Smartwatch Data and Routine Blood Tests Predict Insulin Resistance Before Diabetes Appears
A deep learning model combining wearable data with routine labs flagged insulin resistance (AUROC 0.80–0.88), potentially enabling earlier lifestyle intervention.
Resumen
Insulin resistance often develops years before type 2 diabetes, yet the tests that detect it are costly or rarely ordered. In the WEAR-ME study, 1,165 US adults wore Fitbit or Pixel devices and had routine blood work, with HOMA-IR as the reference measure. Models combining wearable data, demographics, fasting glucose and lipids identified insulin resistance (HOMA-IR >2.9) with an AUROC of 0.80, 76% sensitivity and 84% specificity. A wearable foundation model pretrained on 40 million hours of sensor data further improved results; in a small independent cohort (n=72), adding it raised AUROC from 0.76 to 0.88. The team also built an LLM agent that explains results, which endocrinologists preferred over a base LLM. Findings are promising but need larger, more diverse validation.
Resumen detallado
Insulin resistance (IR) is a core precursor to type 2 diabetes and is linked to fatty liver disease and cardiovascular disease. It affects an estimated 20–40% of people, yet screening usually relies on fasting glucose or HbA1c, which can miss early IR. The gold-standard clamp test is impractical outside research settings, and HOMA-IR requires a lab visit with fasting insulin. A scalable, accessible way to flag IR early could open a window for lifestyle intervention.
The WEAR-ME study was conducted remotely through the Google Health Studies app. It enrolled 1,165 US adults (median age 45, BMI 28, HbA1c 5.4%) who shared Fitbit or Pixel Watch data and ordered blood tests through Quest Diagnostics. HOMA-IR served as ground truth, with IR defined as >2.9 (300 participants), insulin sensitivity as <1.5 (459) and impaired sensitivity in between (406). Wearable time series were summarized or embedded with a masked autoencoder and then passed to tree-based models in fivefold cross-validation across 25 feature combinations. The team also fine-tuned a wearable foundation model pretrained on 40 million hours of data.
HOMA-IR correlated positively with fasting glucose (r=0.57), HbA1c, BMI, triglycerides and resting heart rate (r=0.27), and negatively with HDL, daily steps and HRV. Wearables plus demographics predicted IR with AUROC 0.70; adding fasting glucose raised it to 0.78; adding lipid and metabolic panels reached 0.80 (sensitivity 76%, specificity 84%). Notably, 20% of normoglycaemic participants had IR. In an independent cohort of 72 people, foundation-model embeddings plus demographics outperformed demographics alone (AUROC 0.75 vs 0.66), and adding them to demographics, glucose and lipids improved AUROC from 0.76 to 0.88.
The researchers also built an LLM-based agent (Gemini 2.0 Flash with ReAct) that uses the IR prediction to answer metabolic health questions. Five endocrinologists preferred it over the base model for comprehensiveness (80%), trustworthiness (92%) and personalization (73%). Responses were judged safe in 96% of cases and fully factual in 79%, although interpretation of blood biomarkers was weaker (59%).
Caveats include the modest independent validation cohort, a single HOMA-IR cut-off, reliance on a surrogate rather than clamp-derived IR, and the modest absolute performance of wearable-only models. The agent evaluation used only five participant profiles. Broader, more diverse and prospective validation, including whether predictions drive better outcomes, is still needed before clinical use.
Hallazgos clave
- Wearables, demographics, fasting glucose and lipid/metabolic panels predicted insulin resistance with AUROC 0.80, 76% sensitivity and 84% specificity in 1,165 adults.
- In an independent cohort (n=72), adding foundation-model wearable embeddings raised AUROC from 0.76 to 0.88 over a labs-and-demographics model.
- 20% of normoglycaemic participants (196 of 972) had insulin resistance, showing HbA1c alone misses early metabolic risk.
- Adding fasting glucose to the wearable model doubled R² from 0.212 to 0.435 and cut false positives by 46%.
- An LLM agent using IR predictions was preferred by endocrinologists over a base LLM and rated safe in 96% of responses.
Metodología
Remote, app-based cohort study (WEAR-ME, n=1,165) combining Fitbit/Pixel Watch data, Quest blood biomarkers and demographics, with HOMA-IR as ground truth. Tree-based models and masked-autoencoder and wearable foundation model embeddings were evaluated with fivefold cross-validation, then tested with frozen weights in an independent in-person cohort (n=72). An LLM agent was assessed by five board-certified endocrinologists.
Limitaciones del estudio
The independent validation cohort was small (n=72), and the primary cohort was largely US-based, self-selected and had few people with diabetes (34). HOMA-IR is a surrogate for IR with variable cut-offs, and wearable-only performance was modest. The LLM agent was evaluated on only five participant profiles, with weaker accuracy for interpreting blood biomarkers (59%) and wearable data (79%).
¿Te ha gustado este resumen?
Recibe la última investigación sobre longevidad en tu bandeja de entrada cada semana.
Introduce tu correo electrónico para suscribirte: