Clinical psychopathology-based early relapse prediction model using speech and language in psychosis

Abstract

Introduction: Prediction of psychotic relapse using speech-derived markers promises targeted early intervention. However, the sheer number of speech markers and the 'black box' nature of predictive models challenges clinical translation.

Methods: We propose a psychopathology-based systematic approach to identify likely relapse. We draw on the notion that the predictors of relapse should mark (1) the presence of schizophrenia in its untreated early stages and (2) track disorganization in psychosis. By leveraging Natural Language Processing, we derive 3 lexical, syntactic and narrative markers -semantic similarity, clause complexity, and analytic thinking index from speech samples of people with acute psychosis (n = 68) followed up for subsequent relapses over a year (12 out of 68).

Results: Speech-based model predicted relapse status with strong evidence (Bayes Factor BF₁₀ = 79.5) against the clinical intuition model.

Conclusion: Using a Bayesian approach, this preliminary study demonstrates the utility of psychopathology-guided variable selection for speech-based relapse prediction complementing clinical intuition in practice.

Keywords: Computational linguistics; Disorganization; Early intervention; Impoverishment; Precision psychiatry; Relapse; Thought disorder.

Fig. 1

Psychopathology-based prediction of clinical outcomes using speech data. Left: Computational psychiatry aims to predict clinical outcomes using a data-driven or theory-driven framework. When faced with high-dimensional data with multiple features, data reduction can occur at preprocessing phase (data reduction), or during predictive model building (complexity reduction). Alternatively, mechanistic theories (e.g., drift-diffusion model of evidence accumulation) can provide low dimensional parameters that capture the underlying biophysical or cognitive processes. Even for successful data/theory-driven predictions, clinical sense-making remains a challenge (black or greybox solutions). Right: Psychopathology-based models of clinical outcomes take a top-down approach. We first start with identifying the psychopathological phenomenon that is relevant for the outcome of interest. For example, to model the outcome of future psychotic relapses at the first presentation, the selected speech-NLP features must be more prominent during the initial first episode state (state/trait effect), vary with positive but not negative FTD or explained by medications (causal symptom). This filtration helps to first identify the features that are most likely to relate to the psychopathological phenomena, followed by model evaluation. The resulting predictions are more likely to be clinically accessible, as they follow what a clinician expects to see with the outcome of interest. Of note, the selection process can begin as a data-driven search (e.g., check the NLP feature that best correlates with the psychopathological features of interest, showing periodicity, associated with positive FTD, dissociated from negative FTD and not affected by medication confounds, in the case of psychotic relapses).

Fig. 2

Visual representation of the methodology. Patients referred to the first episode psychosis clinic (PEPP, London, Ontario) from various sources in the community (self-referral, social worker, family doctors or other physicians). 1) Clinical interviews were conducted in the first week of presentation. 2) Speech samples were obtained using the Thought and Language Index procedures. 3) Speech data transcribed and analyzed by three different language processing tools. 4) A 12-month retrospective chart review after program entry and clinical stabilization is conducted to determine whether the patient relapsed (hospitalization within the next year), and a consensus diagnosis is confirmed. 5) Patients were grouped based on relapse or no-relapse; speech and clinical variables were collected at baseline and then correlated with the later relapse status.

Fig. 3

Sequential analysis of construct identification. JASP output of sequential analysis shows the evidential flow in favour of H1 for disorganization scores (left) from the Thought and Language Index (TLI) being different (higher) among those who showed an early relapse later vs. those who did not. The evidence favours a lack of difference between the relapse/non-relapse groups for impoverishment scores (right). Each point indicates a change in the BF when a participant is added in the analysis (n in x axis). BF = Bayesian Factor. H₀ - null hypothesis (of no difference). H₁ - alternative hypothesis.

Fig. 4

Model inclusion probabilities. JASP output of model inclusion probabilities for A) Language-based vs. clinical profile (cannabis-smoking man with premorbid educational deficits and low functioning) B) Combined language and clinical profile-based model v. clinical intuition (intercept only) model. Bars indicated model-averaged posterior inclusion probabilities for the Bayesian linear regression with dashed lines representing the prior probabilities (before the data is evaluated). CAST: Cannabis Abuse Screening Test score, SOFA: Social and Occupational Functioning Assessment Scale score, TLI: Thought and Language Index.