Medical Model Synthesis Architectures: A Case Study

Medicine is rife with uncertainty. Human clinicians regularly need to navigate this uncertainty in contexts of the utmost stakes. AI systems that clinicians may consult, or even rely on, for differential diagnosis ought too to be able to reason under such uncertainty -- and be transparent in how such reasoning is done. Here, we put forward a framework for an AI system that can make practically useful and verifiable clinical predictions under uncertainty. Our framework combines advances in language models with formal probabilistic algorithms to support on-the-fly probabilistic medical model synthesis and reasoning. We present an initial proof-of-concept instantiation of our framework (MedMSA) and explore its potential for differential diagnosis over a series of clinical vignettes.