[Creation of Patient Summaries from Electronic Medical Records and Integration of Medical Knowledge via Large...

In Japan, the adoption rate of electronic medical records (EMR) in general hospitals has surpassed 50%, leading to the rapid digitization of medical information. However, core components such as clinical notes and nursing records remain accumulated as unstructured data in "natural language" (free text). Consequently, their secondary utilization relies heavily on human interpretation, presenting significant challenges. This paper details the development and utility of a system that automatically generates high-precision patient summaries from EMR free text, leveraging large language models (LLMs) and retrieval-augmented generation (RAG) technologies, which have undergone dramatic evolution in recent years. We developed a medical-specific LLM trained on data accumulated at Tohoku University Hospital and conducted demonstration experiments on: 1) the automated generation of discharge summaries from nursing records, and 2) the extraction of complex cases based on clinical trial eligibility criteria. The results demonstrated that the LLM generated summaries of quality comparable to those created by nurses, suggesting the potential for significant improvements in the efficiency of documentation tasks. Furthermore, in the context of drug discovery support, the system successfully identified cases meeting complex clinical conditions that were impossible to extract using conventional search methods. However, the risk of hallucinations inherent in generative AI was also confirmed, making human oversight (human-in-the-loop) indispensable for clinical implementation.