New AI model identifies diabetes risk before abnormal test results appear

Millions may be unaware of their early diabetes risk. AI models show why your blood sugar spikes may matter more than your test results.

In a recent paper published in the journal Nature Medicine, researchers analyzed data from more than 2,400 people across two cohorts to identify patterns of glucose spikes and develop personalized glycemic risk profiles.

They found significant differences in glucose spike patterns between people with type 2 diabetes (T2D) and those with prediabetes or normoglycemia. Their multimodal risk model could help doctors identify prediabetics at higher risk of developing T2D.

People with T2DM experienced more severe nocturnal hypoglycemia and took longer, on average more than 20 minutes, to return to baseline glucose levels after spikes – suggesting key physiological differences.

Diabetes and prediabetes affect a significant proportion of the U.S. adult population, yet standard diagnostic tests such as glycated hemoglobin (HbA1c) and fasting glucose do not capture the full complexity of glucose regulation.

Many factors—stress, microbiome composition, sleep, physical activity, genetics, diet, and age—can influence blood glucose swings, especially postprandial spikes (defined as increases of at least 30 mg/dL within 90 minutes), which occur even in seemingly healthy people.

Previously, these variations have been studied using continuous glucose monitoring (CGM), but their coverage has often been limited to prediabetics and normoglycemic individuals, and studies have often lacked representation of historically underrepresented groups in biomedical research.

To address this gap, the PROGRESS study conducted a nationwide, remote clinical trial that enrolled 1,137 diverse participants (48.1% from groups historically underrepresented in biomedical research) with normoglycemia and T2D over 10 days of CGM, while collecting data on microbiome composition, genomics, heart rate, sleep, diet, and activity.

This multimodal approach allowed for a more nuanced understanding of glycemic control and interindividual variability in glucose excursions.

The aim of the study was to create comprehensive glycemic risk profiles that could improve early detection and intervention for prediabetics at risk of progression to diabetes, offering a personalized alternative to traditional diagnostic measures such as HbA1c.

The researchers used data from two cohorts: PROGRESS (a digital clinical trial in the US) and HPP (an observational study in Israel). PROGRESS enrolled adults with and without T2D who underwent 10 days of CGM while simultaneously collecting data on gut microbiome, genomics, heart rate, sleep, diet, and activity.

Gut microbiome diversity (Shannon index) showed a direct negative correlation with average glucose levels: the less diverse the microbiota, the worse the glucose control in all groups.

Participants also collected stool, blood, and saliva samples at home and shared their electronic medical records. Exclusion criteria included recent antibiotic use, pregnancy, type 1 diabetes, and other factors that could confound CGM or metabolic data. Participant recruitment was conducted entirely remotely via social media and invitations based on electronic medical records.

CGM data were processed in minute intervals, and glucose spikes were defined using preset thresholds. Six key glycemic metrics were calculated, including average glucose, time in hyperglycemia, and spike duration.

Lifestyle data were collected using a food diary app and wearable trackers. Genomic and microbiome data were analyzed using standard methods, and composite metrics such as polygenic risk scores and microbiome diversity indices were calculated.

A model for T2DM risk assessment using multimodal data (demographics, anthropometry, CGM, diet, and microbiome) was then constructed using machine learning and its performance was tested in the PROGRESS and HPP cohorts. Statistical analysis used analysis of covariance, Spearman correlations, and bootstrapping to test significance and evaluate the model.

Of the 1137 included participants, 347 were included in the final analysis: 174 with normoglycemia, 79 with prediabetes, and 94 with T2DM.

The researchers found significant differences in glucose spike metrics between conditions: nocturnal hypoglycemia, spike resolution time, average glucose, and time in hyperglycemia. The largest differences were between T2DM and the other groups, with prediabetics statistically closer to normoglycemia than T2DM for key metrics such as spike frequency and intensity.

Microbiome diversity was negatively correlated with most glucose spike metrics, suggesting a healthy microbiome is associated with better glucose control.

Higher resting heart rate, body mass index, and HbA1c were associated with worse glycemic outcomes, while physical activity was associated with more favorable glucose patterns. Interestingly, higher carbohydrate intake was associated with faster peak resolution, but also with more frequent and intense spikes.

The team developed a binary classification model based on multimodal data that discriminated between normoglycemia and T2DM with high accuracy. When applied to an external cohort (HPP), the model retained high performance and successfully identified significant variability in risk levels among prediabetics with similar HbA1c values.

These results suggest that multimodal glycemic profiling may improve risk prediction and individual monitoring compared with standard diagnostic methods, particularly for prediabetes.

The study highlights that traditional diabetes diagnostics such as HbA1c do not reflect individual characteristics of glucose metabolism.

Using CGM in combination with multimodal data (genomics, lifestyle, microbiome), the researchers found significant differences in glucose excursions between normoglycemia, prediabetes, and T2DM, with prediabetes showing greater similarity to normoglycemia than T2DM on a number of key measures.

The developed machine learning-based risk model, validated in an external cohort, revealed a wide variation in risk among prediabetics with similar HbA1c values, confirming its additional value compared to traditional methods.

Strengths of the study include the decentralized, diverse PROGRESS cohort (48.1% from underrepresented groups) and the collection of “real-world” data. However, limitations include potential bias due to device differences, inaccuracies in self-reporting, difficulties in maintaining a food diary, and the use of hypoglycemic medications.

Larger validation and longitudinal studies are needed to confirm the prognostic benefit and clinical significance.

Ultimately, this study demonstrates the potential of remote multimodal data collection to improve early detection, prediabetes risk stratification, and personalized T2D prevention, paving the way for more precise and inclusive care for patients at risk for diabetes.