Deciphering language disturbances in schizophrenia: A study using fine-tuned language models.

ABSTRACT

This research presents two stable language metrics, namely Successful Prediction Rate (SPR) and Disfluency (DF), to objectively quantify the linguistic disturbances associated with schizophrenia. These novel language metrics can capture both off-topic responses and incoherence in patients' speech by modeling speech information and fine-tuning techniques. Additionally, these metrics exhibit cultural sensitivity while providing a more comprehensive evaluation of linguistic abnormalities in schizophrenia. This research fine-tuned the ELECTRA Pretrained Language Model on a 750 MB text corpus obtained from major Chinese mental health forums. The effectiveness of the fine-tuned language model is verified on a group comprising 38 individuals diagnosed with schizophrenia and 25 meticulously matched healthy controls. The study explores the association between the fine-tuned language model and the Positive and Negative Syndrome Scale (PANSS) items. The results demonstrate that SPR is higher in healthy controls, indicating better language understanding by the pre-trained language model. Conversely, DF is higher in individuals with schizophrenia, indicating more inconsistent language structure. The relationship between linguistic features and P2 (conceptual disorganization) reveals that patients with positive P2 exhibit lower SPR and higher DF. Binary logistic regression using the combined SPR and DF features achieves 84.5 % accuracy in classifying P2, exceeding the performance of traditional features by 20.5 %. Moreover, the proposed linguistic features outperform traditional linguistic features in discriminating FTD (formal thought disorder), as demonstrated by multivariate linear regression analysis.