“FHIR-Starter”: research group wants to structure health data meaningfully
Health data is often not interoperable. Fraunhofer IESE and others want to change this and are developing a large language model for this purpose.
(Image: greenbutterfly/Shutterstock.com)
The data in electronic patient records is currently predominantly unstructured. To change this, the “FHIR-Starter” project is being launched under the leadership of the Fraunhofer Institute for Experimental Software Engineering IESE, which is collaborating with Prof. Sylvia Thun and her working group for health research at the Charité hospital in Berlin and the AI company Insider Technologies.
Software based on a large language model (LLM) and natural language processing will be developed to automatically structure the data so that it can be better used for both healthcare and research. The FHIR medical data standard and the LOINC and SNOMED-CT coding systems will be used.
With the resulting software, doctors will be able to “display laboratory values over time, for example, or automatically create medication lists. With the structured data, the ePA could be completely and meaningfully digitized,” explains Dr. Theresa Ahrens, Head of the Digital Health Engineering department at Fraunhofer IESE. The project also aims to make anonymized data available for research in the future.
The challenges cited by IESE include “ensuring the reliability of the data” to prevent AI hallucinations and “extensive data protection”, which is to be complied with using in-house servers and LLMs based on open source. According to Ahrens, security mechanisms are important for the responsible use of large language models in the healthcare sector. “Fraunhofer IESE has already done important groundwork for this with the development of the so-called Uncertainty Wrapper, which we can build on in this project,” explains Ahrens. The Uncertainty Wrapper is designed to quantify, manage and reduce uncertainties in AI models.
Structured data instead of a “digital scroll”
Even before the current version of the electronic patient record was implemented, there was criticism because the data in the electronic patient record was not structured, but predominantly in PDF format. Opening and examining the individual files was repeatedly described as time-consuming – in this context, concerns were also expressed about a “digital scroll”.
In addition, important information has to be typed out manually to transfer the data to the practice or hospital administration systems, as the IESE press release also points out. This is prone to errors and worsens the quality of care. Similarly, research“can only use data from full texts with difficulty, which weakens Germany as a research location”. “At last! Fax/ Word PDF to FHIR in Germany”, Thun is therefore pleased to report on LinkedIn.
In the future, the software service will offer open interfaces for data transfer. It is unclear whether the developed software will subsequently be available as open source. The three-year project is being funded by the German Federal Ministry for Economic Affairs and Energy with 1.64 million euros as part of the “Generative AI for SMEs” innovation competition.
(mack)