Skip to Content
MDPIMDPI

Advanced

HealthcareHealthcare
PDF
  • Systematic Review
  • Open Access

10 April 2026

20 Pages

Linguistic Markers in At-Risk Mental States Using Natural Language Processing: A Systematic Review

,
,
,
,
,
and
1
Department of Clinical and Health Psychology, Autonomous University of Barcelona, 08193 Bellaterra, Spain
2
Instituto de Psiquiatría y Salud Mental, Instituto de Investigación (IdISSC), Hospital Clínico San Carlos, 28040 Madrid, Spain
3
Department of Applied Mathematics for Information and Communication Technologies, Higher Technical School of Computer Systems Engineering, Polytechnic University of Madrid, 28040 Madrid, Spain
4
MIT Linq, Massachusetts Institute of Technology, Cambridge, MA 02139-4307, USA
Healthcare2026, 14(8), 999;https://doi.org/10.3390/healthcare14080999
This article belongs to the Special Issue Health Informatics and Clinical Decision Support in Psychiatric Care for Vulnerable Populations
Version Notes
Review Reports
Order Reprints

Highlights

What are the main findings?
  • NLP markers, particularly reduced semantic coherence, lower syntactic complexity, and diminished referential cohesion, significantly differentiate individuals with at-risk mental states (ARMS) from healthy controls.
  • Automated linguistic indicators can predict the transition to psychosis with high accuracy levels (ranging from 79% to 100%).
What are the implications of the main findings?
  • NLP serves as an objective, non-invasive, and cost-effective tool that complements traditional semi-structured interviews in clinical settings.
  • The integration of automated speech analysis into early detection protocols could enhance prognostic specificity, enabling more timely and targeted preventive interventions for individuals at clinical high risk.

Abstract

Background/Objectives: In recent years, research on psychosis has increasingly focused on prevention, aiming to implement early interventions that mitigate or reduce its impact. Within this framework, the analysis of linguistic markers in individuals with at-risk mental states (ARMS) has proven valuable for identifying those at risk and predicting psychosis onset. Artificial intelligence tools, particularly natural language processing (NLP), have emerged as effective resources for detecting these language-based indicators. This study aims to synthesize the existing scientific evidence on linguistic markers analyzed through NLP techniques in individuals with ARMS. Methods: A systematic review following the PRISMA 2020 protocol was conducted. Three databases (PubMed, PsycInfo, and Scopus) were searched for published articles from their inception to October 2025. Rayyan software was used to manage references and article downloads. Out of ninety initial search results, fifteen studies involving 1313 participants from diverse groups were included in the review. Results: The findings indicated that alterations in semantic coherence, syntactic complexity, referential cohesion, and speech/content poverty differentiated ARMS individuals from healthy controls. Several of these markers, analyzed with NLP methods, predicted the onset of psychosis with accuracy levels ranging from 79% to 100%, although these findings should be interpreted with caution due to the significant methodological heterogeneity and variability in sample sizes across the included studies. Conclusions: NLP techniques offer a powerful approach for detecting language alterations that distinguish ARMS individuals and provide meaningful predictions of psychosis onset, highlighting their potential as a complement to traditional clinical assessments for early identification and prevention.

