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

Area of Science:

• Computational linguistics
• Psychiatry
• Artificial Intelligence

Background:

• Schizophrenia diagnosis often relies on subjective assessments.
• Objective linguistic metrics are needed for better evaluation of schizophrenia.
• Existing methods may lack cultural sensitivity or comprehensive analysis.

Purpose of the Study:

• To introduce two novel, stable language metrics: Successful Prediction Rate (SPR) and Disfluency (DF).
• To objectively quantify linguistic disturbances in schizophrenia.
• To evaluate the cultural sensitivity and comprehensiveness of these metrics.

Main Methods:

• Fine-tuned the ELECTRA Pretrained Language Model on a Chinese mental health forum corpus.
• Validated the model on 38 schizophrenia patients and 25 healthy controls.
• Analyzed the association between linguistic features and Positive and Negative Syndrome Scale (PANSS) items.

Main Results:

• Healthy controls showed higher SPR; schizophrenia patients showed higher DF.
• Lower SPR and higher DF correlated with conceptual disorganization (P2).
• Combined SPR and DF achieved 84.5% accuracy in classifying P2, outperforming traditional features by 20.5%.

Conclusions:

• SPR and DF are effective, culturally sensitive metrics for assessing schizophrenia-related linguistic abnormalities.
• These metrics significantly improve the objective classification of thought disorders like conceptual disorganization.
• The proposed linguistic features offer a more accurate alternative to traditional methods for schizophrenia assessment.