Towards Artificial General Intelligence (AGI) for healthcare

Community building: build a large-scale and diverse pan-European community of stakeholders with the multidisciplinary expertise united as required to develop the next generation of frontier AI models for healthcare, towards AGI for healthcare, with a view to leveraging as a community the potential of AI Factories. Where relevant, this should build on and strengthen existing EU-funded communities and networks, and could pave the way for a formalised long-term collaboration under one of the available EU instruments.

Roadmap creation: review previous research to identify the most promising AI models and model development approaches. In addition, risk assess and review evidence on safety and efficacy of existing AI models with reference to the AI Act, related regulatory provisions (including any jurisprudence) and ethical and security considerations, so that frontier AI model development can proceed on a well-informed basis. Finally, create a roadmap for developing the next generation of frontier AI models.

Dataset identification, curation, expansion and use: i) identification: identify the most suitable existing datasets for the development of frontier AI models for healthcare, ii) curation: identify how to validate the datasets, ensure dataset interoperability, and convert datasets into formats suitable for frontier AI model development, iii) expansion: identify additional datasets and/or annotations required for frontier AI model development, especially to ensure that datasets are representative and iv) use: identify methods and required infrastructure to allow privacy-preserving use and further expansion of the datasets in alignment with and through the EHDS.

Frontier AI model preparatory activities: mapping approaches for training and evaluating frontier AI models (e.g. approaches to combine foundation models for life sciences and healthcare delivery in order to develop more advanced and multidisciplinary models towards personalised medicine). The approaches should cover all trustworthy AI aspects.