EPFL launches the first open medical LLMs (actu.epfl.ch)

EPFL has launched MeditronFO, the first fully open framework for developing medical large language models (LLMs), enhancing transparency and accountability in AI healthcare applications. This initiative addresses the critical issue of proprietary medical AI systems, which often lack sufficient visibility regarding their training data, methodologies, and decision-making processes. MeditronFO allows researchers to “medicalize” existing open models, incorporating clinician input at every stage, from data curation to model validation. By leveraging publicly available datasets, along with clinician-reviewed synthetic data, the framework ensures that the development process remains grounded in practical medical contexts. The significance of MeditronFO lies in its competitive performance; models built with this framework outperformed their original versions, most notably the Apertus-70B-MeditronFO, which achieved a 6.6 percentage point improvement on medical exams. This launch is part of a broader initiative to conduct clinical trials across various countries, assessing the real-world impact of AI recommendations on patient care. The researchers emphasize that such transparency and active clinician involvement can create a more equitable and locally relevant AI ecosystem in healthcare, challenging the growing reliance on proprietary systems that may not align with specific community needs.