Towards transcribed speech networks as a marker for psychosis risk

Abstract

Recent work has shown that incoherent speech is a powerful predictor of psychosis. Predicting psychosis risk from speech data could revolutionise healthcare for psychosis, for two reasons. First, there is a pressing clinical need for tools to aid diagnosis and monitoring of disease trajectory. Second, speech data is non-invasive, cheap and simple to collect longitudinally. Existing tools have used networks to measure disorganised speech in speech transcripts from psychosis patients. In these networks, nodes are words and edges connect words spoken in succession (Mota, Copelli and Ribeiro, 2017). However, since these networks link only temporally co-occurring words, these tools ignore the semantic content of speech, which is often altered in patients with psychotic disorders. We therefore developed a novel speech network algorithm to capture the semantic content of transcribed speech. Our tool builds on Natural Language Processing libraries (StanfordNLP, Manning et al. 2014; Openie5, Mausam, 2016) to construct speech networks from transcripts of spoken text (e.g. I see a man). Nodes are entities (e.g. I, man). Edges are relations between nodes (e.g. see). The algorithm can capture the information content of a full paragraph of speech, even when semantically connected nodes are separated by several sentences. The algorithm is fast and is robust against artefacts typical for transcribed speech, lending itself to the automated construction of speech networks from large datasets. Having developed our novel speech network tool, we used it to characterise the properties of speech networks from healthy participants describing ambiguous pictures from the Thematic Apperception Test (TAT; Murray, 1943). For example, we assessed the size, clustering and average path length of the networks, and compared our results to randomised networks. Finally, we present preliminary findings showing altered speech network properties in psychosis patients compared to healthy volunteers. Our novel algorithm captures characteristics of speech that are characteristically altered in psychosis but have been difficult to quantify in an automated way. It thereby offers a new perspective on the nature of incoherent speech in psychosis. Ultimately, speech markers derived from this tool could lead to significant advances in disease prediction and clinical practice.

Caroline Nettekoven

Postdoctoral Researcher

I am interested in the neural basis of complex behaviour. To study this, I use neuroimaging techniques, computational modelling of behaviour and brain stimulation.