1. Introduction

In recent decades, the scientific community has witnessed a significant shift toward the prevention of psychotic disorders, given their status as a leading cause of long-term disability in young adults [1]. As functional alterations across multiple domains can be identified during the prodromal phase, clinicians have adopted a preventive approach aimed at reducing the probability of transition to psychosis diagnosis or, at minimum, attenuating the severity of the disease [2,3]. Early detection is clinically relevant not only for identifying individuals at high risk but also for the timely implementation of interventions that maximize the degree of recovery and reduce the duration of untreated psychosis (DUP) [4,5,6]. Absence or delay in treatment often leads to first-episode psychosis (FEP), which is associated with marked declines in education, employment, and social functioning [7], underscoring the need to refine detection strategies toward the earliest possible stages of risk [8].
To identify individuals prior to the emergence of overt symptomatology, frameworks such as the Ultra High Risk (UHR) [9] and Clinical High Risk (CHR) [10] have been developed. The concept of an at-risk mental state (ARMS) is a broad clinical framework used to identify individuals in this vulnerable period. However, ARMS is not a single diagnostic entity; rather, it represents a heterogeneous group characterized by various subthreshold mental health conditions and pre-existing vulnerabilities. The development of ARMS is often preceded by a combination of genetic load (family history of psychotic disorders), premorbid personality traits (such as high levels of schizotypy), and environmental stressors, including childhood trauma or significant social isolation. Furthermore, individuals with ARMS often present with comorbid subthreshold mood disorders, such as depression and anxiety, which interact with emergent cognitive impairments to facilitate the onset of risk symptoms [11]. Semi-structured interviews like the Comprehensive Assessment of At-Risk Mental States (CAARMS) [12] and the Structured Interview for Prodromal Symptoms (SIPS/SOPS) [10] remain the gold standard for ARMS diagnosis, but their predictive specificity is notably low. Previous longitudinal studies have shown that only 29% to 36% of subjects identified as being at clinical high-risk develop psychosis within a three-year follow-up period [13]. This diagnostic gap suggests that relying solely on traditional clinical interviews may be insufficient for accurate prognosis. Consequently, there is an urgent need for objective complementary markers that can improve the sensitivity and specificity of current assessment techniques.
Language abnormalities are increasingly recognized as important predictive markers for the onset of psychosis, with recent studies investigating these alterations as indicators of underlying formal thought disorders [14]. Evidence has shown that specific linguistic features, including pronoun usage, syntactic complexity, speech cohesion, and lexical density, can differ significantly between individuals with ARMS and healthy controls [3,15].
Various artificial intelligence methods have been explored to identify ARMS. Machine learning has been applied to analyze neuroimaging data (MRI/fMRI) to detect structural brain changes, as well as to electronic health records (EHR) and digital phenotyping from smartphone sensors to track behavioral patterns. While these approaches offer objective metrics, they often require expensive infrastructure, invasive procedures, or face significant data privacy hurdles, limiting their applied use in research and, most notably, in clinical environments. The most widely used artificial intelligence method for identifying linguistic markers in individuals with ARMS is natural language processing (NLP) [16], owing to its ability to analyze large volumes of text both automatically and accurately. NLP leverages machine learning algorithms to detect abnormal patterns in speech and writing by examining syntactic, semantic, and pragmatic features. For example, previous studies have shown that NLP methods were able to detect that compared to healthy controls, people with schizophrenia used more pronouns but fewer adverbs, adjectives, and determiners in speech. Additionally, NLP techniques have demonstrated the capacity to identify diagnosis-specific patterns that reflect the distinct etiological mechanisms of different disorders. For instance, while both schizophrenia and first-episode bipolar disorder may present with disorganized speech, NLP-based research indicates that schizophrenia is often characterized by a more pronounced breakdown in semantic coherence and structural simplicity (e.g., incomplete words), reflecting core neurocognitive deficits. In contrast, linguistic alterations in bipolar disorder typically manifest as increased verbal productivity and a ‘flight of ideas’ driven by affective dysregulation [17]. These findings suggest that NLP can capture the unique ‘linguistic signature’ of each disorder, providing insights into their differing underlying psychotic processes [17]. NLP methods were also able to detect that those with FEP demonstrated higher semantic similarity in speech [18]. The NLP approach has also proven effective in predicting and detecting the transition to psychosis [19]. Furthermore, automated NLP techniques are capable of identifying complex and subtle language patterns that may go unnoticed by clinicians assessing ARMS using traditional methods. The markers detected through this approach can uncover anomalies that, according to several studies, help predict the transition to psychosis with substantially higher accuracy than that achieved using traditional clinical interview methods [15,19,20].
The aim of this systematic review is to synthesize the existing scientific evidence regarding linguistic markers in individuals with ARMS using NLP techniques, addressing two distinct clinical questions: (1) the diagnostic identification of the ARMS state (discriminating at-risk individuals from healthy controls) and (2) the prognostic prediction of psychosis transition (discriminating between those who later convert to psychosis and those who do not). Additionally, this review examines how linguistic markers identified in individuals with ARMS compared with those observed in first-episode psychosis (FEP) to evaluate the existence of a linguistic continuum from subthreshold risk to overt illness.
To ensure conceptual clarity, this review uses ‘at-risk mental states (ARMS)’ as the standard clinical term for the population under study, and ‘transition’ to refer to the longitudinal onset of psychosis. Regarding the linguistic markers analyzed via NLP, detailed definitions and clinical examples for each of these linguistic constructs are provided in Table A1 (Appendix A). In addition, a broad and inclusive definition of FEP is applied, encompassing individuals within approximately the first 2–5 years of psychotic symptom onset, regardless of prior treatment exposure, while acknowledging that operational definitions vary across studies. This approach is consistent with the current recommendations for early psychosis research and facilitates meaningful integration of studies with heterogeneous inclusion criteria.

2. Materials and Methods

2.1. Search Strategy

This review was conducted in line with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses statement (available upon request) [21]. The PRISMA checklist is presented in the Appendix A.
A systematic search of published studies was conducted using the following databases: PubMed, Scopus and PsycInfo, from database inception to October 2025. This broad timeframe was chosen to ensure a comprehensive historical mapping of the field, allowing for the identification of core linguistic markers that remain consistent regardless of the evolving computational power. The three databases were selected to prioritize studies with a high level of clinical evidence. While engineering-focused repositories as IEEE Xplore exist, the current review aimed to synthesize findings derived from validated clinical cohorts and standardized psychiatric assessments, for which medical and psychological databases provide the most relevant and rigorously peer-reviewed literature.
The search focused on linguistic markers in ARMS populations analyzed using NLP techniques. Only articles published in English were considered, and no restrictions were applied regarding the year of publication. Additional articles were identified by manually searching the references of retrieved articles and relevant reviews. The general search strategy for each database was designed by combining keywords, the search term repository (thesaurus, MeSH), and Boolean operators (AND, OR). The keywords used in the search strategy for each database are shown in Table 1.
Table 1. Specific search strategies.
The results were imported into the Rayyan platform [22] for duplicate detection and removal, as well as for title, abstract and full-text screening of relevant studies (see Figure 1).
Figure 1. Flowchart of the study selection process for the systematic review.

2.2. Eligibility Criteria

Articles published in English that examined linguistic markers in ARMS populations using NLP and artificial intelligence techniques were included. No restrictions on the sociodemographic characteristics of the analyzed samples were applied. Any research design was considered (observational or experimental, cross-sectional or longitudinal), but single-case studies, literature reviews, systematic reviews, and meta-analyses were excluded. No restrictions were applied based on the year of publication.

2.3. Data Extraction

The first co-authors (Y.Z. and A.C.) were responsible for data extraction from all included articles, and any discrepancies were resolved through discussion with two of the co-authors (J.-B.N. and A.B.). The extracted data encompassed the following: (i) first author and year of publication, (ii) target population and sample size, (iii) ARMS diagnostic tool, (iv) language collection method, (v) language analysis technique, (vi) analyzed language variables, and (vii) main findings. In this review, language collection methods refer to the approaches used to obtain speech samples for subsequent analysis, whereas language analysis techniques denote the tools and/or applications employed to process, examine, and extract features from these samples, including any preprocessing procedures.

2.4. Quality Assessment and Risk of Bias

The methodological quality and risk of bias of the included studies were independently assessed by two reviewers (J.-B.N. and Y.Z.). Any discrepancies were resolved through consensus or by consulting a third senior reviewer (A.B.). The JBI Critical Appraisal Tools for Analytical Cross-Sectional Studies for quantitative papers was used. Studies were evaluated across several domains, including inclusion criteria clarity, setting description, validity and reliability of the linguistic measurements (NLP techniques), identification of confounding factors (such as age, education, or medication), and appropriateness of statistical analysis. No studies were excluded based on quality scores.

