cercor/bhac344
Watts, D. J. , & Strogatz, S. H. (1998). Collective dynamics of ‘small‐world’ networks. Nature, 393 (6684 ), 440–442. 10.1038/30918 9623998
Yang, G. J. , Murray, J. D. , Wang, X.‐J. , Glahn, D. C. , Pearlson, G. D. , Repovs, G. , Krystal, J. H. , & Anticevic, A. (2016). Functional hierarchy underlies preferential connectivity disturbances in schizophrenia. Proceedings of the National Academy of Sciences, 113 (2 ), E219–E228. 10.1073/pnas.1508436113
Yeo, B. T. T. , Krienen, F. M. , Sepulcre, J. , Sabuncu, M. R. , Lashkari, D. , Hollinshead, M. , Roffman, J. L. , Smoller, J. W. , Zöllei, L. , Polimeni, J. R. , Fischl, B. , Liu, H. , & Buckner, R. L. (2011). The organization of the human cerebral cortex estimated by intrinsic functional connectivity. Journal of Neurophysiology, 106 (3 ), 1125–1165. 10.1152/jn.00338.2011 21653723
Yeshurun, Y. , Nguyen, M. , & Hasson, U. (2021). The default mode network: Where the idiosyncratic self meets the shared social world. Nature Reviews Neuroscience, 22 (3 ), 3. 10.1038/s41583-020-00420-w 33219370
Zhang, J. , Wang, J. , Wu, Q. , Kuang, W. , Huang, X. , He, Y. , & Gong, Q. (2011). Disrupted brain connectivity networks in drug‐naive, first‐episode major depressive disorder. Biological Psychiatry, 70 (4 ), 334–342. 10.1016/j.biopsych.2011.05.018 21791259
