Clinical evaluations of speech are a fundamental and highly informative component of psychiatric assessments; however, according to a session at the American Psychiatric Association 2023 Annual Meeting, conventional assessments are subjective and lack precision.
In their presentation titled “Diagnosing Schizophrenia in the 21st Century: Natural Language Processing as an Emerging Biomarker,” presenters discussed how artificial intelligence advances, specifically quantitative voice analysis and natural language processing (NLP), have allowed researchers to objectively detect and characterize alterations in speech and language, which they call “disfluencies.”
While this quantitative approach to speech evaluation is still being explored, the speakers described the technology’s growing utility in psychiatric disorders. They highlighted schizophrenia spectrum disorders (SSD), where strong evidence has supported the ability of NLP to predict onset of psychosis and distinguish patients with psychosis from healthy individuals.
The speakers noted that current computational speech analyses look at the acoustics and phonetics of the physical sound waves that are generated when speaking. These assessments look at characteristics like pausing, rate of speaking, pitch, and volume, and variations within those characteristic, as well as organization, complexity, and content of speech.
Researchers noted a lot of the NLP analysis was built on previous work characterizing speech incoherencies, though they added NLP could also be combined with other assessments like facial expressions, body movements, and other brain signals.
In one presenter’s pilot study, NLP was used to analyze automatically transcribed recordings to objectively identify features of speech and language that could be linked to clinical measures of importance.
The study included 11 controls and 20 patients with SSD who were otherwise clinically stable in terms of thought disorders. Researchers assessed patients’ speech on a word level, a syntactic level, and a sentence-to-sentence organization level and used those features to predict whether those patients had a schizophrenia diagnosis.
According to the presenter, patients with schizophrenia tended to use first-person singular pronouns, while healthy controls tended to use first-person plural pronouns. Patients with schizophrenia had more incomplete words and cut-off sentences and used fewer adverbs and adjectives compared with healthy controls.
The study’s authors used a deep learning, large language model to look at how cohesive sentences were from one to the next. The model found healthy controls tended to remain focused on an interviewer’s initial prompt, whereas patients with schizophrenia tended to deviate more and more as they continued speaking.
Researchers noted that the standard clinical features predicted a diagnosis of schizophrenia with 68% accuracy, while their model based on NLP-derived features showed an accuracy of 87% in identifying patients with schizophrenia.
The study ultimately concluded that “natural language processing-derived features predicted schizophrenia diagnosis more accurately than gold standard clinical rating scales,” suggesting that NLP may be more sensitive to subclinical disturbances in speech and language than clinician interpretations or ratings.