Israeli-led AI model reads glucose data to flag diabetes risk years ahead

A team led by Weizmann Institute computational biologist Eran Segal and Nvidia AI researcher Guy Lutsker, a doctoral student in Segal's lab, described in Nature on 14 January 2026 an artificial-intelligence model that turns continuous glucose-monitor readings into long-range health forecasts. The model, called GluFormer, uses the same transformer architecture that powers large language models, treating a stream of blood-sugar measurements much as a chatbot treats a stream of words.

GluFormer was trained on data from Israel's '10K' project, which has followed more than 14,000 volunteers with continuous glucose monitoring alongside genetic tests, blood and stool analyses, microbiome profiles, sleep studies and detailed medical and lifestyle histories. The collaboration included Nvidia's Israeli research arm, the Israeli health-data company Pheno.AI and partners abroad, among them the Mohamed bin Zayed University of Artificial Intelligence in Abu Dhabi.

With as little as one to two weeks of glucose data, the model could estimate the risk of developing type 2 diabetes up to 12 years later. According to the researchers, about two-thirds of participants who went on to develop diabetes had been flagged as high-risk years in advance, and nearly 70 percent of those who later died of cardiovascular causes had also been identified as high-risk. It outperformed standard glucose-based measures at predicting a range of outcomes, from heart disease and kidney and liver function to visceral fat and sleep disorders.

Segal noted that people labelled 'pre-diabetic' are usually treated as a single group even though existing tests cannot reliably tell who will progress to the disease. Pheno.AI holds the commercial rights and plans to work with health systems on clinical use. The researchers acknowledge that, like many AI systems, the model partly works as a black box.