Researchers at the Weizmann Institute of Science and Nvidia have developed an AI Model Predict Diabetes risk up to twelve years in advance by analyzing continuous glucose monitoring data, according to a study published this week in Nature.
The study highlights an AI Model Predict Diabetes system called GluFormer, which analyzes subtle patterns in blood sugar fluctuations to identify individuals most likely to develop the disease long before standard clinical symptoms appear. The research is led by Prof. Eran Segal, a computational biologist at the Weizmann Institute, with support from Nvidia’s AI infrastructure and researchers.
The findings arrive as health systems worldwide invest heavily in diabetes prevention, particularly among people labeled as pre-diabetic, a group that often receives broad lifestyle advice but uneven follow-up care.
Study Finds Glucose Data Can Flag Diabetes Risk Up to 12 Years Early
The AI Model Predict Diabetes, known as GluFormer, was trained using data from the “10k Project,” a long‑running health study that initially aimed to follow ten thousand participants but now includes about fourteen thousand people. Participants contribute continuous glucose measurements along with genetic data, blood tests, microbiome samples, sleep data, movement tracking, and detailed health questionnaires.
“This is one of the richest datasets in the world linking continuous glucose measurements to long-term health outcomes,” Segal said in an interview. He added that his team were “world pioneers in measuring sugar in healthy and pre-diabetic people over time.”
The researchers tested the model using data from nine independent databases, focusing on people classified as pre-diabetic based on glycated hemoglobin, or A1C, the current standard diagnostic measure.
Algorithm Outperforms Standard A1C Measure in Identifying High-Risk Patients
The study revealed that A1C levels within the pre‑diabetic range were weak indicators of future diabetes risk. In contrast, the AI Model Predict Diabetes demonstrated strong separation between high‑ and low‑risk individuals.
According to the paper, sixty-six percent of participants who later developed diabetes received a risk score of seventy-five or higher on the model’s one-hundred-point scale. Only seven percent of those who developed diabetes scored twenty-five or lower.
“It seems intuitive to think that higher A1C within the pre-diabetic range means higher risk, but that turns out not to be true,” Segal said. “Our algorithm can predict that risk far more accurately.”
The model also showed predictive power beyond diabetes. It identified future cardiac events with even greater accuracy, the researchers reported.
Nvidia-Backed Model May Help Target Preventive Care Resources
Sixty-nine percent of participants who later experienced a heart attack had high risk scores, while none of those with low scores suffered a heart attack during the study period. The researchers say this suggests continuous glucose patterns reflect broader metabolic health.
Guy Lutsker, the study’s lead AI researcher at Nvidia and a doctoral student in Segal’s lab, explained that the AI Model Predict Diabetes works in a way similar to large language models. “From just a week of blood sugar readings, it predicts the probability of future developments,” he said. “It has likely learned something fundamental about diabetes, even if we cannot fully explain it in simple terms.”
Lutsker acknowledged the system functions partly as a “black box,” meaning its internal reasoning is too complex for full human interpretation.
The research was supported by the School of Digital Public Health at the Mohamed Bin Zayed University of Artificial Intelligence in Abu Dhabi. Pheno.AI has acquired commercialization rights and plans to work with health organizations to deploy the technology.
Segal said the model could help focus resources on the twenty to thirty percent of pre-diabetic patients most likely to progress to disease. “The advice may be similar for everyone,” he said, “but the intensity of intervention should not be.”