3. Results

To ensure a structured and comprehensive synthesis of the evidence, the section is organized into three thematic levels. First, we present the study selection process and the results of the methodological quality and risk of bias assessment. Second, we provide an overview of the study characteristics, detailing the clinical populations and diagnostic tools (SIPS/CAARMS) used across the 15 included studies. Finally, we move from descriptive reporting to a thematic synthesis of linguistic markers, grouping the findings into four core domains: semantic coherence, syntactic complexity, referential cohesion, and linguistic poverty.

3.1. Study Selection

A total of 90 articles were identified through searches of the three databases and manual screening of bibliographic references. After removing duplicates, the first author (Y.Z.) independently reviewed the titles and abstracts of the 47 remaining publications, selecting 11 for full-text review (see Figure 1). Subsequently, using the Rayyan platform [22], two independent reviewers (J.-B.N. and A.B.) agreed to include the 11 articles during the full-text screening phase. A search of the bibliographic references of the 11 articles was then done, which revealed 5 potentially relevant articles, from which only one was excluded because of the language. Ultimately, 15 articles were included in this systematic review.

3.2. Methodological Quality and Risk of Bias Assessment of Included Studies

The formal quality assessment indicated that the majority of the studies (n = 13; 86.6%) exhibited high methodological quality, with a low risk of bias. Most studies provided detailed descriptions of the clinical participants (SIPS/SOPS or CAARMS criteria) and utilized robust NLP pipelines. The most common limitation identified was the small sample size in some studies and the lack of explicit strategies to deal with all potential confounders (e.g., the effect of antipsychotic medication on speech production). Despite these limitations, the statistical rigor—specifically the use of cross-validation and permutation testing—was deemed appropriate for the respective sample sizes. A detailed breakdown of the quality assessment for each study is provided in Table A2 (see Appendix A).

3.3. Study Characteristics

The list of included studies and their main characteristics is presented in Table 2. The sample sizes ranged from 24 to 93 ARMS participants, with only one study exceeding this range, including 167 subjects [23]. Seven of the fifteen studies also included groups with first-episode psychosis (FEP) in addition to control groups, while only three studies focused exclusively on ARMS participants. Regarding ARMS diagnosis, eleven studies utilized the Structured Interview for Psychosis (SIPS)—usually in conjunction with the Scale of Prodromal Symptoms (SOPS)—, whereas five studies employed the Comprehensive Assessment of At-Risk Mental States (CAARMS).
Table 2. Studies analyzing linguistic markers in ARMS.
Various methods were employed to collect speech samples, including open-ended narrative interviews and the Thematic Apperception Test (TAT) being the most used. A range of computational methods was applied for language analysis. The Natural Language Toolkit (NLTK) was the most frequently utilized, while other methods, such as Latent Semantic Analysis (LSA), Specific Speech Part Tagging (POS-Tag), S-BERT Bidirectional Sentence Encoder, and Speech Graph Software 2.0, were also employed in several studies, albeit less frequently. The language variables most frequently analyzed were grouped into semantic coherence, syntactic complexity, referential cohesion, and formal thought disorders, with poverty of speech and poverty of content being the most extensively studied (see Table A1, Appendix A for definitions of these variables).
An initial consideration regarding the 15 reviewed articles is the distinction between those that exclusively employ computational language analysis techniques (13 out of 15) and hybrid studies that incorporate manual expert assessment into the computational outputs [26,27].
The main findings indicated the following linguistic markers as the most prominent in individuals with ARMS:
Semantic coherence. Three articles reported reduced semantic coherence in individuals with ARMS compared to healthy controls [19,26,28]. Moreover, Bilgrami et al. [26] found a correlation between semantic coherence and formal positive thinking disorders (tangentiality, circumstantiality, and derailment), as assessed using the Thought, Language, and Communication Assessment Scale (TLC). In another study, semantic coherence with syntactic complexity and speech/content poverty in ARMS participants was associated with measures of brain structure and functional connectivity [25]. In contrast, Dalal et al. [29] suggested most ARMS individuals demonstrated semantic coherence aligned with a typical linguistic profile along with the controls. Additionally, Kizilay et al. [30] reported contradictory findings, showing higher semantic coherence in ARMS individuals than in the control group. Nevertheless, the authors of this study noted that they developed a machine learning model based on NLP-derived features that achieved 79.6% accuracy in distinguishing ARMS individuals from controls.
Syntactic complexity. Three studies reported reduced syntactic complexity in individuals with ARMS compared to control subjects [19,26,30]. Another study found that syntactic complexity was correlated with negative symptoms and appeared sensitive to prodromal symptoms in ARMS participants [25]. Similarly, Bilgrami et al. [26] reported that syntactic complexity was associated with formal negative thought disorder (speech/content poverty), as assessed using the Thought, Language and Communication Assessment Scale (TLC).
Referential cohesion. One study reported reduced referential cohesion in individuals with ARMS compared to control subjects [24]. Additionally, three studies found that referential cohesion was lower in individuals who later developed psychosis compared to the ARMS group [14,16].
Poverty of speech/content. One study reported higher levels of speech/content poverty in individuals with ARMS compared to controls [26]. Furthermore, other study found that speech poverty was higher in ARMS individuals who subsequently transitioned to psychosis than in those who did not [20].

3.4. Linguistic Markers as Predictors of Psychosis

Three of the included studies specifically examined the predictive capacity of certain language variables for psychosis. Bedi et al. [15] demonstrated that semantic coherence, syntactic complexity, and speech/content poverty jointly contributed to accurately predicting the onset of psychosis. Corcoran et al. [19] reported that reduced semantic coherence and syntactic complexity, assessed through the use of possessive pronouns, predicted psychosis with 79% accuracy. Finally, Rezaii et al. [20] found that speech/content poverty, along with the use of words relative to voices and sounds mentioned during the SIPS/SOPS interview, collectively predicted psychosis with 90% accuracy. All three studies employed the Natural Language Toolkit (NLTK), one of the most used NLP tools for language analysis in peer-reviewed studies.
Additionally, retrospective analyses of speech samples in other studies indicated that referential cohesion [14,16] and speech/content poverty [20] were also significant predictors of psychosis.

