Natural Language Processing markers in first episode psychosis and people at clinical high-risk

Abstract

Recent work has suggested that disorganised speech might be a powerful predictor of later psychotic illness in clinical high risk subjects. To that end, several automated measures to quantify disorganisation of transcribed speech have been proposed. However, it remains unclear which measures are most strongly associated with psychosis, how different measures are related to each other and what the best strategies are to collect speech data from participants. Here, we assessed whether twelve automated Natural Language Processing markers could differentiate transcribed speech excerpts from subjects at clinical high risk for psychosis, first episode psychosis patients and healthy control subjects (total N = 54). In-line with previous work, several measures showed significant differences between groups, including semantic coherence, speech graph connectivity and a measure of whether speech was on-topic, the latter of which outperformed the related measure of tangentiality. Most NLP measures examined were only weakly related to each other, suggesting they provide complementary information. Finally, we compared the ability of transcribed speech generated using different tasks to differentiate the groups. Speech generated from picture descriptions of the Thematic Apperception Test and a story re-telling task outperformed free speech, suggesting that choice of speech generation method may be an important consideration. Overall, quantitative speech markers represent a promising direction for future clinical applications.

Semantic speech networks linked to formal thought disorder in early psychosis

“…As expected, in a clustering analysis, the number of nodes and edges in the semantic speech networks clustered with the number of words. The number and size of connected components formed their own community, independent of previously reported NLP measures (14) including syntactic graph measures (11) and semantic coherence. The connected components in semantic speech networks might therefore capture signal beyond information contained in established NLP measures or relating to the amount of speech.…”

“…Basic transcript measures were number of words, number of sentences and mean sentence length. Established NLP measures included Tangentiality, Ambiguous Pronouns, Semantic Coherence (Coherence), On-Topic Score (On Topic) taken from (14). Additionally, syntactic network measures were taken from (14) and included number of nodes in the largest strongly connected component of syntactic networks (LSC), number of nodes in the largest weakly connected component of syntactic networks (LCC), as well as the LSC and LCC normalised to random networks (LSCr, LCCr) (12)…”

“…We compared our semantic speech graph measures to other NLP approaches that previously showed significant group differences in the clinical dataset (14). Specifically, we calculated semantic coherence, tangentiality, on-topic score, ambiguous pronoun count, and connectivity measures from the syntactic speech graphs proposed by (12).…”

“…Healthy controls were matched to the CHR-P and FEP participants for age and gender. Demographics for the three groups have been reported elsewhere (22) (14).…”

“…However, these networks focus on the syntactic characteristics of speech and largely ignore semantic content. Mean-while, measuring semantic coherence has been found to predict psychotic disorders (8, 14), but only provides a single summary measure of an entire speech transcript. Speech excerpts likely include additional semantic information that could be used to build a deeper understanding of how the semantic content of language is altered in schizophrenia; and provide extra power to predict outcome or relapse for individual patients.…”