Electronic Health Record Summarization via LLM-Constructed Knowledge Graphs

With the introduction of electronic health records (EHR) in the medical field, doctors and nurses can examine patients faster and more efficiently than paper records. Despite the advancement in patient documentation technology, one of the main drawbacks for EHRs is the inconsistent format of documents among the different medical specialties, specifically psychiatry, and behavioral health EHRs, as well as those used by a range of behavioral healthcare professionals, tending to be more anecdotal and text-based. With the recent advancement of large language models (LLM), this technology has considerable potential to become a viable solution, as medical professionals could use them to summarize and inquire about patients at record speeds. While LLMs have the potential to revolutionize the medical industry, their issues include their inconsistently formatted responses and their limited knowledge domain. Consequently, they are currently not applicable in high-stakes medical situations, as a single incorrect diagnosis could result in the patient’s injury. We propose using LLM-augmented knowledge graphs to aid in the LLM’s ability to perform QnA tasks and mitigate the possibility of data hallucination. Through prompt engineering, the LLM is able to generate formatted knowledge graphs based on a set of rules that focus on extracting as many relationships involving the patient, including afflictions and previous addictions. Using these graphs, we are provided with better visualizations of the patient’s current and prior issues and reduce the complexity of future inquiries regarding their health via knowledge graph queries.