3.5. Additional Language Measures Beyond NLP in ARMS Research

Two of the included studies examined additional language measures complementary to those obtained through NLP. Baklund et al. [27] compared the validity of NLP measures with basic self-disturbance (BSD) scores, noting that BSD measures are qualitatively different and can serve as additional clinical risk markers in pre-psychotic stages. Similarly, Srivastava [23] analyzed the semantic similarity between the natural speech of ARMS individuals and items from the Inventory of Psychotic-Like Anomalous Self-Experiences (IPASE), which assess anomalous self-experiences. The results showed higher semantic similarity in ARMS individuals compared to controls. These complementary measures provide further insights into linguistic alterations, enriching the predictive and clinical information obtained from NLP analyses.

4. Discussion

This systematic review has synthesized the current scientific evidence on the analysis of linguistic markers in the ARMS population. A key insight from this research is the increasing integration of advanced technologies, particularly NLP methods and machine learning approaches, into clinical psychopathology assessment over the last decade. While various computational methods were identified, including the use of NLP libraries like the Natural Language Toolkit (NLTK) for feature extraction, the application of these features within machine learning models suggests promising, though still emerging, avenues for identifying language patterns characteristic of the ARMS population. These patterns hold potential for early identification and prediction of transition to psychosis. A recent systematic review by García-Molina et al. also evaluated the application of NLP to speech analysis applied to ARMS groups, focusing primarily on the diagnostic and prognostic performance of these models (e.g., accuracy, sensitivity, specificity, and AUC-ROC) [33]. Although both reviews examine NLP studies in at-risk populations, Garcia-Molina et al. primarily evaluated diagnostic and transition accuracy, organizing their synthesis around predictive performance. In contrast, our review places equal emphasis on the linguistic features extracted by these models, systematically examining patterns in semantic, syntactic, and discourse-level markers to clarify the language characteristics associated with psychosis risk. Thus, while the prior review focuses on how well NLP models perform, the present work seeks to elucidate what linguistic signals underlie these predictions.
The primary linguistic markers consistently highlighted across the reviewed studies, derived largely from NLP-based analyses, include alterations in semantic coherence, syntactic complexity, referential cohesion, and the manifestation of formal thought disorders such as poverty of speech/content. Beyond these primary NLP-derived markers, this review also identified studies exploring complementary language measures in ARMS individuals. For instance, Baklund et al. [27] investigated BSD linguistic markers, suggesting they may offer qualitatively different insights, while Srivastava et al. [23] examined the semantic similarity between natural speech and IPASE items. These studies suggest that a multi-modal approach to language analysis, potentially combining computational NLP features with qualitatively distinct measures, might offer a more comprehensive understanding of language alterations in the ARMS population. The discussion of how these complementary measures relate to, or could be integrated with, NLP-derived features warrants further exploration in future research.
The reviewed literature predominantly focused on linguistic markers capable of distinguishing the ARMS population from healthy controls and, in some instances, from individuals experiencing an FEP. Several studies employed retrospective analyses of speech samples from individuals who later transitioned to psychosis (ARMS+) versus those who did not (ARMS−), thereby identifying potential differentiating linguistic features. Others specifically investigated the predictive capacity of certain linguistic markers, reporting varying degrees of accuracy. Addressing the question of which linguistic markers best distinguish the ARMS population from healthy individuals, this review suggests that reduced semantic coherence and diminished syntactic complexity are among the most frequently reported trends, although the findings across studies remain heterogeneous. Secondary to these, poverty of speech/content and reduced referential cohesion also emerged as relevant differentiators.
Semantic coherence, in particular, was the most frequently investigated language characteristic within the included studies of this review. The predominant finding of reduced semantic coherence in ARMS individuals suggests its potential as a robust linguistic marker. This aligns with theoretical perspectives where the analysis of semantic coherence is considered a strategy to identify subclinical manifestations of formal thought disorders, such as derailment, in the ARMS population [34] (See Table A1, Appendix A). Further supporting this, Bilgrami et al. [26] reported an association between semantic coherence and positive thought disorders, as assessed by the TLC. Notably, they also observed that coherence was negatively correlated with negative symptoms, likely reflecting the reduced linguistic complexity and shorter utterances characteristic of individuals with prominent negative symptomatology, which may constrain opportunities for derailment. Given the relative consistency of these findings, semantic coherence appears to be a critical feature to consider in the development of any automated language analysis tool for ARMS assessment. However, the operationalization of “semantic coherence” can vary significantly depending on the NLP techniques employed (e.g., latent semantic analysis vs. graph-based methods vs. transformer-based embeddings), and the impact of this methodological heterogeneity on the robustness of this finding requires careful consideration.
Reduced syntactic complexity also emerged as a robust linguistic marker distinguishing individuals with ARMS from controls. From a psychopathological perspective, studies by Haas et al. [25] and Bilgrami et al. [26] demonstrated its association with negative symptoms. Hence, reduced syntactic complexity may represent a promising linguistic feature for the early identification of subthreshold manifestations of psychosis, particularly within the negative symptom domain, although further validation is required.
Regarding linguistic markers predictive of psychosis, a degree of consensus was observed across studies evaluating semantic coherence, syntactic complexity, and speech/content poverty. These studies reported predictive accuracies ranging from 79% to 100%, though substantial methodological diversity and varying cohort sizes across studies necessitate a cautious interpretation of these outcomes [15,19,20]. It is crucial to interpret these high accuracy figures with caution, considering the often small sample sizes and methodological heterogeneity in the primary studies, which can affect model generalizability. Referential cohesion also showed potential as a predictor, with reduced cohesion observed in ARMS individuals who later transitioned to psychosis (ARMS+) [14,16]. The use of NLP toolkits like NLTK was common across several predictive studies [15,19,20], though the specific features extracted and the machine learning algorithms subsequently applied likely varied, underscoring the need for better methodological transparency and standardization in future predictive modelling research.
