Weizmann Institute, Nvidia develop diabetes prediction model

An article published yesterday in leading periodical “Nature” presents a new model that can predict the risk of diabetes 12 years in advance, based on data collected from continuous glucose monitoring. Prof. Eran Segal, a computational biologist from the Weizmann Institute is the senior researcher who conducted the study.

Segal is best known for his research that formed the basis of the company DayTwo, which predicts the metabolic response of different people to the consumption of different foods. Although it was not commercially successful, Segal’s laboratory continues research in these and related fields and has received widespread global recognition.

Surprising insight from the algorithm

The current study by Segal and his team relies on GluFormer, an AI model trained on data collected in the “10k Project,” which was initially designed to map a series of data from 10,000 people for the purpose of personalized disease prediction, and has exceeded its goal, with 14,000 participants.

Every two years the participants fill out questionnaires, undergo genetic testing, provide blood test results, stool tests to diagnose the microbiome, home sleep tests (manufactured by Israeli company Itamar Medical, now owned by Medtronic), movement tests, details of medical history and medical habits, and also undergo continuous sugar testing. This is one of the richest databases in the world that links continuous sugar measurements to a wide range of variables measured over time.

Segal tells “Globes,” “I think we have been world pioneers in measuring sugar in healthy and pre-diabetic people.”

Based on this data, researchers are working to isolate predictive variables, so that it will be possible to identify healthy people or those with risk factors who need preventive treatment. In the diabetes study that is now being published, the researchers tested the engine on new data from nine databases. The data were from subjects who were defined as pre-diabetic based on the currently accepted index – the A1C index (glycated hemoglobin).

Segal says, “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. Although, on average, those who were in the range defined as pre-diabetic developed diabetes more often, the position within the range hardly predicted the risk, while our algorithm can predict it.”

According to the article, 66% of those who did eventually develop diabetes received a “score” of 75 (out of 100) or higher on Segal’s index. In contrast, only 7% of patients received a score of 25 or lower, evidence of the index’s ability to separate the groups.

Nvidia to collaborate with Sheba on genome hidden codes

Predicts more than diabetes

Although the measurement was based on fluctuations in sugar, it predicted cardiac events even more accurately than diabetes. 69% of those who suffered a heart attack eventually received a score of 75 or higher on the index. In contrast, among those who received a score of 25 or lower on the risk index, not a single heart attack occurred. “If you add additional data to the sugar measurement, you can achieve even better results,” says Segal. The research team also hopes to show how the sugar data can be linked to additional, not necessarily predictable, outcomes, such as future sleep patterns.

According to Segal, this data can help health organizations identify who in the pre-diabetic group really needs significant intervention, and who is unlikely to develop diabetes, even if they do not change their lifestyle. Preventing diabetes in pre-diabetics is a task in which health organizations, especially in the US, are investing a lot of effort, with measures ranging from intensive support for lifestyle changes to drug treatment. The research was also supported by the School of Digital Public Health at MBZUAI (The Mohamed Bin Zayed University of Artificial Intelligence in Abu Dhabi) and Pheno.AI, which acquired the rights to commercialize the technology and will work to bring it to health organizations.

Support from Nvidia

The model developed by the researchers works in a similar way to large language models and is based on Nvidia’s infrastructure and supported by AI experts from within the company.

Just as textual AI tools predict the most likely next word and the one after that, Segal and his team’s model predicts from a week’s worth of blood sugar readings the probability of future developments. “And just as we understand that textual AI has learned something fundamental about language, our model has probably learned something fundamental about diabetes,” says lead researcher Guy Lutsker, an AI researcher at Nvidia and a doctoral student in Segal’s lab at the Weizmann Institute.

Did it learn something that you can put into words? Did you discover a new scientific phenomenon that can be conceptualized?

Lutsker: “No, the model is to some extent a black box. Yes, we were able to extract certain insights from it, but the connections are probably too complex for humans to fully grasp.”

A problem that DayTwo also had and that you will have to overcome is that while your predictions can be accurate and personalized, ultimately the recommendations for everyone are pretty similar, regardless of their exact sugar level – exercise, proper nutrition, and so on.

Segal says, “What makes this work interesting is that only 20%-30% of the group that is defined as pre-diabetic will actually develop diabetes. So maybe the lifestyle recommendations are similar, but the intervention resources need to be concentrated on this group.”

Published by Globes, Israel business news – en.globes.co.il – on January 15, 2026.

Source link