Psychotic disorders, particularly schizophrenia, are chronic illnesses often characterized by recurrent relapses. These relapses significantly impair patients’ social and occupational functioning, increase rates of hospitalization, and impose additional economic burdens on healthcare systems. In clinical practice, the ability to predict which patients are at risk of early relapse is critical for developing targeted interventions. Traditional clinical indicators—such as demographic characteristics, functional status, or substance use—often provide insufficient predictive power and are limited in their capacity to accurately identify patients at risk of early relapse.
In recent years, research in this field has increasingly focused on the computational analysis of speech and language data. The study by Dalal et al. (2025) demonstrates that linguistic features extracted from speech samples may provide stronger predictive information about early relapse in psychosis than conventional clinical intuition. Unlike traditional “black-box” artificial intelligence models, this approach emphasizes the selection of interpretable variables that are theoretically aligned with psychopathology and clinically meaningful.
The central hypothesis of the study is that early relapse may be associated with positive formal thought disorder (FTD). Positive FTD manifests as disorganized thinking, semantic incoherence, and speech abnormalities. The researchers proposed that linguistic features reflecting these cognitive disturbances could serve as indicators of relapse risk.
The study included 68 first-episode psychosis (FEP) patients enrolled in an early psychosis program in Canada. Speech samples were collected prior to prolonged antipsychotic treatment during a three-minute picture description task. These samples were analyzed at three linguistic levels:
(1) Semantic Similarity, measuring the extent to which words in speech are semantically related;
(2) Analytic Thinking Index, reflecting the structural and logical coherence of speech; and
(3) Clause Complexity, quantifying the grammatical complexity of speech.
Patients were followed for one year, and hospitalizations due to psychotic symptoms were considered relapses. During this period, 12 patients experienced early relapse.
The results indicated that patients with higher levels of positive FTD were more likely to experience relapse, whereas negative FTD indicators (e.g., impoverished speech) were not significantly associated with relapse. Analysis of speech features showed that the Analytic Thinking Index was strongly related to positive FTD and thus emerged as a significant predictor of relapse. Semantic Similarity was associated with negative FTD and was excluded from the relapse prediction model. Clause Complexity was moderately related to both types of FTD and emerged as the strongest predictive variable in the model.
The speech-based model outperformed demographic and clinical variables commonly used by clinicians (including gender, education, substance use, and functional status). The Bayes factor indicated that the speech-based model was 79 times more informative than the clinical intuition model. These findings suggest that even short speech samples can provide valuable information for predicting early relapse risk. This highlights the potential of speech analysis as a clinically practical and scalable assessment tool.
The clinical implications of this research are substantial. A three-minute speech sample alone may allow clinicians to predict whether a patient is at risk of relapse within the following year, providing a rapid, low-cost, and non-invasive screening tool for high-demand clinical settings. Furthermore, the use of psychopathology-informed, interpretable variables may enhance the acceptance of AI and computational models in clinical practice and improve their integration into decision support systems.
Several limitations should be noted. The small sample size and low number of relapses constrain the generalizability of the model. Speech was measured only once, and temporal changes or dynamic speech patterns were not captured. Additionally, relapse was defined solely by hospitalization, potentially overlooking milder symptom exacerbations.
In conclusion, the study by Dalal et al. demonstrates that speech and language analysis is a powerful and practical tool for predicting early relapse in psychosis. Psychopathology-based, low-dimensional, and interpretable models may play a significant role in clinical decision support. Future studies with larger samples and real-time speech analysis may further facilitate the clinical adoption of this approach.
Reference
Dalal, T. C., Park, M. T. M., Silva, A. M., Iskhakova, S., Voppel, A., Brierley, N. J., MacKinley, M., Olarewaju, E., & Palaniyappan, L. (2025). Clinical psychopathology-based early relapse prediction model using speech and language in psychosis. Schizophrenia Research: Cognition, 100392. https://doi.org/10.1016/j.scog.2025.100392