Researchers have unveiled a new artificial intelligence model that can predict diabetes and cardiovascular mortality up to 12 years in advance, outperforming current clinical standards by analysing patterns in continuous glucose monitoring data.
The study, published Wednesday in Nature, introduces GluFormer, a generative foundation model built on the same transformer architecture that powers large language models like ChatGPT. The research represents a collaboration between the Weizmann Institute of Science, NVIDIA, Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), and Israeli startup Pheno.AI.
Breakthrough Risk Stratification
GluFormer was trained on over 10 million glucose measurements from 10,812 participants, primarily without diabetes, within the Human Phenotype Project cohort. In a 12-year follow-up study of 580 adults, the model demonstrated remarkable predictive accuracy: 66% of individuals who eventually developed diabetes were identified in the model’s highest-risk quartile, compared to just 7% in the lowest-risk quartile.
For cardiovascular mortality, the results were even more striking. Sixty-nine per cent of cardiovascular-related deaths occurred among those the model classified as highest risk, while no deaths occurred in the lowest-risk group.
“It makes sense to think that within the category of those defined as pre-diabetic, those who have a relatively high A1C would be at increased risk of developing diabetes, and those who have a lower A1C are at reduced risk, but it turns out that this is not true,” said Prof. Eran Segal of the Weizmann Institute and MBZUAI, who led the study. “Our algorithm can predict it.”
From Glucose Patterns to Health Insights
Unlike traditional metrics such as HbA1c blood tests, GluFormer analyses the temporal dynamics of glucose fluctuations rather than static measurements. The model was validated across 19 external cohorts spanning five countries, eight different CGM devices, and diverse conditions including prediabetes, type 1 and type 2 diabetes, gestational diabetes, and obesity.
“Just as we understand that textual AI has learned something fundamental about language, our model has probably learned something fundamental about diabetes,” said Guy Lutsker, lead researcher and AI scientist at Nvidia, who is completing his doctorate in Segal’s laboratory.
The model also predicts outcomes beyond diabetes, including indicators associated with liver and kidney function, blood lipid levels, visceral fat, and sleep disorders.
Toward Precision Metabolic Medicine
Researchers have extended GluFormer to integrate dietary data, allowing it to generate plausible glucose trajectories and predict individual glycemic responses to food, a step toward precision nutrition.
Prof. Gal Chechik, Senior Director of AI Research at Nvidia, said the work “points toward a future in which AI systems can extract clinical insight from patient data at a scale previously unattainable.”
With approximately 10% of the global population living with diabetes and projections suggesting this could exceed 1.3 billion people by 2050, early risk detection tools could reshape preventive care. The global cost of diabetes is projected to reach $2.5 trillion by 2030.
Pheno.AI has acquired the rights to commercialise the technology and will work to bring it to health organisations.
