Can speech reveal mental health status? A study of linguistic features across multiple samples of at-risk adults

An urgent challenge in clinical science is the reliable detection of psychiatric risk in the aftermath of traumatic or stressful events. Current screening tools rely on patient symptom reports that are low in specificity and may confound normative reported distress with symptom escalation. The consequences of this lack of precision are significant, and there is a need for novel tools that will detect not only individuals at risk for psychiatric disorders but also those whose adaptation represents resilience and/or normative recovery pathways. In the current investigation, we applied natural language processing and large language models to transcripts of adult individuals (N = 555) describing recent experiences via semistructured interviews to identify relevant features of language predictive of psychiatric risk. Participants were pooled from five different investigations and were all at elevated risk for psychopathology due highly stressful contexts and/or traumatic event exposure. We applied large language models to characterize the emotional content, overall sentiment, and autobiographical memory type of their interview responses to distinguish individuals with either persistent depressed mood or anhedonia (derived via separate diagnostic interviews). AI-detected features were consistent with prior in-lab emotion assessments. We then tested machine learning models to determine the extent to which each feature would be sensitive to the presence of disorder. The results suggest that emotional content and broad indices of sentiment offer a relatively high level of sensitivity when modeling diagnostic status, including the presence of major depressive disorder as well as any psychiatric disorder. Moreover, results also suggested certain content may be more relevant for risk detection (i.e., discussions of coping rather than a stressful event). Together, these findings advance the potential for applying AI-driven tools in psychiatric surveillance. (PsycInfo Database Record (c) 2026 APA, all rights reserved).