Traditional methods for diagnosing and monitoring type 2 diabetes (T2D), such as episodic glycated haemoglobin (HbA1c) assays, may not fully capture the complex, multifaceted nature of the disease. A new study suggests that a multimodal artificial intelligence (AI) approach, integrating data from wearables, genomics, and the gut microbiome, can provide a more detailed and informative phenotype than HbA1c alone, potentially improving risk stratification for individuals with pre-diabetes.¹ With nearly 98 million adults in the US having pre-diabetes, enhanced risk assessment is crucial.²
The PRediction Of Glycemic RESponse Study (PROGRESS) was a prospective, digital clinical trial conducted in the US. This analysis focused on 347 deeply phenotyped participants, comprising 174 normoglycaemic individuals, 79 with pre-diabetes, and 94 with T2D.
Over a 10-day period, researchers collected a wide range of remote, real-world data, including:
- Continuous glucose monitoring (CGM) data
- Wearable sensor data (activity, sleep, resting heart rate)
- Dietary intake from food logs
- Genomic data (polygenic risk score)
- Gut microbiome composition from stool samples
- Electronic health records
Using this information, an AI model was developed to create multimodal glycaemic risk profiles. The model's findings were then validated in an independent external cohort of 2,069 individuals from the Human Phenotype Project (HPP).
The analysis revealed significant differences in glucose spike metrics across the three glycaemic states. Compared to normoglycaemic individuals, those with T2D had a longer expected time for glucose spike resolution and higher values of nocturnal hypoglycaemia (p<0.001 for both).
Several significant correlations were identified. Greater gut microbiome diversity was associated with lower mean glucose levels (r=−0.301; p<0.001), while higher levels of physical activity correlated with more favourable glucose spike metrics. The AI-generated multimodal glycaemic risk profiles demonstrated substantial variability among individuals who had the same HbA1c value, indicating that the model provides a more nuanced risk assessment than the standard biomarker alone. These findings were consistent upon validation in the HPP cohort.
These findings highlight the potential of integrating diverse, real-world data to move beyond single-point measurements in diabetes care. By creating a more detailed individual phenotype, this AI-driven approach could help clinicians better identify pre-diabetic individuals at high risk of progressing to T2D, allowing for earlier and more targeted interventions. The authors concluded that "such a multimodal approach provides a detailed phenotype that can potentially improve T2D prevention, diagnosis and treatment, and is more informative than HbA1c."¹
The research group plans to conduct a longitudinal study to evaluate the prognostic impact of these glucose spike metrics on adverse clinical outcomes in pre-diabetic and diabetic populations. Future research will also explore the integration of novel tools, such as proteomic organ clocks, into the analysis.
References
1. Carletti M, Pandit J, Gadaleta M, et al. Multimodal AI correlates of glucose spikes in people with normal glucose regulation, pre-diabetes and type 2 diabetes. Nat Med 2025. https://doi.org/10.1038/s41591-025-03849-7
2. National Diabetes Statistics Report. Centers for Disease Control and Prevention, 2024. https://www.cdc.gov/diabetes/php/data-research/index.html
Disclaimer: The information presented in this article is for educational purposes only. Any quotes included reflect the opinions of the individual quoted, and do not necessarily reflect the views of the publisher. The publisher does not guarantee the accuracy or completeness of the content and accepts no responsibility for any errors, or any consequences arising from its use.