Burnishing the blueprint of speech assessment with natural language processing: methods to characterize subtle impairments on individuals in at-risk mental states from a large urban population.

Abstract

Spoken language is a key source of information for thought disorder evaluation. In the last decades, researchers linked psychopathology phenomena to their counterparts in natural language processing (NLP) analysis. Nonetheless, seemingly opposite traits remain unconciliated. For instance, psychotic speech comprises incoherent trails, but also highly associated ones. In order to address some of the remaining gaps, we leveraged procedures from dynamical systems and graph theory. We examined transcribed interviews of 133 individuals — 60 in at-risk mental states (ARMS) and 73 healthy controls — screened from 4,500 quota-sampled citizens in a large metropolis. SIPS was used to assess psychotic symptoms. NLP features were correlated with psychotic traits (Spearman’s ρ) and ARMS status (Wilcoxon signed-rank tests, general linear models and ensemble machine learning algorithms). The general trait (ω), negative, disorganized and general symptoms were correlated with snippets made of consecutive similar words. Namely, their frequency, average/maximum size, heterogeneity and the average number of unrelated words between such snippets. Positive symptoms were associated with adjective use. Average graph centrality was inversely correlated with the general trait. NLP features presented good performance as input in machine learning classification using the AdaBoost model with Random Forests as base learner (F1 score: 0.83, AUC: 0.93, Balanced Accuracy: 0.86).The existence of loosely connected words (e.g. incoherence, looseness, derailment) is well studied. Conversely, NLP models of perseveration (e.g. higher likelihood of chaining together islands of closely related words) and circumstantiality are brought forth in this work. Evidence shows good performance of NLP for clinical decision support in ARMS screening and assessment of subclinical psychosis. We show that a blueprint for speech-based psychometric evaluation is only a few pieces away. We highlight these fields for future research: clanging (a low hanging fruit), environmental context, task-related differences and interpersonal interactions.