Co-designing Ethical Multimodal AI Systems for Mapping T1D Progression

Abstract Type 1 Diabetes (T1D) represents a significant global health challenge, characterized by complex interactions between genetic, environmental, and phenotypic factors. The NIDDK has funded numerous T1D initiatives over the years, generating vast scientific data across multiple modalities. However, there is still no system to fully leverage these large, multimodal datasets to build an AI model to facilitate data-driven knowledge discovery. Advances in AI and multimodal data analysis now offer opportunities to revolutionize our understanding of complex diseases like T1D. In this proposal, we will build MAI-DK (Multimodal AI - from Data to Knowledge), a system that converts multimodal data into a knowledge-generating machine. This system will provide a systematic overview of the molecular mechanisms controlling T1D progression across multiple organs. Achieving this goal will require an interdisciplinary team with expertise in genetics, genomics, pancreas and islet biology, biostatistics, machine learning, AI systems, and ethics. We have assembled a ten-MPI team with established, productive collaborations in these areas. The MAI-DK system will serve the diabetes research community, focusing on T1D pathogenesis and progression, and cater to clinical scientists, basic scientists, data scientists, and developers. To accomplish this, we propose five components: data modalities and QC, multimodal AI models and system, ethics, co-design, and stakeholder engagement. Completing these components will make MAI-DK an attractive destination for T1D research communities, accelerating progress in understanding T1D and informing transformative changes in diabetes prevention and care.