The ELMTEX project presents an innovative solution for using AI and Large Language Models (LLMs) to process clinical documentation. Cost-efficient, data protection-compliant approaches enable clinics to operate AI applications on site while meeting the requirements of the European Health Data Space.

The integration of artificial intelligence (AI) into the healthcare sector opens new possibilities, particularly in the processing of clinical documentation. The ELMTEX project, carried out by the Fraunhofer Institute for Applied Information Technology FIT, is dedicated to the optimization of Large Language Models (LLMs) for applications in the German healthcare sector.

"Our aim is to provide a cost-effective and data protection-compliant solution that enables clinics to run AI applications on site without having to rely on expensive commercial services," says Dr. Carlos Velasco, ELMTEX project manager at Fraunhofer FIT.

Challenges and approaches
Clinical texts place special demands on AI models: they contain complex syntactic structures, numerous abbreviations and temporal relationships between symptoms and diagnoses. The project examined three modeling approaches:

• Naive prompting: Simple queries to obtain information
• Retrieval-Augmented In-Context Learning: Using similar examples to improve results
• LoRA Fine-Tuning: fine-tuning of smaller models with domain-specific data

Metrics such as ROUGE (text similarity), BERTScore (semantic similarity) and entity-level metrics (clinical accuracy) were used for evaluation. The results show that smaller, fine-tuned models perform better than larger models – a key advantage for resource-constrained environments.

Innovative data sets
A central component of the project is a newly developed, annotated data set with 60,000 English and 24,000 German clinical reports. This dataset covers categories such as patient history, diagnoses and treatment measures and has been checked by manual validation and automated procedures. The data enables precise adaptation of the models to the specific requirements of the healthcare sector.

Focus on data protection and interoperability
A key advantage of the developed solution is the possibility of local implementation in clinics. This protects sensitive patient data and at the same time meets the requirements of the European Health Data Space (EHDS). The structured information generated can be seamlessly integrated into existing hospital information systems and supports compliance with the EU AI-Act in terms of transparency and traceability.

Future perspectives
The ELMTEX project is currently being evaluated with clinical teams to further improve its practical applicability. Work is also underway to extend the approaches to other sectors such as manufacturing or finance. Greater synchronization with standardized medical terminology and the expansion of multilingualism are also the focus of future developments.

With its innovative approaches, the ELMTEX project impressively demonstrates how AI-based solutions can not only increase efficiency in the healthcare sector but also meet the challenges of data protection and interoperability – a decisive step towards digitalized healthcare in Europe.

Further information: https://s.fhg.de/elmtex-en