Early Diabetes Detection with AI, Identifying Hidden Risk in Glucose

Aug 4
2 min read

A gray-haired doctor in a white coat talks to a young patient. They are in a medical consultation room, with a monitor displaying anatomical models in the background.

In a breakthrough poised to transform public health, artificial intelligence (AI) is demonstrating an unprecedented ability to detect hidden diabetes risk, long before traditional diagnostic methods can. Two recent studies, presented at the American Diabetes Association's 85th Scientific Sessions, highlight that early diabetes detection with AI will be more accurate and personalized, thanks to the tool's potential.

Millions of people may be missing the opportunity for early diagnosis of diabetes or prediabetes, as standard tools like glycated hemoglobin (HbA1c) or fasting glucose fail to capture the full complexity of glucose regulation. Numerous factors, including stress, microbiome composition, sleep, physical activity, genetics, diet, and age, significantly influence blood glucose fluctuations, particularly post-meal spikes, which have been observed even in seemingly healthy individuals.

One study, published in Nature Medicine, analyzed data from over 2,400 people across two cohorts, including a diverse group from the PROGRESS study with 48.1% of participants from groups historically underrepresented in biomedical research. Researchers used continuous glucose monitoring (CGM) along with a wide range of multimodal data—from the genome and gut microbiome to lifestyle information—to create glycemic risk profiles. They found that reduced gut microbiome diversity directly correlated with poorer glucose control across all groups. Furthermore, higher daily carbohydrate intake, while leading to quicker glucose spike resolution, also triggered more frequent and intense spikes. Their AI model could distinguish normoglycemic individuals from those with Type 2 Diabetes (T2D) with high accuracy and successfully identified substantial variability in risk levels among prediabetic individuals with similar HbA1c values, making it superior to conventional tests.

Simultaneously, AI is making waves in detecting Type 1 Diabetes (T1D). Each year, around 64,000 Americans are diagnosed with T1D, and as many as 40% are unaware of the disease until they experience a life-threatening event requiring hospitalization. By this point, significant and often irreversible damage to insulin-producing cells has already occurred, emphasizing the need for earlier detection. New AI models, trained with millions of health records from claims databases, have managed to identify T1D risk up to a year before clinical diagnosis, with significantly greater accuracy and fewer false positives than current screening methods. One of the most effective models, Bidirectional Encoder Representations from Transformers (BERT), correctly identified 80% of true T1D cases and showed higher accuracy than other models. These models also revealed that 29% of T1D cases had previously been misclassified as Type 2 diabetes or other forms, highlighting a critical diagnostic gap that can delay appropriate treatment.

These findings suggest a future where medicine will be more proactive and personalized, allowing for early interventions before the disease progresses significantly. AI in diabetes not only helps anticipate the disease but also offers a path toward more precise and inclusive healthcare, improving the lives of millions.