Overview
- The WEAR-ME cohort analysis of 1,165 adults using Fitbit and Google Pixel watch data with demographics and blood biomarkers achieved an AUROC of 0.80 with 76% sensitivity and 84% specificity at a HOMA-IR cutoff of 2.9.
- A wearable foundation model pretrained on about 40 million hours of sensor data improved validation performance, reaching AUROC 0.88 when added to lipids, fasting glucose, and demographics compared with 0.76 without wearable data.
- In an independent validation cohort of 72 participants, demographics plus WFM-derived representations scored AUROC 0.75 versus 0.66 for demographics alone.
- An NIH News & Views commentary noted that continuous signals from daily life can reveal physiological strain that episodic clinic tests may miss.
- The team also introduced an LLM-based IR agent built on Gemini 2.0 Flash that endocrinologists reviewed for safety and factual accuracy, while authors cautioned that only about a quarter of enrollees had complete data and results rely on Google/Fitbit devices, underscoring the need for broader validation before deployment.