Toward Automatic Summarization of Hospital Discharge Notes

Traditional methods to perform automatic summarization are rapidly being substituted by methods based on Large Language Models (LLMs). However, it cannot be guaranteed that summaries generated by LLMs contain only content from the source. The goal of my research is to discover new methods that combine language-based graphs and deep learning models to address provenance of content and trustworthiness in automatic summarization. This is particularly important for summarization in health care. Specifically, I focus on automatically generating discharge summaries, a lengthy medical narrative document written by physicians that summarize a hospital in-patient visit. Automatically generating these summaries would allow more time for physicians to be with patients and ameliorate physician documentation burden and burnout. My method uses language-based graphs to model the source data and an alignment algorithm to pair the source and the potential summary; I demonstrate it on both to the publicly available MIMIC-III corpus and clinical notes written by physicians at UI Health (accessed according to IRB protocol 2024-0109).