This topic will contribute to advancing research and providing new evidence on how these models contribute to and support biomedical research and its applicability towards more predictive and personalised medicine, while also defining use conditions, usability requirements and training needs of the researchers. It aims to cover existing gaps related to Generative AI in biomedical research, addressing both capabilities and existing limitations.

Research actions under this topic should include all the following activities, ensuring multidisciplinary approaches and a broad representation of stakeholders in the consortia (e.g. industry, academia, healthcare professionals):

• Develop new or re-purpose existing Generative AI models for biomedical research across various medical fields and/or therapeutic indications. The models should be robust, based on the use of large-scale, complex, and multimodal high-quality data (real and/or synthetic data), such as but not limited to medical imaging, genomics, proteomics, other molecular data, electronic health records, laboratory results, unstructured health data and/or available scientific and public information relevant to biomedical research. The applicants may choose any type of available large-scale biomedical data and/or their combinations and justify their relevance for training and optimisation of the Generative AI tools.
• Develop a proof of concept with at least two use cases relevant for predictive and personalised medicine in different medical fields to demonstrate the scientific added value compared to currently used methods and/or potential future clinical utility of the Generative AI models in biomedical research. The applicants should actively engage potential end users in the development, adaptation and testing of the new/repurposed models, considering sustainability aspects.
• Develop or revise existing methodologies to assess alignment with human values and the use cases of developed and/or repurposed Generative AI models, their applicability, performance, limitations and added value in biomedical research. These methodologies should demonstrate the technical, scientific, and potential future clinical utility, robustness and trustworthiness of the developed or repurposed Generative AI models, in particular:
  • Appropriate performance metrics for continuous evaluation and testing of scientific, technical robustness and relevance of the Generative AI models, as well as risks from misalignment of training data (which may degrade performance, e.g. through including but not limited to hallucinations or confabulations of these models).
  • Appropriate metrics for model intelligibility, robustness, alignment with ethical principles and approaches for ethical evaluation of AI trustworthiness.
  • Appropriate solutions to identify and mitigate potential bias and confounding of Generative AI models and include examples from different perspectives (e.g., representativeness of the data, bias of the trainer, bias of training and validation data, algorithmic discrimination and bias including gender bias etc.).
  • Methods to systematically address and assess ELSI (Ethical, Legal, and Societal Implications) aspects, including data privacy, risk of discrimination/bias (not limited to sex, gender, age, disability, race or ethnicity, religion, belief, minority and/or vulnerable groups).
  • Appropriate techniques to ensure explainability of the model in order to increase users’ trust.
  • Approaches and metrics (where feasible) for the usability of Generative AI models for researchers.

All proposals should demonstrate EU added value by developing and/or using trustworthy and ethical Generative AI models developed in the EU and Associated countries, involving in the consortium EU industrial developers of Generative AI solutions, including leading-edge startups when possible. An open-source approach is encouraged when technically and economically feasible.