OBJECTIVES: The primary goal of this study is to develop and validate a speech-based
digital model to predict psychotic relapses in individuals with early psychosis. The
study specifically aims to:
Test the hypothesis that within-subject changes in speech coherence, connectedness, and
complexity, as measured by natural language processing (NLP) tools, will accurately
identify imminent relapse, up to four weeks before clinical relapse in individuals
receiving care in Early Psychosis Intervention (EPI) programs.
Investigate whether these speech-based relapse prediction models generalize across
different languages (English and French) and are equally predictive in both males and
females, addressing potential sociodemographic and linguistic influences on model
performance.
Explore whether combining acoustic and prosodic features with core NLP-based speech
measures improves the model's sensitivity and specificity for relapse prediction.
METHODS:
This study will employ a longitudinal, prospective design involving 250 first-episode
psychosis (FEP) patients recruited from three Early Psychosis Intervention (EPI) clinics
in Ontario and Quebec. The study aims to develop and evaluate a speech-based relapse
prediction model, with a particular focus on generalizing results across different
languages (English and French) and genders.
Participant Recruitment and Stratification:
Participants: A total of 250 FEP patients, including both English- and French-speaking
individuals, will be enrolled to ensure linguistic diversity. The sample will be
stratified by sex to evaluate model performance across genders.
Language groups: Approximately 60% of the participants will be English speakers and 40%
French speakers, reflecting the population served by the EPI clinics.
Gender representation: The study aims to ensure that at least 40% of participants are
female to assess gender-based differences in model prediction performance.
Baseline Assessments:
At baseline, participants will undergo a comprehensive in-person assessment to collect a
detailed profile for each patient. This will include psychiatric symptomatology using the
Positive and Negative Syndrome Scale (PANSS), Calgary Depression Scale and the Personal
and Social Performance (PSP) scale, and cognitive functioning. Additionally,
socioeconomic variables, historical and current medication usage, substance use (e.g.,
cannabis), and treatment adherence will also be recorded to provide a full clinical and
treatment profile for each participant.
Speech Sampling and Data Collection:
Monthly Speech Samples: After the baseline assessment, participants will provide monthly
speech samples over the course of 24 months. These speech samples will be collected using
web-based prompts that include open-ended tasks, such as picture description or recall
narratives, designed to elicit spontaneous speech.
Attrition and Speech Sample Estimates: Given an expected attrition rate of 35-50%, it is
estimated that by the end of the study, 840-960 speech samples will be obtained from
English-speaking participants and 660-870 speech samples from French-speaking
participants.
Speech Analysis:
The collected speech samples will be analyzed using natural language processing (NLP)
methods to extract key features associated with psychosis, including coherence (Measured
by lexical predictability), Connectedness (Assessed using speech graph analysis) and
Complexity (evaluated using the Analytic Thinking Index (ATI)). These NLP-derived speech
metrics will be tracked over time to predict imminent psychotic relapses and compared
across subgroups to assess the impact of language and gender on the predictive accuracy
of the relapse model.
Data Analysis and Generalization:
The primary objective is to determine whether speech-based relapse prediction models
generalize across different languages and genders. To achieve this, model performance
will be evaluated across subgroups:
Linguistic subgroup analysis will compare the model's performance in English- and
French-speaking participants.
Gender-based analysis will assess whether the predictive power of the speech-based model
varies between male and female participants.
This analysis will ensure that the final model can be generalized across diverse
populations and adapted for use in different clinical settings.