To investigate the association of computational speech and language markers with psychotic symptom severity in individuals with schizophrenia spectrum disorders.
Approach:
Study Design: Longitudinal, multicenter cohort study involving 773 brief speech recordings from 356 participants in the Netherlands and 165 English-language recordings from 72 acute care inpatients in the US.
Data Analysis: Automated pipeline extracted over 100 acoustic, semantic, syntactic, and sentiment features from speech, which were then reduced to interpretable components using principal component analysis.
Key Findings:
Clinically meaningful estimations of symptom severity were achieved, with mean absolute errors (MAEs) of 2.85 and 3.22 for positive and negative symptoms in the Dutch cohort, respectively. Item-level MAEs ranged from 0.21 to 0.94.
In the US cohort, MAEs for thought disturbance and withdrawal were 3.00 and 2.44, respectively.
Positive symptoms correlated with longer utterances and altered discourse organization, while negative symptoms were linked to reduced speech output and flatter acoustic profiles.
Interpretation:
Speech may serve as a biosocial marker of psychosis, providing insights from low-level acoustics to higher-level semantics.
Limitations:
Extraneous factors like antipsychotic medication may influence speech characteristics, complicating symptom attribution.
Generalizability of findings needs further validation across diverse settings and populations.
Conclusion:
Future studies should focus on the robustness, interpretability, and actionable thresholds of prediction models in real-world clinical settings.
