Personen und Einrichtungen

Datum

Beschreibung

Large language models (LLM) are a type of artificial intelligence algorithm that is versatile in the generation and processing of natural language without prior training for specific tasks. Their use does not depend on training corpora that are partly unavailable outside the English language. They are thus ideally placed to deal with large amounts of written text produced daily in clinical settings, foremost among them in psychiatry. Their use may include extraction of unstructured information into machine readable form for research or even clinical purposes: In our proof-of-concept study, locally deployed LLMs automatically detected suicidality in real world clinical notes of acute psychiatric patients from a large German hospital. Llama 2, an open source LLM, was able to detect suicidality with a sensitivity of 83.0% and specificity of 91.8% in a fine-tuned version for the German language without any further medical training and only basic prompt-engineering. We adapted an LLM based information extraction pipeline for the analysis of various psychopathological characteristics beyond our current approach, including mood, anxiety and psychosis with the potential of predicting disease trajectories or supporting therapy decisions. In the European Union at least, clinical applications of cloud-based LLMs are currently strictly limited due to the sensitive nature of patient data, as mental diseases are especially often considered to carry social stigma. Our approach demonstrates an effective implementation using only local computing resources, without the need for personal data to leave the hospital.