image: CGMformer is first self-supervised pretrained on CGM data to gain fundamental knowledge of the glucose dynamics, and then applied to a multitude of downstream clinical applications. The extractable contextual time point and individual embeddings can be used as an intrinsic representation for daily glucose profiles in clinical applications including screening, subtyping, and postprandial glucose prediction and dietary suggestion.
Credit: ©Science China Press
A team of researchers led by Dr. Yong Wang (Academy of Mathematics and Systems Science, Chinese Academy of Sciences), Dr. Huating Li, Dr. Weiping Jia (Shanghai Sixth People’s Hospital Affiliated to Shanghai Jiao Tong University School of Medicine), and Dr. Luonan Chen (Center for Excellence in Molecular Cell Science, Shanghai Institute of Biochemistry and Cell Biology, Chinese Academy of Sciences) has developed CGMformer, a deep learning model that learns from large-scale continuous glucose monitoring (CGM) data to improve diabetes screening, risk assessment, and personalized treatment. This study leverages the recent breakthrough in artificial intelligence (AI) and offers a new approach to understanding glucose metabolism, providing deeper insights into individual metabolic health and disease progression.
Type 2 diabetes (T2D) is a complex metabolic disorder influenced by genetics and lifestyle factors. Traditional diagnostic methods, such as fasting blood glucose and HbA1c tests, provide only a partial view of glucose regulation, often missing subtle but important fluctuations that signal early disease onset. To overcome these limitations, the research team developed CGMformer, an AI model trained on large-scale CGM data to extract individual glucose dynamics.
“We tokenize CGM data and glucose values were discretized into distinct glucose levels and ordered by time points to mimic a sentence structure. The transformer architecture in NLP allows us to utilize attention mechanism to handle long-range dependencies with ease. Pretraining in large scale CGM data creates an intrinsic representation for individual glucose dynamics and transfer learning boosts diverse downstream tasks. Together, those efforts lead to the outperformance over existing dynamics, statistical and machine learning analysis”, says Ms. Yurun Lu, developer of CGMformer from the Academy of Mathematics and Systems Science, Chinese Academy of Sciences.
CGMformer captures a continuous picture of glucose fluctuations, identifying patterns that may indicate early metabolic dysfunction, long before traditional laboratory tests would detect abnormalities.
"CGM data offers a unique window into glucose metabolism, but interpreting these patterns manually is challenging. Our model leverages deep learning to uncover hidden glucose dynamics, accurately and stably represent individual’s metabolic state, and eventually improve early detection and risk prediction," says Dr. Yong Wang.
Beyond diabetes screening, CGMformer enables more personalized risk assessments, classifying individuals into different metabolic subtypes based on their glucose dynamics. This includes the identification of previously overlooked high-risk individuals, such as those with normal BMI but impaired glucose regulation. These findings could reshape how prediabetes and early-stage diabetes are diagnosed and managed, allowing for timely interventions tailored to each person’s metabolic profile.
In addition to screening and risk assessment, the researchers introduced CGMformer_Diet, an extension of the model that predicts how different foods will affect an individual’s blood glucose levels. By integrating CGM data with dietary intake, the model enables personalized nutrition strategies to maintain stable glucose levels. Through AI-driven simulations, the researchers demonstrated how adjusting macronutrient intake—such as reducing carbohydrates or increasing protein—could optimize postprandial glucose responses. These findings could help guide personalized dietary recommendations, offering practical tools for individuals managing diabetes or seeking to prevent its onset.
"By predicting an individual’s response to food, we can provide a proactive and personalized approach to nutrition, helping people make informed choices to support their metabolic health." explains Dr. Huating Li, T2D expert from Shanghai Diabetes Institute and Shanghai Sixth People’s Hospital.
The development of CGMformer marks a significant advancement in AI-powered healthcare, offering a comprehensive view of glucose metabolism that goes beyond traditional diagnostic methods. As continuous glucose monitoring becomes more widely adopted, integrating AI models like CGMformer into clinical practice could enable more accurate, early detection of metabolic disorders and more effective, personalized interventions.
As AI and wearable health technologies continue to evolve, models like CGMformer pave the way for smarter, data-driven approaches to metabolic health, with the potential to extend beyond diabetes to other chronic conditions.
See the article:
A pretrained transformer model for decoding individual glucose dynamics from continuous glucose monitoring data
https://doi.org/10.1093/nsr/nwaf039
Journal
National Science Review