Smartwatch predicts insulin resistance
A machine‑learning model that combines smartwatch streams (heart rate, activity) with routine blood biomarkers and demographics can predict insulin resistance with nearly 90% accuracy — flagging metabolic dysfunction before standard clinical tests change Nature and Science News. This suggests wearables + clinical data could be a practical early‑warning system for targeted prevention.
The WEAR‑ME study enrolled 1,165 participants with a median age of 45 years, median BMI of 28 kg m−2 and median HbA1c of 5.4%, according to the Nature paper. (nature.com) Researchers used HOMA‑IR with a cut‑off of 2.9 as the ground truth and report a multimodal model AUROC of 0.80, with sensitivity 76% and specificity 84% when combining wearable, demographic and routine blood biomarker data. (nature.com) The team fine‑tuned a wearable foundation model (WFM) pretrained on 40 million hours of sensor data, and in an independent validation cohort (n = 72) adding WFM representations to demographics + fasting glucose + a lipid panel raised AUROC from 0.76 to 0.88. (nature.com) The cohort’s smartwatch signals came from Fitbit devices and Pixel Watches, and the authors note that the strongest predictors in their models were clinical and demographic inputs rather than the wearable streams alone. (sciencenews.org) Lead author Ahmed A. Metwally and multiple co‑authors are affiliated with Google Research, where the group first described the approach on the Google Research blog in August 2025. (nature.com) The manuscript was received by Nature on 10 October 2024, accepted on 23 January 2026 and published online in mid‑March 2026, marking a multi‑year effort from initial proposal to peer‑reviewed publication. (nature.com) The paper also reports an LLM integration to contextualize results and generate personalized recommendations from model outputs, a step the authors present as part of a scalable pipeline for translating model signals into actionable guidance. (nature.com)
