Insulin Resistance (IR), a condition where the body’s cells do not properly respond to insulin, a vital hormone for regulating blood sugar, is a major precursor to type 2 diabetes. This condition affects hundreds of millions of people worldwide, and its prevalence is on the rise. Early diagnosis of insulin resistance is crucial, as lifestyle changes can often reverse it, preventing or delaying the onset of type 2 diabetes. However, current methods for accurately measuring IR, such as the euglycemic insulin clamp or HOMA-IR, are often invasive, expensive, or not readily available in routine examinations. These barriers make early diagnosis and intervention difficult, especially for individuals unknowingly at risk.
In the paper “Predicting Insulin Resistance from Wearables and Routine Blood Biomarkers,” we explore a suite of machine learning models that have the potential to predict IR using data readily available to many individuals, including data from wearable devices (e.g., resting heart rate, step count, sleep patterns) and routine blood tests (e.g., fasting glucose, lipid panel). This approach demonstrates strong performance in the studied population (N=1,165) and an independent validation cohort (N=72), particularly in high-risk individuals such as those who are obese and have sedentary lifestyles.
Furthermore, we introduce an Insulin Resistance Literacy and Understanding Agent (an early IR Agent) built upon advanced Gemini large language models (LLMs). This agent helps in understanding insulin resistance, facilitating interpretation, and providing personalized and safe recommendations. This work offers the potential for early detection of individuals at risk of type 2 diabetes, thereby facilitating the early implementation of preventative strategies. It is important to note that the models, predictions, and the Insulin Resistance Literacy and Understanding Agent are intended for informational and research purposes only.
Predicting Insulin Resistance Using Digital Biomarkers and Routine Blood Tests
We designed a study, titled WEAR-ME, to explore the potential for predicting insulin resistance (via HOMA-IR prediction) using readily accessible data. To automate the data collection process for routine blood biomarkers, we collaborated with Quest Diagnostics. The study was conducted with Institutional Review Board (IRB) approval, and all 1,165 participants from across the United States, via the Google Health Studies app, provided electronic informed consent and HIPAA authorization prior to enrollment.
The participant cohort was diverse in terms of age, gender, geography, and Body Mass Index (BMI); the mean BMI was 28 kg/m², mean age was 45 years, and mean HbA1c was 5.4%. Participants consented to share wearable data from Fitbit or Google Pixel Watch devices (e.g., resting heart rate, step count, sleep patterns), routine blood biomarker results (e.g., fasting glucose and insulin, lipid panel), and demographic and survey information. These data were collected pseudonymously to preserve privacy.
Using this rich, multimodal dataset, which we call “WEAR-ME data,” we developed and trained deep neural network models to predict HOMA-IR scores. Our goal was to investigate how accurately this key IR marker could be estimated using various combinations of available data. This approach allows us to seek new ways for early detection of insulin resistance, which could lead to timely lifestyle interventions to prevent or delay the onset of type 2 diabetes.
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, sedentary lifestyle, and genetic factors.
Illustration of our proposed modeling pipeline for predicting HOMA-IR, and interpreting the results with the Insulin Resistance Literacy and Understanding Agent.
Powerful Predictions from Combined Data Sources
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 each source alone. This finding underscores the importance of a holistic view of health, where diverse data from various sources can complement each other to provide a more accurate picture of an individual’s metabolic status.
Specifically, the combination of wearable data and demographic information showed acceptable predictive power for classifying insulin resistance (auROC = 0.70). Adding fasting glucose results to this combination added significant value to the model, dramatically increasing performance (auROC = 0.78). Fasting glucose is a common and relatively inexpensive blood test that provides important information about blood sugar regulation.
Finally, combining wearable data, demographic information, and full routine blood panel results (including lipids and other markers) yielded the best outcomes. This model accurately predicted HOMA-IR values (R² = 0.50) and effectively classified individuals with insulin resistance (auROC = 0.80, sensitivity = 76%, specificity = 84%, where a HOMA-IR value of 2.9 or higher was considered the threshold for insulin resistance). These results demonstrate the high potential of this combined approach to provide a robust and non-invasive screening tool for insulin resistance.
Left: Performance evaluation of IR prediction (classification). Right: Visualization of precision-recall curves for selected feature sets. Mean values are indicated by colors, with gray regions around each line representing the standard deviation across five folds.
Most Important Factors Influencing Prediction Accuracy
Our results indicate that features extracted from wearable data, such as resting heart rate, were consistently among the most important predictors, alongside BMI and fasting glucose. This finding emphasizes the high significance of lifestyle-related signals captured by wearable devices.
Body Mass Index (BMI), a simple yet powerful measure of obesity, plays a vital role in predicting insulin resistance, as obesity is one of the strongest risk factors for insulin resistance and type 2 diabetes. This suggests that even without complex measurements, basic body information can provide valuable insights.
