From clinical records to conversations: generative artificial intelligence for symptom extraction and assessment

Recent advances in large language models (LLMs) enable privacy-preserving clinical applications in mental health. The authors’ recent work demonstrates the feasibility of deploying locally hosted generative AI models for information extraction within hospital environments [1,2]. Using prompt engineering strategies, core psychopathological domains were systematically extracted from electronic health records, enabling prediction of psychiatric diagnoses and stratification of patients according to risk, including the likelihood of psychiatric emergencies [Verhees et al., in review].

Furthermore, the authors conducted simulation studies in which state-of-the-art models (Llama 3 and GPT-4o) assumed the roles of both clinicians and patients, demonstrating strong alignment with clinician assessments in the evaluation of psychiatric interviews. Preliminary results from an ongoing clinical validation study of an AI-driven chatbot for the assessment of depressive symptoms will be presented.

As further aim, the authors propose a clinically validated, generative AI based conversational agent that serves as a broadly applicable clinical decision support (CDS) system capable of accurately and continuously assessing mental health symptoms.

References:

[1] Wiest IC, Verhees FG, Ferber D, Zhu J, Bauer M, Lewitzka U, et al. Detection of suicidality from medical text using privacy-preserving large language models. The British Journal of Psychiatry 2024; 225: 532–7.

[2] Mikolas P, Huth F, Bröckel-Bundt K, Berndt C, Martini J, Maicher B, et al. Predicting functional outcome using machine learning and large language models: A transdiagnostic approach in a young, help-seeking population (In press JMIR Mental Health). 2026. doi:10.2196/preprints.84424.