Google AI Predicts Insulin Resistance from Wearables & Blood Tests

Type 2 diabetes, a condition affecting hundreds of millions globally, continues its concerning rise in prevalence. A critical precursor to this disease is insulin resistance (IR), a state where the body’s cells fail to respond effectively to insulin, the hormone vital for blood sugar regulation. Early detection of IR is paramount, as timely lifestyle adjustments can often reverse the condition, thereby preventing or significantly delaying the onset of type 2 diabetes. However, current gold-standard methods for accurately measuring IR, such as the euglycemic insulin clamp or the Homeostatic Model Assessment for Insulin Resistance (HOMA-IR), are often invasive, expensive, or not routinely available during standard check-ups. These limitations create substantial barriers to early intervention, especially for individuals unknowingly at risk.

Researchers at Google are now exploring a novel approach: leveraging data already accessible to many people, specifically from wearable devices and common blood tests, to estimate IR risk. In their recent work, they introduce a suite of machine learning models designed to predict IR using readily available data points such as resting heart rate, step count, and sleep patterns from wearables, alongside routine blood test results like fasting glucose and lipid panels. This methodology demonstrated robust performance across a study population of 1,165 participants and an independent validation cohort of 72, showing particular efficacy in high-risk groups, including individuals with obesity and sedentary lifestyles.

To explore this potential, Google Research designed the WEAR-ME study, partnering with Quest Diagnostics to automate the collection of routine blood biomarkers. Over 1,100 remote participants from across the US enrolled via the Google Health Studies app, a secure platform for digital health studies, providing electronic informed consent and HIPAA authorization. This diverse cohort, with a median BMI of 28 kg/m² and an average age of 45, consented to share data from their Fitbit or Google Pixel Watch devices (pseudonymized for privacy), results from routine blood tests performed at Quest Diagnostics, and basic demographic information and health questionnaires.

Using this rich, multimodal dataset, advanced machine learning models, specifically deep neural networks, were developed and trained to predict HOMA-IR scores. The goal was to assess how accurately this key IR marker could be estimated using various combinations of the available data. The results, evaluated using the area under the receiver operating characteristic curve (auROC) — a standard measure of classification accuracy — indicated that combining data streams significantly enhanced prediction accuracy compared to using any single source alone. While wearables combined with demographics showed some predictive power (auROC = 0.70), the addition of fasting glucose substantially boosted performance (auROC = 0.78). The most accurate predictions were achieved by combining wearables, demographics, and routine blood panels, yielding an auROC of 0.80 for classifying individuals with IR (with a sensitivity of 76% and specificity of 84% for a HOMA-IR value of 2.9 or higher) and an R² of 0.50 for predicting HOMA-IR values directly. Notably, features derived from wearable data, such as resting heart rate, consistently ranked among the most important predictors, alongside BMI and fasting glucose, underscoring the value of lifestyle-related signals.

Recognizing that individuals with obesity and sedentary lifestyles are particularly susceptible to type 2 diabetes, the models were specifically evaluated in these subgroups. The results were compelling: accuracy improved in obese participants (sensitivity = 86% vs. 76% overall) and was even higher in sedentary participants (sensitivity = 88%). In the most critical group—individuals who were both obese and sedentary—the model performed exceptionally well, achieving a sensitivity of 93% and an adjusted specificity of 95%, suggesting its potential to effectively identify those who could benefit most from early lifestyle interventions.

To ensure the generalizability of these findings, the best-performing model was tested on a completely independent validation cohort of 72 participants. This cohort, whose data included wearable information from Fitbit Charge 6 devices and blood biomarkers acquired in-person, demonstrated that the trained models maintained strong predictive performance, with a sensitivity of 84% and specificity of 81%. It is crucial to note, however, that this remains a research prototype, and its safety and effectiveness for any health-related purpose have not yet been established.

Beyond mere prediction, the researchers also explored how to make this information actionable for individuals. They developed the Insulin Resistance Literacy and Understanding Agent, an AI prototype built on the state-of-the-art Gemini family of large language models. This agent aims to provide personalized, contextualized answers about metabolic health, grounded in the individual’s study data and predicted IR status. With user consent, the agent can access specific data points, search for up-to-date information, and perform calculations. Board-certified endocrinologists evaluated the IR Agent’s responses, overwhelmingly preferring them over a base model, finding them significantly more comprehensive, trustworthy, and personalized. This highlights the transformative potential of integrating predictive health models with advanced AI to empower individuals with a deeper understanding of their health.

This research marks a significant step towards more accessible and scalable screening for type 2 diabetes risk. The approach offers several advantages: leveraging data many people already possess, enabling early detection even before blood sugar levels become abnormal (as many normoglycemic participants in the study were found to have IR), and providing a potentially scalable screening method compared to specialized IR tests. Furthermore, its strong performance in high-risk subgroups and potential for integration into personalized health tools underscore its promise for proactive metabolic health management. Future work will include longitudinal validation of these models, exploring the impact of interventions, incorporating genetic and microbiome data, and refining models to ensure equitable performance across diverse populations. It is important to reiterate that while promising, these models and the IR Agent are currently for informational and research purposes only and are not approved medical devices.