SANTA CLARA — Researchers from the Weizmann Institute of Science, Tel Aviv-based Pheno.AI, and NVIDIA have collaborated to develop GluFormer, a cutting-edge AI model designed to predict an individual’s glucose levels and other health metrics. This model uses past continuous glucose monitoring data to forecast future health outcomes, offering significant potential for both patients and clinicians.
GluFormer leverages AI technology to enhance the use of glucose monitoring data, potentially accelerating the diagnosis of conditions like prediabetes and diabetes. According to experts from Harvard Health Publishing and NYU Langone Health, this predictive capability could aid in identifying health anomalies, forecasting clinical trial results, and anticipating patient outcomes for up to four years.
The model goes a step further by incorporating dietary intake data, enabling it to predict how glucose levels will respond to specific foods and dietary changes, thus paving the way for precision nutrition. For individuals at high risk of diabetes, accurate glucose predictions could lead to earlier preventive care, improving patient outcomes while mitigating the economic impact of diabetes, which is expected to reach $2.5 trillion globally by 2030.
Given that diabetes currently affects around 10% of the world’s adult population, with projections suggesting this figure could double by 2050, AI models like GluFormer hold immense promise. Diabetes is a leading cause of death globally, contributing to complications such as kidney damage, vision loss, and heart disease.
Built on transformer architecture, the same neural network framework used by OpenAI’s GPT models, GluFormer is tailored to predict glucose levels instead of text. Gal Chechik, senior director of AI research at NVIDIA, explained that the transformer model is ideal for tracking biological processes and predicting future medical tests based on previous results.
The AI model was trained on 14 days of continuous glucose monitoring data from over 10,000 non-diabetic participants, collected every 15 minutes via wearable devices. This data was part of the Human Phenotype Project, led by Pheno.AI, which focuses on leveraging large-scale health data to improve human health.
As the lead author of the study, Guy Lutsker, an NVIDIA researcher and Ph.D. student at the Weizmann Institute of Science, noted that advancements in generative AI and the availability of extensive health data played a critical role in the success of the research. “These factors put us in a unique position to extract valuable medical insights,” he said.
The team validated GluFormer across 15 datasets and found it to be effective in predicting health outcomes for a variety of groups, including those with prediabetes, type 1 and type 2 diabetes, gestational diabetes, and obesity. The model’s training and inference were accelerated using a cluster of NVIDIA Tensor Core GPUs.
Beyond glucose levels, GluFormer can predict additional health metrics, including visceral adipose tissue, systolic blood pressure, and the apnea-hypopnea index, all of which are linked to diabetes and other health conditions.
Read the GluFormer research paper on Arxiv.
