Insulin resistance prediction from wearables and routine blood biomarkers

Summary

This Nature study describes WEAR-ME, a large multi-centre effort (n = 1,165) that combines consumer wearable data (Fitbit/Pixel watches), routine blood biomarkers and demographics to predict insulin resistance (measured by HOMA-IR). The team trained and evaluated multimodal models — including gradient-boosted regressors, masked-autoencoder representations and a pretrained wearable foundation model (WFM) — to predict continuous HOMA-IR and binary insulin-resistance status. The best models combine wearable embeddings, fasting glucose and routine lipid/metabolic panels and reach R2 ≈ 0.50 and AUROC up to ≈0.87 (cross-validation) and 0.88 on an independent validation cohort. The authors also built an LLM-based “IR agent” that uses model outputs plus personal data to provide tailored explanations; clinicians rated its responses as more comprehensive and trustworthy than a baseline LLM.

The study shows that simple wearable metrics (resting heart rate, HRV, step counts) correlate with HOMA-IR, but a WFM trained on minute-resolution signals substantially increases wearable contribution to prediction and improves generalisability. The approach could enable scalable pre-screening to prioritise clinical insulin testing and earlier lifestyle or therapeutic interventions.

Key Points

Large cohort (n = 1,165) with wearables, routine blood tests and HOMA-IR ground truth enabled multimodal model training and evaluation.
Wearable-derived signals (higher resting heart rate, lower HRV, lower step counts) correlate with higher HOMA-IR and insulin resistance status.
A wearable foundation model (WFM) trained on minute-resolution sensor data improved prediction vs simple aggregates — wearable feature importance rose from ~43% to ~82% with the WFM.
Best multimodal model (wearables via WFM + demographics + fasting glucose + lipid/metabolic panel) achieved R2 ≈ 0.50 and AUROC ≈ 0.87 (cross-val) and 0.88 on an independent cohort.
Adding fasting glucose to wearables + demographics doubled R2 in some experiments and materially increased true positives while reducing harmful false positives.
Independent validation (n = 72 with complete data) confirmed added value of wearables and WFM-derived embeddings for unseen participants.
Developed an “IR agent” (LLM + tools + IR model) to give personalised, grounded metabolic advice; endocrinologists rated it more comprehensive, trustworthy and personalised than a base LLM.
Authors position the model as a scalable screening tool to prioritise who should get clinical insulin testing (HOMA-IR), not as a replacement for laboratory diagnostics.

Context and relevance

Insulin resistance (IR) is common and often precedes type 2 diabetes; early detection enables lifestyle and pharmacological interventions that can reverse or slow progression. Current clinical practice rarely measures insulin routinely (HOMA-IR requires insulin and fasting glucose), so early-stage IR is often missed. This work leverages the increasing prevalence of consumer wearables together with blood tests already collected in routine exams to create a scalable screening pathway. It fits into broader trends: using continuous personal sensors and foundation-model representations to extract clinically meaningful signals, and combining ML outputs with LLM-based interpretation tools to help users and clinicians understand results. Practical implications include earlier targeted lab testing, better risk stratification for metabolic therapies and potential integration into remote-monitoring services.

Why should I read this?

Quick version: if you care about spotting metabolic risk early (or you build health tech), this paper shows wearables actually add predictive muscle — especially when you use a proper foundation model. It isn’t just correlation fluff: the team validated models, tested robustness, and built a clinical-style LLM assistant. Saves you hours of digging through methods-heavy ML papers and gives clear practical signals about what data and models matter.

Author style

Punchy: big cohort, real-world devices, and a foundation-model boost — the paper delivers a concrete, deployable pipeline rather than a toy experiment. If you work in digital health, primary care screening or device-derived biomarkers, read the details: the WFM and IR agent sections show the real engineering and validation steps necessary to move from signals to clinical utility.

Limitations & takeaways

Limitations highlighted by authors: HOMA-IR is a proxy (not the clamp gold standard), participants with complete data may be biased, most devices were from a single manufacturer, and insulin assays still require clinic visits. The model is best seen as a triage/screening tool to prioritise formal insulin testing and tailored interventions.

Source

Source: https://www.nature.com/articles/s41586-026-10179-2