MedGellan is a lightweight, annotation-free framework for generating clinical guidelines from medical records. It uses a large-scale language model (LLM) to generate clinical guidelines for diagnosis prediction from medical records, which doctors then use to make their own diagnoses. It uses a Bayesian-inspired prompting strategy to respect the temporal order of clinical data. Initial experiments have shown that LLM-generated guidelines using MedGellan improve diagnostic performance, particularly in terms of recall and F1 score. Given that errors in medical decision-making can have serious consequences, and that full automation is challenging, we present a hybrid framework that combines human supervision and machine intelligence.
