A Synthetic Multi-modal Dataset for Type 2 Diabetes

Abstract

Abstract Type 2 diabetes mellitus (T2DM) is a widespread chronic disease that impacts multiple organ systems. Artificial intelligence (AI) has shown promise in advancing type 2 diabetes mellitus (T2DM) research by enabling the integrated analysis of complex, multimodal data. However, progress is constrained by limited access to such multimodal datasets as contractual restrictions and concerns around patient privacy often limit the dissemination of existing resources. Here, we present the development and evaluation of a publicly accessible synthetic version of AI-READI, a large multimodal dataset for T2DM research. This synthetic dataset, constructed through generative AI, contains 93,856 retinal fundus images, 56,512 optical coherence tomography slices, and 10,518 tabular data records that combine clinical diagnoses, laboratory measurements, and data from wearable sensors. Our evaluation illustrates that the synthetic data captures many of the relationships within each modality of the original AI-READI dataset while being strongly resilient against membership inference attacks aimed at compromising participant privacy. This synthetic dataset is designed to support education, innovation, and reproducibility in T2DM research and is released under a permissive license that supports broader reuse than the original AI-READI dataset.