Additionally, although direct comparisons between ARMS and FEP were limited across the included studies, the available evidence suggests partial overlap in key linguistic alterations, particularly reduced semantic coherence and syntactic complexity. These shared features may support the hypothesis of a linguistic continuum from subthreshold risk states to overt psychotic illness. However, the extent to which these alterations reflect a progressive worsening versus qualitatively distinct phenomena remains unclear. Further longitudinal and cross-diagnostic studies are needed to better characterize the trajectory and specificity of these linguistic changes across the psychosis spectrum.
A notable limitation of most of the included studies was their small sample sizes. This is particularly pertinent for studies employing machine learning techniques, where small datasets can lead to model overfitting and limit the generalizability of findings, necessitating replication in larger independent cohorts. Furthermore, the inclusion of participants on antipsychotic medication in some studies introduces a potential confounder, as medication effects could influence language production and thus the observed results.
Methodological heterogeneity across analytical approaches poses a significant challenge. To ascertain the predictive utility of language variables for psychosis transition, studies have employed diverse statistical analyses and machine learning models. This methodological variability, while reflective of an exploratory phase in the field, complicates the direct comparison of findings and the formulation of generalized conclusions regarding the most effective computational strategies. The path from identifying “linguistic markers” to establishing them as validated clinical “biomarkers” is long and requires rigorous testing for reliability, specificity, and clinical utility, which is currently underexplored. In addition, it is important to note that the study of linguistic markers in the ARMS population is still very recent, and consequently, the available scientific evidence remains limited. Furthermore, only articles published in English were included, meaning that potentially relevant evidence in other languages may have been overlooked.
These findings may have relevant clinical implications for the detection and prevention of psychosis, as well as for the treatment of the ARMS population. Linguistic markers can reveal the presence of subclinical symptoms in early stages before the full onset of clinical symptoms, allowing intervention to be applied to individuals who are at potential risk of transitioning to psychosis in order to prevent or mitigate its impact. The application of AI and machine learning allows for the analysis of large linguistic datasets and the detection of subtle alterations that might be missed by human evaluators, thereby enhancing the accuracy and efficiency of identifying these markers. Such tools could complement existing clinical assessments, potentially improving the diagnostic precision. However, the practical implementation of this methodology into routine clinical settings will require addressing challenges such as data privacy, integration with electronic health records, and clinician training.
Some critical considerations should be considered when interpreting the conclusions of this work. First, it is necessary to consider the evolution of NLP technology. The included studies span from early models based on Latent Semantic Analysis (LSA) and part-of-speech tagging [15] to modern transformer-based models (e.g., BERT, SBERT) [23,26,30]. While earlier statistical methods established foundational markers like semantic coherence, they often lacked the context-awareness of current architectures. Interestingly, predictive accuracies have remained high across eras; however, modern models offer superior robustness and handle the inherent complexity of natural language more effectively, potentially reducing the risk of overfitting seen in smaller earlier cohorts. A second critical consideration refers to whether the reported deficits in linguistic markers are specific to ARMS or reflect a general signature of cognitive decline. Similar NLP-derived alterations, such as reduced syntactic complexity and semantic density, have been documented in neurodegenerative conditions like Alzheimer’s and Parkinson’s diseases [35,36]. However, important qualitative differences exist. In Alzheimer’s, linguistic impairment is typically driven by semantic memory loss and lexical retrieval deficits (often resulting in ‘empty speech’), whereas in ARMS, the markers are primarily associated with the breakdown of logical thought trajectories and formal thought disorder. While Parkinson’s disease also affects verbal fluency, it is often characterized by motor–speech impairments (dysarthria) and executive dysfunction rather than the semantic incoherence found in the psychosis spectrum. Future research using NLP should focus on multi-diagnostic comparisons to further refine the specificity of these ‘digital biotypes’. A third and final consideration is about the number of published studies applying natural language processing to clinical high-risk (CHR) populations, which remains limited and concentrated within a small number of research groups. This relative scarcity also reflects the early developmental stage of the field. Importantly, large-scale initiatives such as the Accelerating Medicines Program in Schizophrenia have recently made available a substantial multilingual corpus of de-identified clinical transcripts through the NIH National Data Archive, including longitudinal outcome data and extracted linguistic features. These resources are expected to generate a rapid increase in NLP-based studies of CHR populations in the coming years.
In this context, the current review is intended not as a definitive synthesis of a mature literature but as a conceptual and methodological framework to support the next phase of research. By critically evaluating existing linguistic metrics, theoretical assumptions, and sources of heterogeneity across studies, this review aims to provide a foundation for interpreting future multi-center large-scale findings, facilitating cross-study comparability, and guiding the development of clinically meaningful linguistic biomarkers. As increasingly complex NLP approaches—including large language models—are applied to expansive datasets, such grounding may be particularly important to ensure that advances in predictive performance remain aligned with psychopathological theory and clinical relevance.
For future research, several avenues are apparent. There is a pressing need for studies with larger and more diverse samples to validate the current findings and to train more robust machine learning models. Standardization of both language data collection methods and NLP/machine learning analytical pipelines is crucial for improving the comparability and reliability of results. This could involve developing benchmark datasets and encouraging comparative studies of different computational approaches on the same data. Furthermore, longitudinal studies are essential to more definitively establish the predictive validity of linguistic markers for psychosis transition. Investigating the interpretability of complex machine learning models used in this domain is also important to ensure clinical acceptance and understanding. While the results of this systematic review warrant cautious interpretation, they provide a foundational basis for directing further investigation into linguistic markers that show significant promise for identifying ARMS individuals and predicting the onset of psychosis. Lastly, emerging studies have begun to examine the emotional valence of language in ARMS populations. For instance, recent findings indicate that narratives produced by ARMS group contain a higher proportion of incongruent negative emotional words compared to healthy controls [31], suggesting affective disturbances in spontaneous speech. Future research could expand on this line of inquiry by systematically quantifying positive and negative emotional word use and by exploring how such affective–linguistic patterns interact with established markers such as semantic coherence and syntactic complexity.

