Introduction to Insulin Resistance and Type 2 Diabetes

Type 2 diabetes is a chronic condition that currently affects hundreds of millions of people worldwide, and its prevalence is on the rise. One of the most significant precursors to this disease is insulin resistance (IR). In this condition, the body’s cells do not respond properly to insulin, a hormone vital for regulating blood sugar. Early detection of insulin resistance is of paramount importance because it can often be reversed with lifestyle changes, preventing or delaying the onset of type 2 diabetes.

However, current methods for accurately measuring insulin resistance, such as the “gold standard” (euglycemic insulin clamp) or the Homeostatic Model Assessment of Insulin Resistance (HOMA-IR), which require specific insulin blood tests, are often invasive, expensive, or not readily available in routine check-ups. These barriers create significant challenges for early diagnosis and timely interventions, especially for individuals unknowingly at risk.

But what if we could use data already available to many people, such as data from wearable devices and routine blood tests, to estimate the risk of insulin resistance? In the paper “Insulin resistance prediction from wearables and routine blood biomarkers,” we explore a set of machine learning models that have the potential to predict insulin resistance using data from wearable devices (such as resting heart rate, step count, sleep patterns) and routine blood tests (such as fasting blood glucose, lipid panel).

This approach demonstrates strong performance in the study population (N=1,165) and an independent validation cohort (N=72), particularly in high-risk individuals such as those with obesity and sedentary lifestyles. Furthermore, we introduce the Insulin Resistance Education and Understanding Agent (an IR prototype agent) built upon the advanced Gemini family of large language models (LLMs) to aid in understanding insulin resistance and facilitate safe, personalized interpretation and recommendations. This work offers the potential for early detection of individuals at risk of type 2 diabetes, thereby facilitating earlier implementation of preventative strategies.

Image: Insulin resistance prediction from wearables and routine blood biomarkers.

Predicting Insulin Resistance Using Digital Biomarkers and Routine Blood Tests

We designed a study called WEAR-ME to investigate the potential of predicting insulin resistance (via HOMA-IR prediction) using readily available data. To automate the data collection process for routine blood biomarkers, we partnered with Quest Diagnostics. 1,165 participants from across the United States remotely enrolled in the WEAR-ME study through the Google Health Studies app, a secure and user-friendly platform for digital studies. The study was conducted with Institutional Review Board (IRB) approval.

All participants provided electronic informed consent and specific HIPAA authorization through the Google Health Studies app prior to enrollment. This cohort was diverse in terms of age, gender, geography, and Body Mass Index (BMI). Participants had a mean BMI of 28 kg/m², a mean age of 45 years, and an HbA1c of 5.4%. Participants agreed to share the following data:

• Wearable device data: Data from their Fitbit or Google Pixel Watch devices (such as resting heart rate, step count, sleep patterns) which were de-identified in a pseudo-anonymous manner to protect participant privacy.
• Routine blood biomarkers: Results from routine tests (such as fasting blood glucose and insulin, lipid panel) specifically requested for this research during an in-person visit at Quest Diagnostics.
• Demographic information and surveys: Baseline information and health questionnaires (completed at the beginning and end of the study) including age, weight, height, ethnicity, race, and gender data, along with questions about general health perceptions (fitness, diet) and history of diabetes or other comorbid conditions.

Using this rich, multimodal dataset (which we refer to as “WEAR-ME data”), we developed and trained deep neural network models to predict HOMA-IR scores. Our goal was to see how well we could estimate this key indicator of insulin resistance using various combinations of available data. This novel approach paves the way for earlier diagnosis and intervention.

Image: Metabolic subphenotypes of type 2 diabetes. Chronic insulin resistance is a precursor to approximately 70% of type 2 diabetes cases and results from a combination of obesity, an inactive lifestyle, and genetic factors.

