Assessing dimensions of thought disorder with large language models: The tradeoff of accuracy and consistency

Natural Language Processing (NLP) methods have shown promise for the assessment of formal thought disorder, a hallmark feature of schizophrenia in which disturbances to the structure, organization, or coherence of thought can manifest as disordered or incoherent speech. We investigated the suitability of modern Large Language Models (LLMs - e.g., GPT-3.5, GPT-4, and Llama 3) to predict expert-generated ratings for three dimensions of thought disorder (coherence, content, and tangentiality) assigned to speech samples collected from both patients with a diagnosis of schizophrenia (n = 26) and healthy control participants (n = 25). In addition to (1) evaluating the accuracy of LLM-generated ratings relative to human experts, we also (2) investigated the degree to which the LLMs produced consistent ratings across multiple trials, and we (3) sought to understand the factors that impacted the consistency of LLM-generated output. We found that machine-generated ratings of the level of thought disorder in speech matched favorably those of expert humans, and we identified a tradeoff between accuracy and consistency in LLM ratings. Unlike traditional NLP methods, LLMs were not always consistent in their predictions, but these inconsistencies could be mitigated with careful parameter selection and ensemble methods. We discuss implications for NLP-based assessment of thought disorder and provide recommendations of best practices for integrating these methods in the field of psychiatry.