5. Conclusions

This systematic review highlights the potential of linguistic marker analysis in the ARMS population to identify subclinical symptoms in prodromal phases of psychosis using different NLP methods. It has been observed that the use of this type of techniques has made it possible to detect alterations in language, such as reduced semantic coherence, reduced syntactic complexity and poor speech/content, which may help distinguish individuals with ARMS from healthy controls and contribute to the prediction of psychosis onset. These findings point to a promising and potentially fruitful direction for improving clinical evaluation and early detection. Likewise, the study of linguistic markers using NLP techniques also contributes to the development of solid theoretical models on the relationship between language alterations and subthreshold symptoms in the ARMS population. NLP methods provide automated high-throughput analysis but should be regarded as complementary to gold-standard clinician assessments. Traditional expert evaluations, though time-intensive, offer the high clinical interpretability that current black-box computational models often lack.
While the findings are promising, the limitations identified in the included studies (e.g., small sample sizes, methodological heterogeneity) warrant a cautious interpretation of the results. Nonetheless, the evidence reviewed provides a valuable foundation to inform and guide more rigorous future research in this rapidly evolving field.

Author Contributions

A.C. (Alba Carrió) and Y.Z. contributed equally to the manuscript as first authors and participated in investigation, methodology, visualization and writing the original draft of the manuscript. J.-B.N. and A.B. contributed equally to the manuscript as senior (last) authors and were responsible for conceptualization, funding acquisition, project administration, supervision, reviewing and editing of the manuscript, and providing final approval of the version. J.S.-L.-J., E.G., and A.C. (Ana Calvo) were involved in the critical review and editing of the manuscript, and E.G. also contributed to funding acquisition. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the Ministry of Research and Universities of the Government of Catalonia to the Research Group COGen-MH “Cognition and Gender: Implications for Mental Health” (Code: 2021 SGR 000534). This project has received funding from the European Union’s Horizon Europe research and innovation programme under the Marie Skłodowska-Curie grant agreement No 101126533. Co-funded by the European Union.

Institutional Review Board Statement

Not applicable.

Data Availability Statement

No new data were created or analyzed in this study.

Conflicts of Interest

The authors declare no conflicts of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results.

Abbreviations

The following abbreviations are used in this manuscript:
ARMSIndividuals with At-Risk Mental States
NLPNatural Language Processing
PRISMAPreferred Reporting Items for Systematic Reviews and Meta-Analyses
FEPFirst-Episode Psychosis
CHRClinical High Risk for Psychosis
DUPDuration of Untreated Psychosis
UHRUltra High Risk

Appendix A

Table A1. Definition and examples of the language variables analyzed.
Table A2. Quality assessment of included studies using JBI Critical Appraisal Tools.

