ADVANCE Center Synthetic AI-READI Dataset for T2DM Research

Synthetic AI-READI is a publicly accessible, multimodal synthetic dataset designed to accelerate research, education, and innovation in artificial intelligence and machine learning for Type 2 Diabetes Mellitus (T2DM). Developed from the NIH Bridge2AI AI-READI program (https://fairhub.io/datasets/1 ), the dataset was generated using generative AI methods oriented to preserve important statistical and clinical characteristics of the original AI-READI cohort.

• 93,856 synthetic retinal fundus photographs - high-resolution images preserving clinically relevant features
• 56,512 synthetic OCT image slices - optical coherence tomography retinal images for structural analysis
• 10,518 synthetic tabular records - participant-day observations combining clinical measurements and wearable-device data

The multimodal linkage across retinal imaging and wearable-device data within a single synthetic cohort is uncommon among synthetic medical datasets, enabling researchers to develop and test cross-modal AI architectures without assembling disparate data sources.

• Model Development and Testing: Train and validate AI/ML models on realistic synthetic data before accessing real patient cohorts, reducing time-to-prototype significantly
• Benchmarking: Evaluate model performance across imaging and tabular modalities using a standardized, reproducible dataset
• Education and Hackathons: Provide students and participants with privacy-safe biomedical data that requires no IRB approval for classroom or competition use
• Methodological Research: Investigate generative modeling techniques, synthetic data quality metrics, and multimodal fusion approaches

After subscribing through AWS Marketplace and obtaining your access keys, the dataset can be loaded into Amazon SageMaker for model training and experimentation. Tabular data can be queried and analyzed using Amazon Athena, and image data can be stored and accessed via Amazon S3. The dataset formats are compatible with standard ML frameworks including PyTorch and TensorFlow running on AWS infrastructure.

The dataset originates from the federally funded NIH Bridge2AI AI-READI cohort, a well-characterized source population. The generative AI methods used to produce the synthetic data are designed to preserve statistical relationships while mitigating re-identification risk. The synthetic nature of the data means no real patient information is directly contained in the released files, enabling broader collaboration without the access restrictions typically associated with clinical datasets.

Further details on generation methodology and validation metrics are available in: Jackson N, Espinosa Dice N, Yan C, Li Z, Jiang X, Lee A, Malin B. "A synthetic multi-modal dataset for type 2 diabetes." Scientific Reports. 2026; In press.

This synthetic dataset is intended to supplement - not replace - the original AI-READI dataset. While it preserves many important statistical relationships found in the source data, it should not be used for clinical decision-making or as a substitute for access to real-world clinical data.

To access the dataset, request your keys at https://hiplab.vumc.org/synthetix  (select AI-READI Dataset). After receiving your keys, follow the onboarding instructions provided to begin loading data into your AWS environment.