ViPATalk meets AI: Advancing Multimodal Virtual Anamnesis Simulation using LLM Fine-tuning
Gesellschaft für Informatik (GI) · 01.01.2025
DOI: 10.18420/delfi2025_35 ↗Kurz gefasst
Fachartikel in der Zeitschrift Gesellschaft für Informatik (GI), erschienen am 01.01.2025. Verfasst von Martin Sboron und 3 weiteren Autor:innen.
Abstract (Originalsprache)
Virtual simulated patients (VSP) have become an education tool for medical students and are implemented by frameworks of varying degrees of immersion. The path to free-form interaction between the user and the VSP has been of ever-growing interest. For a step in this direction, we extend the ViPATalk framework by Lippitsch et al. by replacing the current rule-based system for question categorization with a fine-tuned Large Language Model (LLM). We demonstrate that smaller LLMs perform similarly to larger models, when trained in the same domain, and that recent base models perform similarly to their language-specific fine-tuned counterparts. Further, our findings indicate that generated (i.e. synthetic) training data can be helpful when little data is available for certain categories.
Schlagwörter
Zitieren
APA
Sboron, M., Thews, O., Schmid, & Thomas (2025). ViPATalk meets AI: Advancing Multimodal Virtual Anamnesis Simulation using LLM Fine-tuning. Gesellschaft für Informatik (GI). https://doi.org/10.18420/delfi2025_35
Vancouver
Sboron M, Thews O, Schmid, Thomas. ViPATalk meets AI: Advancing Multimodal Virtual Anamnesis Simulation using LLM Fine-tuning. Gesellschaft für Informatik (GI). 2025. doi:10.18420/delfi2025_35
BibTeX
@article{Sboron2025,
author = {Sboron, Martin and Thews, Oliver and Schmid and Thomas},
title = {ViPATalk meets AI: Advancing Multimodal Virtual Anamnesis Simulation using LLM Fine-tuning},
journal = {Gesellschaft für Informatik (GI)},
year = {2025},
doi = {10.18420/delfi2025_35}
}
EUVIMED