References

  1. Schmidt, S.J.; Schultze-Lutter, F.; Schimmelmann, B.G.; Maric, N.P.; Salokangas, R.K.R.; Riecher-Rössler, A.; van der Gaag, M.; Meneghelli, A.; Nordentoft, M.; Marshall, M.; et al. EPA Guidance on the Early Intervention in Clinical High Risk States of Psychoses. Eur. Psychiatry 2015, 30, 388–404. [Google Scholar] [CrossRef] [Scilit]
  2. Fusar-Poli, P.; Rocchetti, M.; Sardella, A.; Avila, A.; Brandizzi, M.; Caverzasi, E.; Politi, P.; Ruhrmann, S.; McGuire, P. Disorder, Not Just State of Risk: Meta-Analysis of Functioning and Quality of Life in People at High Risk of Psychosis. Br. J. Psychiatry 2015, 207, 198–206. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  3. Olivares, R.I.; Figueroa, B.A. Análisis de Alteraciones Del Discurso En Estados Mentales de Alto Riesgo de Psicosis (EMAR): Una Revisión Sistemática. Rev. Chil. Neuropsiquiatr. 2021, 59, 343–360. [Google Scholar] [CrossRef] [Scilit]
  4. Bird, V.; Premkumar, P.; Kendall, T.; Whittington, C.; Mitchell, J.; Kuipers, E. Early Intervention Services, Cognitive–Behavioural Therapy and Family Intervention in Early Psychosis: Systematic Review. Br. J. Psychiatry 2010, 197, 350–356. [Google Scholar] [CrossRef] [Scilit]
  5. Larsen, T.K.; Melle, I.; Auestad, B.; Haahr, U.; Joa, I.; Johannessen, J.O.; Opjordsmoen, S.; Rund, B.R.; Rossberg, J.I.; Simonsen, E.; et al. Early Detection of Psychosis: Positive Effects on 5-Year Outcome. Psychol. Med. 2011, 41, 1461–1469. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  6. Killackey, E.; Yung, A.R. Effectiveness of Early Intervention in Psychosis. Curr. Opin. Psychiatry 2007, 20, 121–125. [Google Scholar] [CrossRef] [Scilit]
  7. Rinaldi, M.; Killackey, E.; Smith, J.; Shepherd, G.; Singh, S.P.; Craig, T. First Episode Psychosis and Employment: A Review. Int. Rev. Psychiatry 2010, 22, 148–162. [Google Scholar] [CrossRef] [Scilit]
  8. Salazar de Pablo, G.; Guinart, D.; Armendariz, A.; Aymerich, C.; Catalan, A.; Alameda, L.; Rogdaki, M.; Martinez Baringo, E.; Soler-Vidal, J.; Oliver, D.; et al. Duration of Untreated Psychosis and Outcomes in First-Episode Psychosis: Systematic Review and Meta-Analysis of Early Detection and Intervention Strategies. Schizophr. Bull. 2024, 50, 771–783. [Google Scholar] [CrossRef] [Scilit]
  9. Yung, A.; Phillips, L.; McGorry, P.D. Treating Schizophrenia in the Prodromal Phase; CRC Press: Boca Raton, FL, USA, 2004; ISBN 9781135428983. [Google Scholar]
  10. Miller, T.J.; McGlashan, T.H.; Rosen, J.L.; Cadenhead, K.; Ventura, J.; McFarlane, W.; Perkins, D.O.; Pearlson, G.D.; Woods, S.W. Prodromal Assessment With the Structured Interview for Prodromal Syndromes and the Scale of Prodromal Symptoms: Predictive Validity, Interrater Reliability, and Training to Reliability. Schizophr. Bull. 2003, 29, 703–715. [Google Scholar] [CrossRef] [Scilit]
  11. Thompson, A.; Marwaha, S.; Broome, M.R. At-Risk Mental State for Psychosis: Identification and Current Treatment Approaches. BJPsych Adv. 2016, 22, 186–193. [Google Scholar] [CrossRef] [Scilit]
  12. Yung, A.R.; Yung, A.R.; Pan Yuen, H.; Mcgorry, P.D.; Phillips, L.J.; Kelly, D.; Dell’olio, M.; Francey, S.M.; Cosgrave, E.M.; Killackey, E.; et al. Mapping the Onset of Psychosis: The Comprehensive Assessment of At-Risk Mental States. Aust. New Zealand J. Psychiatry 2005, 39, 964–971. [Google Scholar] [CrossRef] [Scilit]
  13. Fusar-Poli, P.; Bonoldi, I.; Yung, A.R.; Borgwardt, S.; Kempton, M.J.; Valmaggia, L.; Barale, F.; Caverzasi, E.; McGuire, P. Predicting Psychosis: Meta-Analysis of Transition Outcomes in Individuals at High Clinical Risk. Arch. Gen. Psychiatry 2012, 69, 220. [Google Scholar] [CrossRef] [Scilit]
  14. Spencer, T.J.; Thompson, B.; Oliver, D.; Diederen, K.; Demjaha, A.; Weinstein, S.; Morgan, S.E.; Day, F.; Valmaggia, L.; Rutigliano, G.; et al. Lower Speech Connectedness Linked to Incidence of Psychosis in People at Clinical High Risk. Schizophr. Res. 2021, 228, 493–501. [Google Scholar] [CrossRef] [Scilit]
  15. Bedi, G.; Carrillo, F.; Cecchi, G.A.; Slezak, D.F.; Sigman, M.; Mota, N.B.; Ribeiro, S.; Javitt, D.C.; Copelli, M.; Corcoran, C.M. Automated Analysis of Free Speech Predicts Psychosis Onset in High-Risk Youths. Schizophrenia 2015, 1, 15030. [Google Scholar] [CrossRef] [Scilit]
  16. Morgan, S.E.; Diederen, K.; Vértes, P.E.; Ip, S.H.Y.; Wang, B.; Thompson, B.; Demjaha, A.; De Micheli, A.; Oliver, D.; Liakata, M.; et al. Natural Language Processing Markers in First Episode Psychosis and People at Clinical High-Risk. Transl. Psychiatry 2021, 11, 630. [Google Scholar] [CrossRef] [Scilit]
  17. Tang, S.X.; Kriz, R.; Cho, S.; Park, S.J.; Harowitz, J.; Gur, R.E.; Bhati, M.T.; Wolf, D.H.; Sedoc, J.; Liberman, M.Y. Natural Language Processing Methods Are Sensitive to Sub-Clinical Linguistic Differences in Schizophrenia Spectrum Disorders. NPJ Schizophr. 2021, 7, 25. [Google Scholar] [CrossRef] [Scilit]
  18. Arslan, B.; Kizilay, E.; Verim, B.; Demirlek, C.; Demir, M.; Cesim, E.; Eyuboglu, M.S.; Ozbek, S.U.; Sut, E.; Yalincetin, B.; et al. Computational Analysis of Linguistic Features in Speech Samples of First-Episode Bipolar Disorder and Psychosis. J. Affect. Disord. 2024, 363, 340–347. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  19. Corcoran, C.M.; Carrillo, F.; Fernández-Slezak, D.; Bedi, G.; Klim, C.; Javitt, D.C.; Bearden, C.E.; Cecchi, G.A. Prediction of Psychosis across Protocols and Risk Cohorts Using Automated Language Analysis. World Psychiatry 2018, 17, 67–75. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  20. Rezaii, N.; Walker, E.; Wolff, P. A Machine Learning Approach to Predicting Psychosis Using Semantic Density and Latent Content Analysis. NPJ Schizophr. 2019, 5, 1–12. [Google Scholar] [CrossRef] [Scilit]
