Israeli scientists’ AI model can predict diabetes many years before symptoms appear

Researchers presented the GluFormer neural network, capable of predicting the occurrence of diabetes and cardiovascular pathologies. The development was based on the transformer architecture, which is usually used in large language models (for example, ChatGPT). Instead of words, the system was trained on an array of ten million blood sugar measurements taken from tens of thousands of people. Thanks to this approach, AI has learned to recognize hidden patterns in glucose fluctuations that go undetected in standard clinical tests.

The effectiveness of the model was confirmed during long-term tests. Analyzing twelve years of data, GluFormer successfully identified two-thirds of patients who later developed diabetes and also predicted deaths from cardiovascular complications. The system showed high consistency of results across different populations and across different medical devices used for testing. In addition to diabetes, the neural network is able to assess the risks of kidney disease, liver disease and sleep disorders, based on indicators of metabolic health.

The authors of the work emphasize that such early detection of threats can radically change preventive medicine. This will allow doctors to intervene in the situation before irreversible changes occur in the body. The use of such technologies will not only save millions of lives, but also significantly reduce the economic burden on healthcare systems around the world.

Co-author of the model, Professor Gal Chechik, notes the significance of the work done: “The success of GluFormer in predicting diabetes and disease risks demonstrates the significant potential for integrating artificial intelligence into medical research.” According to him, this opens the way to a future where AI can extract clinically relevant data from observation sets on a scale that was previously unattainable by humans.