Fasting glucose, as a traditional blood biomarker, is also strongly associated with insulin resistance. These three factors – wearable data, BMI, and fasting glucose – appear to form a powerful combination for building accurate predictive models. These insights not only help us build better models but also suggest pathways for more targeted interventions for individuals at risk of insulin resistance.
Sankey diagram showing the relative feature importance (SHAP values) for each of the proposed non-linear direct XGBoost regression models.
Improved Performance in High-Risk Groups
Given that 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 goal was to understand if our model could more effectively identify these at-risk populations, where early detection could have the greatest impact.
Among obese participants, the model showed higher accuracy compared to the overall population (sensitivity = 86% vs. 76%). For sedentary participants, the accuracy was even higher than the obese subgroup (sensitivity = 88%). This finding emphasizes the importance of wearable data in detecting insulin resistance risk in individuals who may be predisposed to the condition due to physical inactivity.
Notably, in the group that was both obese and sedentary, the model demonstrated excellent performance (sensitivity = 93%, adjusted specificity = 95%). These results suggest that our approach can be particularly effective in identifying those who would benefit most from early lifestyle interventions.
Results of classification performance for various lifestyle stratifications.
Real-World Validation and Generalizability
To ensure that our findings were not solely specific to our initial dataset, we tested our best-performing model (trained on WEAR-ME data) on a completely independent validation cohort (N=72). This cohort was recruited through a separate study with IRB approval, and participants shared wearable and blood biomarker data.
Our results in this validation cohort indicate that our trained models maintained their robust predictive performance (sensitivity = 84%, specificity = 81%), proving their potential for generalizability. This means the model can perform effectively across different populations and using data collected in varying ways. This external validation increases confidence in the model’s capabilities and paves the way for further research and, ultimately, clinical application.
However, it must be noted that this is still a research prototype, and its safety and efficacy for any health-related purpose have not yet been proven. We are committed to continuing the research and development of these models to ensure their accuracy, safety, and efficacy at scale, with the understanding that any health-related technology requires rigorous testing and regulatory approvals.
Overview of the independent validation cohort study. We compare the models’ accuracy from the initial training and testing cohort with the external validation cohort, demonstrating its potential for generalizability.
Beyond Prediction: Towards Understanding and Preventive Actions
Predicting insulin resistance risk is valuable, but how can we make this information understandable and actionable for individuals? We explored how to integrate our prediction models with large language models (LLMs) to empower users in better understanding their metabolic health. We developed an Insulin Resistance Literacy and Understanding Agent (an early IR Agent) built upon advanced Gemini LLMs.
When queried about metabolic health, the IR Agent provides personalized, contextualized responses for educational purposes, grounded in the individual’s study data and predicted IR status. With user consent, the agent has the capability to access specific data points, search for up-to-date information, and perform computations. This interactive capability makes the agent a powerful tool for educating and empowering individuals.
We asked 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 demonstrates the potential of combining health prediction models with LLMs to empower individuals with a better understanding of health. It is important to note that interaction with these models or the IR Agent is solely for informational and educational purposes and does not have a medical aspect.
Illustration of the proposed agent architecture, which leverages the HOMA-IR prediction model to assess insulin resistance risk for user education.
Overview of the Insulin Resistance Literacy and Understanding Agent (IR Agent). Illustration of our proposed IR Agent (left), along with results (win rate) of our IR Agent against a baseline model as evaluated by endocrinologists (right).
Conclusion and Future Work
Our research demonstrates that machine learning models combining readily available wearable data and routine blood biomarkers have the potential to effectively predict insulin resistance. This approach offers significant advantages, including high accessibility, early detection (even in individuals with normal blood glucose), and high scalability for screening. It also shows strong performance in high-risk subgroups and potential for integration into personalized health tools.
This work paves the way for earlier and more accessible screening for type 2 diabetes risk, potentially enabling timely lifestyle interventions that can prevent or delay the disease, especially for those unknowingly progressing toward it.
Future work involves longitudinal validation of these models, investigating the impact of interventions, incorporating genetic and microbiome data, as well as further refining the models for specific populations to ensure equitable performance across diverse groups. We believe this line of research holds significant promise for proactive and personalized metabolic health management.
Disclaimer: While our proposed approach, including the IR Agent, shows promise for various health applications, this research specifically addresses the critical need for early insulin resistance detection and does not present the models discussed herein as approved medical devices or solutions. The models and IR Agent are not medical devices and have not been cleared, approved, or reviewed by the U.S. Food and Drug Administration (FDA) or any other regulatory agency. 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 here is a collaborative effort by Google Research and partnering teams. The following researchers contributed to this work: Ahmed Motawlley, Ali Heydari, Daniel McDuff, Alexandru Sola, Zeinab Esmailpour, Anthony Faraone, Menglian Zhou, David Bisavath, Conor Heneghan, Shwetak Patel, Kathy Speed, and Javier L. Prieto. Google collaborated with Quest Diagnostics to allow eligible participants to share their biomarker data received as part of a free blood test.
