Schizophrenia is an illness of disordered thought, with symptoms often reflected in disturbances in language and communication (1). An impairment of verbal communication is one of several diagnostic features of schizophrenia, with a strong posited genetic component (2), but not all patients with schizophrenia exhibit clinically identifiable disordered speech. Speech disturbances, referred to as formal thought disorders (FTD), can be classified into positive or negative FTD. Positive FTD includes phenomena such as derailment, tangentiality, or in more severe cases, neologisms or even complete incoherence (schizophasia). Alternatively, negative FTD captures the characteristic poverty of speech that many patients experience (1). While several scales have been developed with the goal of identifying these elements of speech, such as the scale for Thought, Language, and Communication (TLC) (3) or the Thought Language Index (TLI) (4), many of the speech disturbances in psychosis are too subtle to be captured by clinicians during a cross-sectional clinical interaction (5).

Recent work has focused on identifying subtler forms of speech variation in naturalistic speech among schizophrenia samples, a goal that has been aided by the proliferation of automated linguistic analysis tools (6, 7). The utilization of these automated speech analysis software programs allows complex analysis of speech without the burdens (and expense) of manual scoring. Automated linguistic analyses have allowed researchers to identify disturbances in multiple levels of speech in schizophrenia, from phonological, morphological, syntactic, and pragmatic levels (8), and have been utilized in predicting psychosis onset in at risk populations (9).

While it is intuitive that social and vocational outcomes may relate to one’s verbal abilities, the body of research demonstrating this link in schizophrenia have several limitations that preclude the use of linguistic features in functional prognostication. First, much of this work has been based on language impairments in experimental, rather than naturalistic, paradigms where the semantic space is defined by the researcher (10) [e.g., using verbal fluency tests (11, 12)]. Secondly, even in studies assessing unconstrained speech, objective aspects of conversations are not considered; instead, the clinically judged construct of thought disorder is employed. While studies have observed associations between functional outcomes and negative FTD (specifically poverty of speech and content) (13, 14), other studies have reported that positive, but not negative, elements of FTD are related to functional outcome (15). These inconsistencies in the extant literature may be related to the difficulties surrounding the clinical assessment of formal thought disorder. Thirdly, most studies to date make cross-sectional correlations between functioning and verbal assessments; there is a notable lack of longitudinal data to clarify whether the verbal deficits temporally precede (and thus lie on the causal pathway of) poor functioning seen in schizophrenia. Furthermore, functional outcomes in many prior studies have been conflated with severity of psychopathology when using tools such as Global Assessment of Functioning (15), and a lack of satisfactory definition of social dysfunction (16, 17). In addition, exposure to antipsychotics over a long period of time alters the nature of speech and our ability to assess FTD (18), thus necessitating the study of minimally treated or drug-naive subjects. Demonstrating this relationship will be of critical value in improving clinical decisions during early intervention based on long-term prognostic outlook, which at present is challenging to assess. To address this crucial gap, we sought to identify linguistic features of speech in an untreated FEP sample using a computational linguistic approach called parts-of-speech tagging implemented through Cohmetrix (19), and the Linguistic Inquiry Word Count (LIWC) (20).

While the feature space for selecting linguistic variables in relation to functional outcomes is relatively large, we focus exclusively on the variables that we have previously studied in an overlapping sample and demonstrated to have clinical relevance. In a prior cross-sectional analysis on this sample of untreated subjects, Mackinley et al. (21) used Coh-Metrix automated speech analysis software (19) to compare FEP patients and healthy controls on a number of variables at the word, sentence, and higher-order level. In this study, patients showed reduced speech production (number of words) and higher pronoun use compared to their healthy control counterparts but did not differ in a variety of other higher-order linguistic metrics (narrativity, formality, referential cohesion, or deep cohesion). Five types of connectives were analyzed in this earlier study including: causal connectives (words used to connect a cause to an effect), logical connectives (words linking two logically connected elements), temporal connectives (words to put ideas in order of time), contrastive connectives (words to compare and contrast ideas), and additive connectives (words used to add information, e.g., “additionally,” “moreover”). The use was analyzed using data driven principal factor analysis, two factors, and one with a positive loading on “all connective types” and the second “acausal temporal connective factor” reflecting reduced use of causal and contrastive connectives, but higher use of temporal linkages and additive connections appeared. While patients and healthy controls employed these connective factors in a comparable manner during the picture description tasks, patients with higher connectives use had higher scores on clinically rated conceptual disorganization (21). This suggests that aberrant linguistic connective use may contribute to the clinician’s detection of disorganized thought.

