Home | Publications | SSB+25a

LLM-Gestützte Extraktion Klinischer Daten: Potenziale Und Herausforderungen

MCML Authors

Peter Schüffler

Prof. Dr.

Associate

Computational Pathology

Abstract

Background: Large language models (LLM) can automatically process clinical free-text documents, extract key information, and thereby reduce reading effort and documentation-related workload. High-quality data and targeted model control are essential for practical applicability.<br>Material and methods: Various approaches to information extraction are presented. Additionally, 24 unstructured pathological reports of bone and soft tissue tumors are processed using the local, generic LLM Llama 4 Scout with three different prompt variants and compared in terms of extraction quality.<br>Results: Prompt design had a substantial impact on model behavior. Prompts with clear parameter definitions and examples achieved the most reliable results. Typical LLM-specific errors, such as hallucinations and misclassifications, were also observed.<br>Summary: LLM can support clinical staff by rapidly and systematically extracting relevant content from free-text documents. Safe and effective use requires high-quality data, precise inputs, and close collaboration between medical and technical experts.

article SSB+25a