[Keeping track of things: large language models for patient synopses : Source-bound system for clinical information systems].
Authors
Affiliations (3)
Affiliations (3)
- Klinik für Diagnostische und Interventionelle Radiologie, Universitätsklinikum Freiburg, Freiburg, Deutschland. [email protected].
- Department of Radiology, Medical Center, University of Freiburg, Hugstetterstr. 55, 79106, Freiburg, Deutschland. [email protected].
- Klinik für Diagnostische und Interventionelle Radiologie, Universitätsklinikum Freiburg, Freiburg, Deutschland.
Abstract
The increasing documentation burden in electronic health records makes it difficult to obtain a rapid overview of relevant prior information in complex disease courses. Especially in highly digitized settings such as radiology and interdisciplinary case conferences, large volumes of heterogeneous documents must be reviewed under time pressure. The aim was to develop a large language model (LLM)-based patient synopsis to accelerate information retrieval while ensuring transparency, physician oversight, and data protection. At the University Medical Center Freiburg, a system was developed for the automated integration of clinical documents from multiple primary systems, including the electronic health record (EHR), picture archiving and communication system (PACS), and other subsystems. It uses retrieval-augmented generation (RAG), metadata harmonization, vector search, and agentic retrieval for the context-sensitive selection of relevant content. The system operates with role-based access control, end-to-end source attribution, and processing in a sovereign cloud without persistent data storage. The system enables patient-specific queries, structured longitudinal summaries, and automated, source-grounded summaries to support preparation for interdisciplinary case discussions. All statements remain traceable to primary documents, while medical interpretation and decision-making remain the responsibility of the treating physician. Interoperable information systems and a data protection framework in accordance with the General Data Protection Regulation (GDPR), the German Social Code Book V (SGB V), and the EU Artificial Intelligence Act are prerequisites for implementation. LLM-based patient synopses can support text-intensive clinical workflows, provided that controlled retrieval, clear source attribution, physician validation, and an interoperable, regulation-compliant system architecture are in place. Prospective evaluations are required before routine clinical use.