Detecting formal thought disorder by deep contextualized word representations.

Area of Science:

• Computational linguistics
• Psychiatry
• Artificial intelligence in healthcare

Background:

• Schizophrenia symptoms, including formal thought disorder (FTD), can be objectively measured using computational linguistics tools.
• Previous methods relied on individual linguistic indicators, with varying accuracy in detecting FTD.

Purpose of the Study:

• To compare the accuracy of neural network-based utterance embeddings (ELMo) against traditional coherence models for detecting FTD in schizophrenia.
• To evaluate the performance of ELMo and coherence models against a clinician's assessment.

Main Methods:

• Utilized Embeddings from Language Models (ELMo) to represent speech data from schizophrenia patients (N=35) and healthy controls (N=35).
• Compared ELMo performance to a coherence model (Bedi et al., 2015) and a clinician's ratings using the Scale for the Assessment of Thought, Language and Communication (TLC).

Main Results:

• ELMo achieved 80% accuracy in distinguishing schizophrenia patients from controls, outperforming the coherence model (70%) and the clinician (74%).
• Both ELMo and TLC assessments demonstrated sensitivity to disorganization symptoms in patients.
• Language model-based text representations were more effective than FTD-based assessments.

Conclusions:

• Neural network-based language models, specifically ELMo, offer a more accurate computational tool for detecting formal thought disorder in schizophrenia.
• These advanced linguistic analysis methods can serve as valuable complements to traditional clinical evaluations.
• Objective linguistic measures enhance the understanding and assessment of disordered language in psychiatric conditions.