Smartwatch Signals and Blood Tests Team Up for Early Warning on Insulin Resistance

Insulin resistance is a metabolic state in which insulin becomes less effective at regulating blood glucose, often progressing silently until prediabetes or type 2 diabetes emerges. Routine screening can miss early disease and the reference test is costly and time-intensive, limiting widespread use. Early identification enables lifestyle interventions that can reverse risk trajectories. To help address this gap, researchers have introduced a multimodal approach that combines smartwatch data with routine blood tests to improve early detection of insulin resistance.

The approach was evaluated in the Wearables for Metabolic Health (WEAR-ME) study, which aggregated remotely collected smartwatch signals with standard laboratory measures of cholesterol, insulin, and glucose, plus health and lifestyle questionnaires. Investigators trained deep neural networks on data from 1,165 individuals. They then validated model performance using cross-validation and an independent cohort of 72 participants. The findings were published in Nature on March 16, 2026.

The system integrates wearable-derived features with demographic and laboratory information to generate an insulin resistance risk prediction. Fine-tuning the model with a wearable foundation model (WFM) pretrained on 40 million hours of sensor data further enhanced accuracy. A model that combined WFM-derived representations with demographics outperformed a demographics-only baseline, with area under the receiver operating characteristic curve (AUROC) of 0.75 versus 0.66. Adding WFM representations to an optimized model containing demographics, fasting glucose, and a lipid panel improved AUROC to 0.88 from 0.76.

The team also developed an insulin resistance literacy and understanding agent (IR agent) to communicate results to users. The IR agent uses a reason-and-act (ReAct) framework built on a large language model (LLM), in this case Gemini 2.0 Flash, to plan responses, search for current information, perform calculations, and interface with the prediction models. Endocrinologists assessed the agent’s explanations, rating 79% of responses as completely factually accurate and 96% as safe. The agent accurately referenced and interpreted blood test values.

If supported by additional validation, this multimodal strategy could enable scalable, at-home screening and triage pathways for metabolic risk, while reserving resource-intensive testing for those most likely to benefit. Earlier identification of insulin resistance may facilitate timely counseling on weight management, exercise, and diet, with the goal of preventing progression to diabetes and avoiding downstream complications.