Researchers Release Open Diabetes Datasets to Accelerate AI Development

Researchers have introduced Glucose-ML, a publicly accessible collection of 10 longitudinal diabetes datasets designed to address a longstanding bottleneck in digital health technology: the scarcity of high-quality, shareable data for training robust AI models.

The Collection#

The Glucose-ML assemblage spans more than 300,000 days of continuous glucose monitor (CGM) readings, encompassing 38 million glucose measurements gathered from over 2,500 participants across four countries. The participating population reflects the full spectrum of glucose metabolism disorders—including individuals with type 1 diabetes, type 2 diabetes, prediabetes, and people without diabetes. These datasets were released between 2018 and 2025, ensuring recency and relevance to current research.

Addressing a Critical Gap#

AI algorithms have become integral to modern diabetes management tools, yet researchers have long faced obstacles in accessing datasets large and clean enough to build reliable, generalizable models. By releasing these datasets openly alongside comparative analysis and code, Glucose-ML aims to enable the creation of transparent, reproducible algorithms that can withstand real-world variation.

Benchmark Findings#

The team conducted a case study focused on short-term blood glucose prediction—a foundational task in diabetes AI—testing the same algorithm across all 10 datasets. Results revealed a striking insight: identical algorithms produced substantially different performance metrics when trained and evaluated on different data sources. This finding underscores how dataset selection profoundly shapes model outcomes and highlights the need for systematic guidance when choosing training data.

Implications for Researchers#

The study yields practical recommendations for developing robust AI solutions not only in diabetes management but across the broader health technology landscape. Researchers now have direct links to each dataset and access to the team's open-source code, positioning Glucose-ML as a resource for standardizing and improving AI-driven healthcare innovation.