Language disturbance is a biomarker for psychosis. Prior studies utilizing natural language processing (NLP) methods to analyze speech find that semantic coherence predicts psychosis onset in high-risk youth, and linguistic complexity/idea density is related to psychiatric symptoms and functioning in psychosis. Few studies have examined the relationship of NLP-derived variables to neuropsychological language variables in the early psychosis spectrum. We aimed to examine the relationship among schizotypal traits, NLP, and neuropsychological language variables. We hypothesized semantic coherence would be related to semantic-based language measures; idea density would be related to executively-mediated language measures; and all language variables would be negatively related to schizotypal traits. Although language skills are integral to social functioning, scant research exists exploring the differential relationships of NLP and neuropsychological language measures to social functioning. We also explored whether neuropsychological variables accounted for additional variance in social functioning, after controlling for NLP variables. We hypothesized neuropsychological variables would jointly account for additional variance in social functioning.

We recruited 80 participants high and low in schizotypal traits from undergraduate courses, and 15 participants who recently experienced a first psychotic episode from a psychosis intervention center. On average, participants were 21.44 (4.18)-years-old and completed 14 (1.57) years of education. A majority were female (67%) and White (82%). Participants completed the following measures/tasks: Schizotypal Personality Questionnaire, Similarities, Proverb Test, semantic fluency, phonemic fluency, Stroop Word Reading, GF: Social (examiner-rated social functioning) and Social Skills Performance Assessment (SSPA; performance-based) and a personal narrative task. Narrative data were transcribed and analyzed via NLP methodologies to derive semantic coherence (latent semantic analysis; bidirectional encoder representations from transformers [BERT]) and idea density (number of ideas divided by word count) variables. We utilized age-normed neuropsychological scores and raw total scores for other measures. Biological sex and verbal processing speed were entered as in step one of regression models (covariates) due to group differences.

In correlational analyses, NLP variables were unrelated to neuropsychological variables and schizotypal traits. Proverb Test and semantic and phonemic fluency were negatively related to total schizotypal traits. In regression analyses, after controlling for covariates, NLP variables jointly accounted for significant variance in GF: Social, p = .05, ΔR² = .07. Greater semantic coherence via BERT was associated with better social functioning. Controlling for all other variables, neuropsychological variables did not account for additional variance in the GF: Social. In the second model, after controlling for covariates, NLP variables jointly accounted for significant variance in SSPA, p = .01, ΔR² = .10; Greater semantic coherence via BERT was associated with better social functioning. Controlling for all other variables, neuropsychological variables accounted for marginally more variance in SSPA, p = .08, ΔR² = .07; better Proverb Test performance was associated with better SSPA performance, p = .01.

NLP and neuropsychological language variables were unrelated to one another, but differentially related to social functioning. These findings underscore the utility of both metrics in language analysis. Future work should replicate these results in larger samples, incorporate longitudinal designs, and increase recruitment in other demographic groups.