In a recent study published in Nature, researchers used continuous glucose monitors (CGMs) and machine learning to uncover hidden patterns in prediabetes and type 2 diabetes (T2D).
Study: Prediction of metabolic subphenotypes of type 2 diabetes via continuous glucose monitoring and machine learning. Image Credit: Inna Kot/Shutterstock
By analyzing glucose data from at-home oral glucose tolerance tests (OGTTs), they identified specific metabolic subtypes like insulin resistance (IR) and β-cell dysfunction with impressive accuracy. These insights could lead to more personalized approaches for preventing and managing diabetes.
Background
Type 2 diabetes is a global issue, affecting over 537 million people. Despite its prevalence, the way we diagnose and manage diabetes hasn’t changed much in decades. Tools like OGTTs measure blood sugar but don’t dive into the “why” behind abnormal results.
What’s clear is that T2D isn’t the same for everyone. For some, insulin resistance is the main issue, while others may have problems with β-cell function or hormone regulation. These metabolic differences often begin long before diabetes is officially diagnosed, making early detection a critical challenge—especially outside of specialized labs.
That’s where this study comes in. By combining CGMs with machine learning, researchers developed a way to analyze glucose dynamics and identify these unique metabolic patterns—making early detection and targeted interventions more realistic than ever.
Metabolic Phenotyping via OGTT and Machine Learning
The study was part of the Precision Diets for Diabetes Prevention trial and included 36 participants from the San Francisco Bay Area. These participants (aged 30–70, with BMIs ranging from 23–40 kg/m2) were carefully screened to rule out major health issues like organ disease or advanced diabetes. Thirty-two people completed all tests, and a second group of 24 participants helped validate the results.
Participants wore CGMs to monitor glucose levels during both at-home and clinic-based OGTTs. The researchers also conducted advanced metabolic tests to measure things like muscle IR, β-cell function, and hepatic IR. On top of that, they analyzed genetic data to calculate diabetes risk scores.
Using machine learning, the team looked at glucose patterns from the OGTTs—things like how high glucose spiked, how fast it fell, and how long it took to stabilize. These patterns were then matched with the underlying metabolic traits, creating a clearer picture of each person’s unique glucose regulation challenges.
Experimental Results
The study highlighted just how much glucose responses vary from person to person, even among those with similar HbA1c levels (a common marker for long-term blood sugar control). Here’s what stood out:
- Everyone’s different: Glucose patterns varied widely, from the height of glucose spikes to how fast levels dropped and the overall shape of the curve.
- Metabolism is complex: Traits like muscle IR, β-cell dysfunction, and hepatic IR didn’t consistently match standard blood sugar markers, showing how intricate glucose regulation can be.
- Four key subtypes: Researchers identified four main metabolic subtypes: muscle IR, β-cell dysfunction, impaired incretin effect, and hepatic IR. Some participants even showed overlapping traits, like a combination of muscle and hepatic IR.
The machine learning models performed exceptionally well in predicting these metabolic traits based on glucose data, achieving:
- 95 % accuracy for muscle IR.
- 89 % accuracy for β-cell dysfunction.
- 88–90 % accuracy for incretin-related issues.
The models also validated well in a second group of participants, confirming their potential for broader application.
Insights from Glucose Curve Analysis
The researchers also identified four key metabolic subtypes: muscle IR, β-cell dysfunction, impaired incretin effect, and hepatic IR. Interestingly, these subtypes didn’t always line up with standard measures like HbA1c, underscoring the complexity of glucose regulation.
By applying machine learning to glucose time-series data from OGTTs, the team was able to classify these subtypes more effectively than traditional methods. They also showed how CGMs used during at-home OGTTs can serve as a practical, scalable alternative to more invasive and time-consuming tests.
One of the most promising findings was the potential for targeted therapies. For example, GLP-1 agonists could be particularly effective for individuals with insulin resistance or incretin deficiencies. This precision-based approach has the potential to improve treatment outcomes significantly.
Conclusion
This study shows just how much CGMs and machine learning can reveal about what’s happening under the surface in prediabetes and type 2 diabetes. By focusing on specific issues like insulin resistance, β-cell dysfunction, incretin problems, and hepatic IR, the researchers uncovered unique metabolic subtypes that traditional methods often miss. Even better, the machine learning models outperformed conventional approaches, offering a smarter, more precise way to predict these traits and guide tailored treatments.
What’s especially exciting is how accessible this method could be. At-home OGTTs paired with CGMs provide a practical and cost-effective way to catch early signs of glucose regulation issues without requiring invasive tests or frequent trips to the clinic. While it’s still early days and larger studies are needed, this approach could be a game-changer for diabetes prevention and personalized care.
Journal Reference
Metwally et al., 2024. Prediction of metabolic subphenotypes of type 2 diabetes via continuous glucose monitoring and machine learning. Nature Biomedical Engineering. DOI:10.1038/s41551-024-01311-6 https://www.nature.com/articles/s41551-024-01311-6
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.