Model Performance and Key Predictive Factors

Our results, using the Area Under the Receiver Operating Characteristic curve (auROC) metric, show that combining data streams significantly improved prediction accuracy compared to using any single source alone:

• Wearables + Demographics: Showed predictive power (auROC = 0.70) for classifying insulin resistance. This indicates that even without blood information, lifestyle data can provide valuable insights.
• Adding fasting glucose to Wearables + Demographics: The results of this routine blood test were highly valuable, dramatically improving performance (auROC = 0.78). This combination highlights the importance of core metabolic markers.
• Wearables + Demographics + Routine Blood Panels: Achieved the best results with high accuracy in predicting HOMA-IR values (R² = 0.50) and effectively classifying individuals with insulin resistance (auROC = 0.80, Sensitivity = 76%, Specificity = 84%, where a HOMA-IR value of 2.9 or higher was used to identify an individual as insulin resistant).

These results clearly demonstrate that utilizing a multimodal approach to data has the potential for significant improvement in the early detection of insulin resistance. By combining continuous, non-invasive data from wearable devices with more precise laboratory information, we can obtain a more comprehensive picture of an individual’s metabolic health.

Importantly, our results indicate that features derived from wearable device data, such as resting heart rate, consistently ranked among the most important predictors, alongside BMI and fasting blood glucose. The feature importance results highlight the value of recording lifestyle-related signals. These findings emphasize that wearables are not just tools for activity tracking but can serve as vital sensors for overall health.

Image: Schematic of our proposed model for HOMA-IR prediction and interpretation of results with the Insulin Resistance Education and Understanding Agent.

Image: Left: Performance evaluation of IR prediction (classification). Right: Visualization of the precision-recall curve for selected feature sets. Average values are colors, with the gray areas around each line indicating the standard deviation across the five folds.

Focus on High-Risk Groups and Model Validation

Since individuals with obesity and sedentary lifestyles are particularly vulnerable to developing type 2 diabetes, we specifically evaluated our model’s performance in these subgroups. The results showed a significant improvement in model accuracy within these groups, demonstrating the true potential of this approach for early detection in targeted populations:

• Obese participants: The model showed improved accuracy compared to the general population (Sensitivity = 86% vs. 76%). This increased accuracy allows us to identify high-risk individuals in this group with greater confidence.
• Sedentary participants: The model’s accuracy was even higher than the obese subpopulation (Sensitivity = 88%). This indicates that physical activity patterns, even regardless of weight, are a significant factor in predicting insulin resistance.
• Obese and sedentary participants: The model performed exceptionally well in this crucial group (Sensitivity = 93%, adjusted Specificity = 95%; adjusted specificity here focuses on minimizing misclassification of truly insulin-sensitive individuals as resistant). This finding confirms the value of our approach in identifying individuals who would benefit most from early lifestyle interventions.

The results of this experiment indicate that our approach can be particularly effective in identifying those who might benefit most from early lifestyle interventions. This can help clinicians and individuals take more targeted preventive measures before the disease progresses.

To ensure that our findings were not merely specific to our initial dataset, we tested our best-performing model (trained on WEAR-ME data) on a completely independent validation cohort (N=72) collected through a separate IRB-approved study. In this study, participants shared their wearable device data using a Fitbit Charge 6, and blood biomarker data was obtained in-person at the study center in San Francisco. This cohort had a mean BMI of 30.6 kg/m² and a mean age of 44.5 years. Our results on the validation cohort demonstrate that our trained models maintained their strong predictive performance (Sensitivity = 84%, Specificity = 81%), indicating their potential generalizability.

This external validation is a critical step in confirming the reliability of our models in real-world scenarios. However, as this is a research prototype, its safety and effectiveness for any health-related purpose have not yet been proven. For the actual deployment of this technology, rigorous testing, validation, and regulatory approval are required.

Image: Sankey diagram showing the relative feature importance (SHapley Additive exPlanations [SHAP] values) for each of the proposed nonlinear XGBoost models for direct regression.

Image: Results of classification performance for various lifestyle stratifications.