In an overlapping cross-sectional sample, we (22) analyzed the picture description speech samples using the Linguistic Inquiry Word Count (LIWC) software package (20) to determine the relative proportion of content words and function words. From this parts-of-speech tagging, we determined Pennenbaker’s Analytic Thinking scores (higher scores suggesting a well-formed hierarchical thinking style suitable for academic expressions, and lower scores suggesting a narrative style which is more intuitive and episodic in nature) (23). A higher analytic score (more categorical thinking style) is linked with academic success due to this linguistic style’s use in academic and professional settings (23). We observed that compared to HC, patients showed reduced analytic thinking in their speech. Further, among FEPs, reduced analytic thinking related to higher clinical metrics of disorganization (22). This suggests that less structured, less content-based speech may contribute to the clinician’s detection of disorganized thought. Thus, it is possible that among FEP patients, analytic thinking styles are associated with later academic and occupational success; however, little evidence to assess this question has been gathered.

With longitudinal functional outcome data from this cohort, we aim to ascertain the role of connectives, analytic thinking index, total number of words, and frequency of pronouns on vocational status and social and occupational functioning ascertained after 6-to-12 months of treatment in an early intervention setting. The selected linguistic variables tap on distinct aspects of message generation and grammatical encoding in Bock and Levelt’s language processing model (24). At the generation level, total number of words (verbosity) relates to the production plan (25); at the functional level, lexical selection influences the frequency of pronouns, while positional processing involving the assembly of constituent words influences the connective use and analytic thinking index. Given the prior observations that “negative FTD” relates more strongly to functional outcomes than “positive FTD,” we expected a reduction in total number of words used during a picture description will be predictive of later functional outcomes. To this end, we used a Bayes network (a directed acyclic graph) to (1) identify dependencies among the baseline linguistic variables and vocational status or social functioning after six to 12 months of treatment in an early intervention program for psychoses and (2) parameterize these dependencies in terms of conditional probability distributions. In the network, the dependencies are represented as connections (edges) between nodes (variables) identified through a prototypical constraint-based algorithm (26, 27). Parameters (conditional probability distributions) are found via maximum likelihood estimation (28). We assessed the contribution of other explanatory variables such as parental socioeconomic status and speed of cognitive processing using probabilistic models of functional outcome. We quantified social functioning using the widely used Social and Occupational Functioning Assessment Scale (SOFAS) as a continuous measure, and a macroeconomic indicator of productivity in young adults reflecting participation in active Employment Education or Training (EET vs. not-EET or NEET) status as a categorical measure, as employed in our previous brain imaging study (29).

This study sheds light on how the way we speak when experiencing acute psychosis may provide insight into our occupational/functional outcomes in the first year of early intervention. We report three major findings: (1) Speech production (total number of words spoken) during a three-minute descriptive task at the time of first presentation with psychosis, explained significant variance in NEET status after 6–12 months of treatment; (2) measures of parental socioeconomic status and processing speed did not explain this relationship; and (3) the linguistic features included in our analysis (connectives, pronoun use, and analytic thinking scores) formed their own causal network (i.e., inter-related) but were not related to vocational or social outcomes. Thus, the ability to find a productive vocational status following the experience of psychosis relates to the number of words an individual manages to deploy during a discursive task of describing a picture to another person, irrespective of parental social background, one’s personal speed of processing information and linguistic style of expression, and the severity of core symptoms (PANSS-8 total). These findings supply an objectively detectable and intuitive speech metric that requires no clinical judgment as a prognostic marker of functional outcome. This takes rater-related factors out of consideration when considering prognosis, potentially complementing clinical decisions that may require an assessment of longer-term outcomes (e.g., duration of case management, employment, and placement support).

Individuals with robust speech production had a “protective” effect with respect to functional deterioration. While those with median speech production (113 words) still had an above chance level of poor vocational outcomes (65% NEET), the effects of high speech production on vocational outcomes were far more positive; our modeling would predict that patients with speech production on the upper tail of the distribution (211 words) to have a 99% chance of being vocationally active. There are several hypotheses that could explain this association between the abundance of speech production with good vocational outcome. First, patients with high speech production are far less likely to have broader dysfunction in other negative domains. Poverty of speech has been consistently associated with affective flattening (45) and reduced symptom remission in negative domains (46), as well as likely a marker for underlying cognitive deficits (47). While this contributes a strong case for why lower speech production is likely to impair vocational prospects, it fails to make an affirmative case for good outcomes among those producing higher speech. It is likely that the benefit of speech production to good vocational prospects related to patients with more speech production being rated as more socially adept and desirable by peers and employers. In both healthy control and patient samples, social skills are highly correlated with gaining and retaining competitive employment (48). Among the patient population, the social threshold for employment may in fact be more pronounced as patients are more likely to be involved in the service sector and routine/non-technical occupations where customer or client relations are of primary importance. This speculation warrants further investigation. While “verbosity” may not be readily modifiable among clinical samples, social skills training as part of employment support in first-episode psychosis clinics may yield more robust results among patients who are on the cusp of functioning.

