AI Unlocks Early Diabetes Warning Signs by Analyzing Glucose Spikes

A new AI model developed by Scripps Research can detect early diabetes risk by analyzing real-time glucose spikes and other health data that traditional tests like HbA1c often miss. By using continuous glucose monitors (CGMs), gut microbiome insights, diet, and activity levels, the model provides a more detailed view of metabolic health, helping identify individuals at risk before symptoms appear.

Key Takeaways:

• Traditional HbA1c tests are limited in predicting early diabetes risk.
• Scripps Research developed an AI model that uses CGM data, gut microbiome, diet, and activity levels to assess diabetes progression risk more accurately.
• The PROGRESS study enrolled 1,000+ participants remotely, allowing for a fully self-guided clinical trial.
• Glucose spike recovery time is a critical marker for identifying metabolic health issues before symptoms appear.
• The model can differentiate high-risk prediabetic individuals from those less likely to progress to diabetes, even if they have similar lab results.
• Future applications may allow individuals using CGMs at home to monitor their risk in real-time, enabling early interventions.

Doctors have traditionally relied on a lab test called HbA1c to diagnose type 2 diabetes and prediabetes, measuring a person’s average blood sugar over the past few months. While useful, HbA1c has its limits—it doesn’t reveal who is on the brink of developing diabetes, nor can it differentiate between individuals who share the same test score but have different risk profiles.

A new approach developed by researchers at Scripps Research aims to change that. By harnessing artificial intelligence and combining data from wearable glucose monitors with insights from diet, physical activity, genetics, and gut health, the team has created a model that spots early warning signs of diabetes long before symptoms surface.

Moving Beyond HbA1c: A Deeper Look at Metabolic Health

This innovative model, detailed in Nature Medicine, uses information from continuous glucose monitors (CGMs) to track subtle shifts in blood sugar dynamics. Unlike the static snapshot provided by HbA1c, this AI-driven method captures how the body responds to daily life—meals, exercise, and even sleep patterns—offering a more refined view of metabolic health.

“Two people might have the exact same HbA1c, yet their bodies could be handling glucose in very different ways,” explains Giorgio Quer, co-lead author and Director of AI at Scripps Research. “By monitoring the fine details—how quickly glucose levels normalize after a spike, overnight glucose patterns, dietary habits, and gut microbiota—we can distinguish who is silently edging towards diabetes.”

A Groundbreaking Virtual Clinical Trial: The PROGRESS Study

The project, named the PROGRESS (Prediction of Glycemic Response) Study, enrolled over 1,000 participants across the U.S. through a fully virtual clinical trial. Participants, ranging from healthy individuals to those with prediabetes and diabetes, wore Dexcom G6 CGMs for 10 days. They logged their meals and workouts, and mailed in samples of blood, saliva, and stool. Researchers also tapped into their medical records to gather a comprehensive health profile.

“This was an entirely remote, self-guided study,” notes Ed Ramos, co-lead author and Senior Director of Digital Clinical Trials at Scripps. “We designed a framework that let participants complete everything from home—attaching their own sensors, collecting samples, and shipping them back. It’s a new frontier in clinical research.”

AI Model Identifies Hidden Risk Through Glucose Dynamics

The AI model was trained to sift through this diverse dataset, identifying distinct patterns that correlate with diabetes risk. One standout metric was the time it took for blood sugar levels to stabilize after a spike.

Individuals with type 2 diabetes often needed over 100 minutes to bring their glucose back to baseline, whereas healthier individuals recovered far more rapidly. Additionally, participants with a richer diversity of gut bacteria and higher levels of physical activity showed better glucose control, while those with elevated resting heart rates were more prone to poor sugar regulation.

Detecting Risk in Pre-Diabetic Individuals Before It’s Too Late

What makes this model particularly valuable is its ability to detect risk even when traditional lab results appear normal. Among prediabetic participants, the AI could discern who mirrored diabetic metabolic patterns and who remained closer to a healthy profile, despite similar HbA1c levels. This granularity could enable doctors to personalize interventions, prioritizing lifestyle changes or early treatments for those most at risk.

Looking Ahead: Real-World Impact and Personalized Prevention

The researchers are continuing to monitor participants over time to see if the model’s predictions align with actual disease progression. They’ve also successfully tested the model on an independent patient dataset from Israel, boosting its potential for widespread clinical adoption.

Looking ahead, the team envisions a future where this AI tool could be a standard part of diabetes risk assessment—either through doctors’ offices or directly by individuals using CGMs at home. By offering real-time feedback, it could empower people to understand how their daily habits influence their health and take proactive steps to prevent disease.

“Diabetes doesn’t develop overnight,” Quer emphasizes. “It’s a gradual process, and with the technology we now have, we can detect it earlier and intervene in smarter, more personalized ways.”