Designing LLM-powered agents to manage clinical text reports in medical institutions

Large language models (LLMs) based on Transformers have revolutionized the field of artificial intelligence, allowing the analysis of large corpora of text in a way that was previously unconceivable. These models revealed unprecedented capabilities in tackling problems for which they were not specifically trained, performing zero-shot and few-shot predictions in a variety of contexts. In the medical domain, large amounts of data are stored in Electronic Health Records (EHRs) in the form of unstructured text. Identifying relevant information in EHRs is not straightforward and usually involves a time-consuming manual inspection by clinicians. In this context, LLMs could be employed to build intelligent systems to extract meaningful data from EHRs automatically. Moreover, open-source models can be integrated locally so that no privacy concerns arise. In this work, we present an agent model powered by LLMs allowing clinicians to perform powerful manipulations of EHRs, like Retrieval Augmented Generation (RAG), summarization, and structured data extraction. The chatbot design is especially convenient since it makes the interaction with the model straightforward and requires no technical knowledge by the clinical staff. We implement the system in the Neurological Sciences Institute of Bologna, with the goal of integrating it in the everyday clinical practice and leading the integration of LLMs in the Italian healthcare landscape.