From Rehab to Recap: Generating Discharge Reports for Rehabilitation Clinics with LLMs

Healthcare professionals spend a large part of their working time on documentation. Discharge reports from rehabilitation clinics require an extensive amount of time to write, review and finalize by a team of clinicians and administrative staff. Natural Language Generating (NLG) can be used to speed up this process, while maintaining a high content quality. Due to a pre-specified structure of discharge reports, high variability of content and heterogeneous information types, a custom approach is necessary. Our main contribution is a hybrid framework that combines a locally deployed Large Language Model (LLM) and a context-specific prompting approach with conventional control elements inside a web interface accessed by practitioners and administrative staff, tied together in an assistant application. We compare multiple LLM architectures and prompting aspects regarding linguistic styles in detail. Based on a user study, we showcase how our approach achieves a high satisfaction by healthcare professionals.