Loss of human life by suicide - lived experiences of suicide survivors and professionals.

Mental status assessment in psychiatry has always relied on spoken or written language to infer signs of mental illness. The way a person speaks gives insight into how their mind works, into underlying biological and cognitive processes, emotional states, and, in cases of significant changes in someone’s speech, can represent the onset or worsening of mental illness. This is particularly true for psychotic disorders where, in the case of schizophrenia, disorganized speech is at the core of the illness, and most traditional assessments of psychotic symptoms have an emphasis on linguistic symptoms such as impoverished or incoherent speech. In recent years, Natural Language Processing (NLP) has gained traction in mental health research, and numerous studies have employed NLP techniques to model and analyze speech patterns of individuals with psychosis, trying to find linguistic markers of the illness and its course. The potential is transformative: automated systems that integrate speech collection, automatic speech recognition, and real-time analysis of linguistic markers promise earlier and more effective intervention through automated symptom monitoring. However, real-world implementation requires rigorous benchmarking. In this study, we therefore evaluate the performance, robustness, and generalizability of commonly used NLP markers of psychosis to accelerate their translation into clinically meaningful applications. By systematically evaluating these issues, this work provides a reference point for future NLP studies in psychosis and beyond. Many prior NLP studies have focused on proof-of-concept demonstrations, showing that newly developed metrics can (i) differentiate individuals with psychosis from healthy controls, (ii) correlate with clinical ratings of symptom severity, and (iii) predict psychosis onset in high-risk individuals. These approaches have helped establish that NLP metrics can be linked to clinically relevant phenomena. NLP studies to date have aimed to find markers of psychosis in speech collected through various speech elicitation tasks (e.g., open questions vs. picture description tasks) as well as data from electronic health records or social media posts. Analysis of social media data specifically raises ethical concerns as users have not consented to analysis of their data for characteristics that could indicate a mental illness. Therefore, we will focus on NLP analysis of speech data collected from individuals with psychosis who have provided informed consent. We also used several elicitation tasks in each dataset, as the metric performance has previously been reported to vary across different speech elicitation tasks.