AI model from Nvidia’s Israel research center and Weizmann Institute predicts diabetes years ahead
Researchers say GluFormer outperforms standard clinical tools using glucose data alone.
A new artificial intelligence model developed by researchers in Israel and at Nvidia is showing an unprecedented ability to predict diabetes and other life-threatening diseases more than a decade before they are typically diagnosed, according to a peer-reviewed study published on Wednesday in the scientific journal Nature.
The model, known as GluFormer, analyzes long-term blood sugar patterns collected through continuous glucose monitoring systems and uses them to forecast future disease risk. In trials, it outperformed existing clinical tools, including the widely used HbA1c blood test, in predicting diabetes and cardiovascular disease as far as 12 years in advance.
The research was carried out by a collaboration between Nvidia’s Artificial Intelligence Research Center in Israel, the Weizmann Institute of Science, the Israeli startup Pheno.AI, and additional academic and clinical partners. Its publication in Nature places the work among a small group of AI-driven medical studies that have passed the journal’s stringent peer-review process.
At its core, GluFormer is a foundation model built on the Transformer architecture that underlies large language models such as GPT and Gemini. Rather than text, however, it was trained on more than 10 million glucose measurements collected from continuous glucose monitors, drawn from data on 10,812 individuals, most of whom were not diabetic at the time.
When researchers applied the model to glucose data collected 12 years earlier from a cohort of 580 adults, GluFormer correctly identified 66% of individuals who later developed diabetes and 69% of cardiovascular-related deaths among those classified as highest risk. The model’s performance remained strong across 19 external databases, representing different populations, devices, and medical conditions.
Beyond diabetes, the researchers found that the system was able to anticipate a broad range of health outcomes linked to metabolic health, including indicators associated with cardiovascular disease, kidney and liver function, blood lipid levels, visceral fat, and sleep disorders. In each case, the model consistently outperformed other prediction methods based on glucose monitoring alone.
Researchers involved in the study argue that the ability to identify disease risk many years earlier could reshape how preventive care is delivered. Earlier risk detection could allow physicians to intervene before irreversible damage occurs, help tailor treatment strategies in clinical trials, and reduce the long-term economic burden of chronic disease. The global cost of diabetes alone is projected to reach $2.5 trillion by 2030, according to estimates cited in the study.
“GluFormer’s success in predicting diabetes and disease risk further demonstrates the significant potential of integrating artificial intelligence into medical research,” said Prof. Gal Chechik, Senior Director of AI Research at Nvidia. He said the work points toward a future in which AI systems can extract clinical insight from patient data at a scale previously unattainable, supporting earlier detection and more informed medical decisions.
The backdrop to the research is a growing global diabetes crisis. About 10% of the world’s population currently lives with diabetes, and researchers estimate that by 2050 the number of people affected could exceed 1.3 billion. The disease is already among the leading causes of death worldwide and is associated with severe complications, including kidney failure, vision loss, and heart disease.
The development of GluFormer was led by Prof. Eran Segal of the Weizmann Institute and MBZUAI University, alongside Chechik, with contributions from researchers at Pheno.AI, clinicians from Schneider Children’s Medical Center, and academic teams from Tel Aviv University and Bar-Ilan University. Training and testing were conducted using Nvidia’s artificial intelligence infrastructure.
