CocktailAI: Discharge Summary Generator Using Template Document and Retrieval-Augmented Generation, and Hybrid-Cloud Hospital Information System Platform to Enable It

Abstract

The recent advancement of artificial intelligence (AI) technologies initiates health data science research. The hospital information system of university hospitals should enable easy access to such new technologies without harming data safety to accelerate the research. Kyoto University Hospital developed hybrid-cloud hospital information system to let researchers apply any cloud-enabled tools to clinical data. The platform successfully initiated AI research including a discharge summary generator. Emerging generative AI technology is expected to be a silver bullet to decrease documentation tasks of clinical staff. On the other hand, the straight-forward application of generative AI may cause several errors including the hallucination. We simply introduce retrieval-augmented generation to find the key information from inpatient records to finalized template document including the key information. The developed system, named CocktailAI, is applied for daily clinical activity to generate discharge summary for referral letters for a month. 56% of generated documents are published almost as it is and other 36% of generated documents are applied with a few additional sentences. In total, 92% of referral letter writing task was supported by the developed system.