Image: Overview of the independent validation cohort study. We compare model accuracies from the initial training and testing cohort with the external validation cohort and demonstrate its potential generalizability.

Beyond Prediction: Towards Understanding and Preventive Actions

Predicting the risk of insulin resistance is valuable, but how can we make this information understandable and actionable for individuals? We explored the integration of our predictive models with large language models (LLMs) to empower users to better understand their metabolic health. We developed the Insulin Resistance Education and Understanding Agent (an IR prototype agent), built upon the advanced Gemini family of LLMs. When asked a question about metabolic health, the IR agent provides personalized and contextually relevant responses for educational purposes, based on the individual’s study data and their predicted IR status. With user consent, this agent has the ability to access specific user-provided data points, search for up-to-date information, and perform calculations. It is important to note that interaction with the models or the IR agent is solely to demonstrate how such a tool could assist users in exploring their results for informational and educational purposes.

We engaged five board-certified endocrinologists to evaluate the IR agent’s responses compared to a baseline model. They strongly preferred the IR agent’s responses, finding them significantly more comprehensive, trustworthy, and personalized. This highlights the potential of combining health predictive models with LLMs to empower individuals with a better understanding of their health. This step creates a bridge between complex data and understandable information for the general public, engaging them in their health management journey.

Image: Overview of the Insulin Resistance Literacy and Understanding Agent (IR Agent). An illustration of the proposed IR agent (left), along with the results (win rate) of our IR agent against the base model as evaluated by endocrinologists (right).

Conclusion and Future Work

Our research demonstrates that machine learning models, by combining wearable device data and routine blood biomarkers, have high potential for effectively predicting insulin resistance, a key precursor to type 2 diabetes. This approach offers several advantages:

• Accessibility: Uses data that many people already have or can easily obtain.
• Early detection: Identifies risk even before blood sugar levels become abnormal; for example, we found many participants with normal blood glucose (with HbA1c < 5.7) in our study who already had insulin resistance.
• Scalability: Offers a potentially more scalable screening method than specialized insulin resistance tests.
• Personalization: Shows strong performance in high-risk subgroups and potential for integration into personalized health tools.

This work opens the door for earlier and more accessible screening for type 2 diabetes risk, potentially enabling timely lifestyle interventions that could prevent or delay the disease, especially for those unknowingly progressing towards it. These advancements can contribute to improving public health on a wide scale.

Future work includes longitudinal validation of these models (following individuals over time), investigating the impact of interventions, incorporating genetic and microbiome data, and further refining the models for specific populations to ensure fair performance across diverse groups. We believe that this line of research holds promise for proactive and personalized metabolic health management and could revolutionize type 2 diabetes prevention.

Disclaimer:
While our proposed approach, including the IR agent, holds promise for various health applications, this research specifically addresses the critical need for early detection of insulin resistance and does not present the models discussed here as approved medical devices or solutions. The models and IR agent are not medical devices. They have not been cleared, approved, or reviewed by the U.S. Food and Drug Administration (FDA) or any other national or international regulatory body. This work is not intended to be and should not be used as a substitute for professional medical advice, diagnosis, or treatment. Actual deployment of such technologies would require rigorous testing, validation, and regulatory approval.

Acknowledgements:
The research described herein is a collaborative effort by Google Research and collaborating teams. The following researchers contributed to this work: Ahmed A. Metwally, A. Ali Heydari, Daniel McDuff, Alexandru Solot, Zeinab Esmaeilpour, Anthony Z. Faranesh, Menglian Zhou, David B. Savage, Conor Heneghan, Shwetak Patel, Cathy Speed, and Javier L. Prieto. Google partnered with Quest Diagnostics, the world’s leading provider of diagnostic information, to enable eligible participants to share their biomarker data, received as part of a free blood test including a comprehensive metabolic panel and measurements of cholesterol, triglycerides, and insulin levels.

Source: Google Research Blog: Insulin resistance prediction from wearables and routine blood biomarkers