  21. Page, M.J.; McKenzie, J.E.; Bossuyt, P.M.; Boutron, I.; Hoffmann, T.C.; Mulrow, C.D.; Shamseer, L.; Tetzlaff, J.M.; Akl, E.A.; Brennan, S.E.; et al. The PRISMA 2020 Statement: An Updated Guideline for Reporting Systematic Reviews. BMJ 2021, 372, n71. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  22. Ouzzani, M.; Hammady, H.; Fedorowicz, Z.; Elmagarmid, A. Rayyan-a Web and Mobile App for Systematic Reviews. Syst. Rev. 2016, 5, 210. [Google Scholar] [CrossRef] [Scilit]
  23. Srivastava, A.; Selloni, A.; Bilgrami, Z.R.; Sarac, C.; McGowan, A.; Cotter, M.; Bayer, J.; Spark, J.; Krcmar, M.; Formica, M.; et al. Differential Expression of Anomalous Self-Experiences in Spontaneous Speech in Clinical High-Risk and Early-Course Psychosis Quantified by Natural Language Processing. Biol. Psychiatry Cogn. Neurosci. Neuroimaging 2023, 8, 1005–1012. [Google Scholar] [CrossRef] [Scilit]
  24. Gupta, T.; Hespos, S.J.; Horton, W.S.; Mittal, V.A. Automated Analysis of Written Narratives Reveals Abnormalities in Referential Cohesion in Youth at Ultra High Risk for Psychosis. Schizophr. Res. 2018, 192, 82–88. [Google Scholar] [CrossRef] [Scilit]
  25. Haas, S.S.; Doucet, G.E.; Garg, S.; Herrera, S.N.; Sarac, C.; Bilgrami, Z.R.; Shaik, R.B.; Corcoran, C.M. Linking Language Features to Clinical Symptoms and Multimodal Imaging in Individuals at Clinical High Risk for Psychosis. Eur. Psychiatry 2020, 63, e72. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  26. Bilgrami, Z.R.; Sarac, C.; Srivastava, A.; Herrera, S.N.; Azis, M.; Haas, S.S.; Shaik, R.B.; Parvaz, M.A.; Mittal, V.A.; Cecchi, G.; et al. Construct Validity for Computational Linguistic Metrics in Individuals at Clinical Risk for Psychosis: Associations with Clinical Ratings. Schizophr. Res. 2022, 245, 90–96. [Google Scholar] [CrossRef] [Scilit]
  27. Baklund, L.; Røssberg, J.I.; Møller, P. Linguistic Markers and Basic Self-Disturbances among Adolescents at Risk of Psychosis. A Qualitative Study. EClinicalMedicine 2023, 55, 101733. [Google Scholar] [CrossRef] [Scilit]
  28. Nettekoven, C.R.; Diederen, K.; Giles, O.; Duncan, H.; Stenson, I.; Olah, J.; Gibbs-Dean, T.; Collier, N.; Vértes, P.E.; Spencer, T.J.; et al. Semantic Speech Networks Linked to Formal Thought Disorder in Early Psychosis. Schizophr. Bull. 2023, 49, S142–S152. [Google Scholar] [CrossRef] [Scilit]
  29. Dalal, T.C.; Liang, L.; Silva, A.M.; Mackinley, M.; Voppel, A.; Palaniyappan, L. Speech Based Natural Language Profile before, during and after the Onset of Psychosis: A Cluster Analysis. Acta Psychiatr. Scand. 2025, 151, 332–347. [Google Scholar] [CrossRef] [Scilit]
  30. Kizilay, E.; Arslan, B.; Verim, B.; Demirlek, C.; Demir, M.; Cesim, E.; Eyuboglu, M.S.; Uzman Ozbek, S.; Sut, E.; Yalincetin, B.; et al. Automated Linguistic Analysis in Youth at Clinical High Risk for Psychosis. Schizophr. Res. 2024, 274, 121–128. [Google Scholar] [CrossRef] [Scilit]
  31. Mota, N.B.; Ribeiro, M.; Malcorra, B.; Argolo, F.; Lopes-Rocha, A.C.; Ara, A.; Gondim, J.M.; Cecchi, G.; Loch, A.A.; Corcoran, C.M. Attenuated Symptoms Are Associated with Connectedness and Emotional Expression in Narratives Based on Emotional Pictures in a Brazilian Clinical High-Risk Cohort. Psychiatry Res. 2025, 348, 116469. [Google Scholar] [CrossRef] [Scilit]
  32. Kim-Dufor, D.-H.; Walter, M.; Krebs, M.-O.; Haralambous, Y.; Lenca, P.; Lemey, C. Deeper Insight into Speech Characteristics of Patients at Ultra-High Risk Using Classification and Explainability Models. Front. Psychiatry 2025, 16, 1595197. [Google Scholar] [CrossRef] [Scilit]
  33. García-Molina, J.T.; Downey, M.; Méndez, E.; Figueroa-Barra, A. Diagnostic and Transition Accuracy of Natural Language Processing in High Risk for Psychosis Individuals: A Systematic Review. Asian J. Psychiatry 2025, 112, 104695. [Google Scholar] [CrossRef] [Scilit]
  34. Argolo, F.; Magnavita, G.; Mota, N.B.; Ziebold, C.; Mabunda, D.; Pan, P.M.; Zugman, A.; Gadelha, A.; Corcoran, C.; Bressan, R.A. Lowering Costs for Large-Scale Screening in Psychosis: A Systematic Review and Meta-Analysis of Performance and Value of Information for Speech-Based Psychiatric Evaluation. Braz. J. Psychiatry 2020, 42, 673–686. [Google Scholar] [CrossRef] [Scilit]
  35. Eyigoz, E.; Mathur, S.; Santamaria, M.; Cecchi, G.; Naylor, M. Linguistic Markers Predict Onset of Alzheimer’s Disease. EClinicalMedicine 2020, 28, 100583. [Google Scholar] [CrossRef] [Scilit]
  36. García, A.M.; Carrillo, F.; Orozco-Arroyave, J.R.; Trujillo, N.; Vargas Bonilla, J.F.; Fittipaldi, S.; Adolfi, F.; Nöth, E.; Sigman, M.; Fernández Slezak, D.; et al. How Language Flows When Movements Don’t: An Automated Analysis of Spontaneous Discourse in Parkinson’s Disease. Brain Lang. 2016, 162, 19–28. [Google Scholar] [CrossRef] [Scilit]
  37. Bearden, C.E.; Wu, K.N.; Caplan, R.; Cannon, T.D. Thought Disorder and Communication Deviance as Predictors of Outcome in Youth at Clinical High Risk for Psychosis. J. Am. Acad. Child Adolesc. Psychiatry 2011, 50, 669–680. [Google Scholar] [CrossRef] [Scilit]
  38. Figueroa-Barra, A.; Del Aguila, D.; Cerda, M.; Gaspar, P.A.; Terissi, L.D.; Durán, M.; Valderrama, C. Automatic Language Analysis Identifies and Predicts Schizophrenia in First-Episode of Psychosis. Schizophrenia 2022, 8, 53. [Google Scholar] [CrossRef] [Scilit]
  39. Ortega, L. Syntactic Complexity Measures and Their Relationship to L2 Proficiency: A Research Synthesis of College-level L2 Writing. Appl. Linguist. 2003, 24, 492–518. [Google Scholar] [CrossRef] [Scilit]
  40. Vilarrasa, A.B. Vallejo. Introducción a la Psicopatología y la Psiquiatría; Elsevier: Amsterdam, The Netherlands, 2025; ISBN 9788491138303. [Google Scholar]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Article Metrics

Citations

Article Access Statistics

Journal Statistics
Multiple requests from the same IP address are counted as one view.
Effects of Aquatic Exercise on Type 2 Diabetes Management in Adulthood: A Systematic Review and Meta-Analysis, Including Evidence on the Use of Wearable Devices
Previous
Frailty Transition and Risk of New-Onset Arthritis Among Adults Aged 45 Years and Older: A Longitudinal Analysis of CHARLS
Next