Structuring Medical Data with AI: Progress on the FHIR-Starter Project

Details

The 'FHIR-Starter' project, led by Fraunhofer IESE, aims to structure and standardize medical data. Currently, non-structured text data, such as physician reports and clinical notes in PDF format, are prevalent, creating significant challenges related to time and effort required for manual reading or input into EHR/hospital information systems. This project utilizes Large Language Models (LLMs) and Natural Language Processing (NLP) to automatically analyze these unstructured medical documents and convert them into standardized data formats. By leveraging FHIR—a key international standard for healthcare data exchange—along with coding systems like LOINC and SNOMED-CT, the system aims to facilitate seamless data exchange between different systems. The research team is addressing two major challenges: mitigating AI 'hallucinations' (the tendency of models to generate false information) and ensuring GDPR compliance when handling sensitive medical data. To enhance security, they are adopting open-source LLMs that run on the user's own servers, and have developed proprietary tools like the 'Uncertainty Wrapper' to quantify and manage model uncertainties. The software has diverse potential applications, including providing anonymized data for research or automatically displaying longitudinal records and generating medication lists in future electronic patient records (ePA). The project runs for three years starting February 2025, expected to significantly contribute to the digitalization of medical information systems.