What does this finding mean for the study of speech, language, and communication in psychosis? As noted earlier, clinicians’ rating of disorganization tracks the deviations in grammatical encoding (connectives, pronoun use and analytic thinking index), but the social outcomes relate more with the aspects of message generation or production plan (number of words). Interestingly, a causal network exists among the variables relevant to functional/positional processing, distinct from the word count. We expect future studies to parse the large feature space of computational linguistic metrics to provide further clarity on the message production vs. grammatical encoding components in psychosis.

Our study has several strengths including the assessment of minimally treated FEP subjects, the use of objective linguistic analysis and careful control of known confounders. Nevertheless, several limitations warrant consideration. The use of a binarized NEET status has a few limitations that warrant consideration, including failing to capture “underemployment,” and its inability to capture the complexity of biological, psychological, and social factors underlying vocational outcome, and the potential instability of this metric for patients who experience a relapse of psychotic illness. Despite these limitations, the consistency between NEET status and SOFAS score (which includes a broader definition of functioning) suggests that this construct is indeed a valid measure of functioning, and simple vocational status remains a relevant goal for patients undergoing mental health treatment.

Further limitations include the lack of sufficient longitudinal speech data to assess the stability of ‘verbosity’ over time in this sample, and the lack of information on many mediators of educational/vocational success, e.g., parental support, workplace mentorship, motivational factors, or ratings of social desirability. As a result, our findings pertaining to the value of word counts in forecasting later functioning should be considered complementary information rather than being the best of all baseline predictors of functioning. Such a conservative interpretation also fits with effect-size noted in the primary Bayesian analysis (BF > 2 relating number of words to NEET status). Nevertheless, the use of acyclic graph models on longitudinal data allows us to draw causal inferences (49) from observational design. Finally, two limitations in clinical follow-up are present: the lack of longitudinal antipsychotic medication and the degree of clinical severity at the time of our follow-up vocational assessment. Antipsychotic medications may reduce articulation speed and reduce sentence length in patients with psychosis; thus patients with superior verbal output at baseline may be well positioned to offset any adverse treatment effects from antipsychotic medication, achieving superior vocational outcomes. However, in our sample, we were not able to assess the effect of long-term antipsychotic exposure on speech or clinical severity. Despite this gap, our data revealed that patients with high speech production at baseline tended to do better over time, despite no evidence of systematic differences in antipsychotic exposure or clinical severity at baseline. In future analyses, assessment of the association between baseline speech production, as well as longitudinal antipsychotic exposure and clinical response with vocational outcomes may allow researchers to parse the relationships between these variables which may provide more clarity on the mechanism underlying our observed relationship.

To conclude, we call for including the rate of word production during routine clinical assessments of first-episode psychosis. Our results suggest that this approach, while inexpensive and not requiring exhaustive training, may carry prognostic value above what is currently captured in general clinical practice.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by Western University Research Ethics Board, London, Ontario, Canada. The patients/participants provided their written informed consent to participate in this study.

MM prepared the first draft, recruitment, data collection, and data analysis. RL and AS designed the analysis and implemented statistical procedures and critical review of the paper. HG, JR, and PS: contributed to recruitment and collecting the data, and critical review of the paper. LP conceived the study, designed the analysis, collected the data, and performed critical review of the paper. All authors contributed to the article and approved the submitted version.

This study was funded by CIHR Foundation Grant (FDN 154296) to LP; AMOSO Opportunities fund to LP; Bucke Family Fund and support from the Chrysalis Foundation to LP. LP acknowledges salary support from the Tanna Schulich Chair of Neuroscience and Mental Health (Schulich School of Medicine, Western University: 2019–2022) and a Chercheur-Boursier Award from the Fonds de recherche du Québec-Santé (FRQS, 2022); program support from Monique H. Bourgeois Chair in Developmental Disorders and the Graham Boeckh Foundation (Douglas Research Centre, McGill University). We thank the Jonathon and Joshua Memorial Scholarship for their generous support of MM.

LP receives book royalties from Oxford University Press, editorial stipend from the Canadian Medical Association Journal, and income from the SPMM MRCPsych course. LP has received investigator-initiated educational grants from Otsuka, Janssen, and Sunovion, Canada (2017) and speaker fee from Otsuka and Janssen, Canada (2019), and Canadian Psychiatric Association (2019). LP and MM received support from Boehringer Ingelheim to attend an investigator meeting in 2017. All other authors report no potential conflicts of interest. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.