Leveraging the recent competence of Large Language Models in language generation, we developed an innovative method for summarizing clinical reports, specifically targeted at ICU patient reports at the University Hospital of Zurich (USZ). To enhance the utility of the summaries, we augmented the reports with relevant facts extracted from SNOMED CT, ensuring their interoperability in line with the ongoing harmonization efforts of SPHN. We established several best practices and fine-tuned various LLMs for this purpose. The qualitative results from experiments conducted on available benchmarks demonstrated the proposed methods' significant impact on the summaries' quality. Additionally, we provided a user interface enabling medical professionals to qualitatively evaluate the generated summaries for each LLM.

Building on this success, we secured a larger grant from DIZH (volume 600K CHF) to extend this idea toward developing a more comprehensive medical information retrieval system.

Clinical physicians spend about 40% of their work time for reading and writing patient documentation. We will employ Natural Language Processing (NLP), SNOMED CT, and Large Language Models (LLM) to generate concise, accurate, and interoperable summaries of patients’ records, thus saving time, effort, and resources.