We analyzed linked wearable and laboratory data from 16,760 people to understand whether everyday physiological signals could help identify who may benefit from metabolic screening. Using up to six months of wearable-derived data collected before each person’s first recorded laboratory result, together with age as context, the model meaningfully prioritized people whose HbA1c or fasting-glucose result indicated glycemic dysregulation. The pooled model achieved an AUC of 0.726, with different physiological patterns and performance among women and men. The finding supports a practical two-step workflow for providers and health systems: use passively collected wearable data to surface a potential indication of metabolic risk, then use established laboratory testing and clinician assessment for confirmation.
Key takeaways
- Wearable-derived signals collected before the first recorded laboratory result, together with age, achieved an AUC of 0.726 in the pooled glycemic model. Performance was 0.750 among women and 0.703 among men.
- Day-to-day variability and six-month trends in heart rate, heart-rate variability, sleep, activity, blood oxygen, and related measures contributed additional signals.
- Wearable data can help identify who may benefit from outreach and testing. HbA1c, fasting glucose, clinician assessment, and the patient’s full context remain necessary for diagnosis and care decisions.
Why metabolic risk is often found late
Prediabetes and type 2 diabetes can develop quietly. Screening usually depends on a person visiting a clinician, meeting testing criteria, and completing a blood draw. That makes detection episodic: a patient’s metabolic health may be changing for months or years between tests. Wearables offer a different view. Heart rate, sleep, activity, and related signals are collected repeatedly in daily life. These measurements are not diagnostic, but together they may reveal patterns that help a care team decide who should be screened next. We asked a simple question:
Can wearable-derived signals collected before a person’s first recorded metabolic blood test help distinguish people whose laboratory results indicate glycemic dysregulation?
The data
The linked dataset included 16,760 people with consumer-device history and clinical laboratory records associated at the individual level. The device streams included daily measurements such as:
- Heart rate and heart-rate variability
- Sleep duration, stages, and nocturnal physiology
- Steps, distance, and other activity measures
- Blood oxygen, respiratory rate, weight, and related signals where available
- Platform-derived metabolic measures, including glucose where available
The laboratory records included HbA1c, fasting glucose, lipids, liver markers, renal markers, and inflammatory markers.
The broader analysis cohort contained 15,520 people with sufficient linked coverage and a first laboratory result. For the headline glycemic model, 10,720 people met the device-coverage and outcome requirements. At least 14 days of device data were required within the lookback window.
At the first recorded test, 13.4% of the modeled group met the study’s definition of glycemic dysregulation. Status was determined from HbA1c when available and otherwise from fasting glucose, using established screening thresholds.
Figure 1. Study design. For each person, device signals from up to 180 days before the first recorded laboratory results were summarized as typical level, variability, and trend. No device data from on or after the lab date were used.
What did the wearable data tell us?
The pooled glycemic model achieved an AUC of 0.726. The model achieved 0.750 among women and 0.703 among men. In plain language, the model could meaningfully rank a population by the likelihood that a first recorded laboratory result would indicate glycemic dysregulation. That is useful for deciding where to focus outreach and testing. It is not accurate enough to diagnose or rule out disease. The model used wearable-derived features and age. Laboratory values were used to define the outcome, but they were not used as prediction inputs.
<table header-row="true"> <tr> <td>Glycemic model</td> <td>People</td> <td>Prevalence</td> <td>Cross-validated AUC</td> </tr> <tr> <td>Women</td> <td>4,365</td> <td>10.3%</td> <td>**0.750**</td> </tr> <tr> <td>Men</td> <td>6,354</td> <td>15.5%</td> <td>**0.703**</td> </tr> <tr> <td>Pooled</td> <td>10,720</td> <td>13.4%</td> <td>**0.726**</td> </tr> </table>  ## Why changes over time matter A single average can hide an important story. Two people may have the same average step count, but one may be steadily becoming less active. Likewise, two people may have the same average heart rate, but one may show increasingly irregular daily patterns. Those trajectories can indicate vastly diverging profiles of health. For each device measure, the analysis considered three views: - **Baseline:** What was normal for the person across the window? - **Variability:** How much did the measurement change from day to day? - **Trend:** Was it generally rising, falling, or remaining stable? Variability and changes over time were among the model’s more important features. Examples included changing blood-oxygen patterns, step-count variability, sleep-duration variability, and heart-rate dynamics. This agrees with a growing research literature showing that circadian structure and longitudinal patterns can be more predictive together than episodic readings.[\[1-4\]](/p/d02d9955818e83ba8780016a0e3609f9?pvs=25#3ffd9955818e8146a079d324776d6929) The directions were also physiologically coherent across the cohort: - More and better sleep was generally associated with **lower risk.** - Higher heart-rate variability and lower resting or nocturnal heart rate were generally associated with **lower risk.** - Higher body weight was generally associated with **greater risk** where weight data were available. - Activity levels and improving activity trends contributed to **risk reduction**. ## Where the signal differed between women and men The model did not rely on exactly the same physiological patterns for everyone. Among men, the more informative device signals centered on autonomic and cardiovascular dynamics, including heart-rate structure, heart-rate variability, activity patterns, REM sleep, and total sleep. Among women, sleep depth and body composition had more impact. Deep-sleep duration was the leading non-age device feature, followed by weight measures where available. Skin-temperature variability and heart-rate dynamics also contributed. <empty-block/>  Prior research has also reported sex-specific associations between wearable-derived heart-rate patterns and metabolic syndrome.[\[3\]](/p/d02d9955818e83ba8780016a0e3609f9?pvs=25#3ffd9955818e8146a079d324776d6929) Our results extend that observation in this cohort, and underscore the need for further prospective testing in broader and more representative populations. ## What this says about data from different wearables The analysis operated on harmonized, real-world device data rather than requiring one specialized clinical instrument. Common physiological measures were standardized into a consistent data layer before analysis. This result supports the use of data drawn from a variety of consumer wearable sources to identify and surface potential indications of metabolic dysregulation for follow-up. It is evidence that meaningful signals can remain after heterogeneous device data are normalized and analyzed together. For providers and hospital systems, that points toward a more practical population-health model: work with data patients may already generate, rather than requiring every participant to adopt the same new device. There is an important boundary. This study did not compare performance brand by brand, show that every device measures each signal equally well, nor test how the model performs on a completely unseen device source. Published research likewise finds encouraging comparability for some measures while warning that device methods remain heterogeneous.[\[5,6\]](/p/d02d9955818e83ba8780016a0e3609f9?pvs=25#3ffd9955818e8185b521ca7568f74dc6) The result therefore validates a harmonized multi-source data strategy at the cohort level. Device-specific validation, source-mix analysis, and testing on unseen device sources remain necessary before claiming universal interoperability. ## Why this matters for providers and health systems The most immediate opportunity is not automated diagnosis. It is earlier and more targeted engagement. A provider or health system could use a wearable-derived risk signal to help: - Identify people who may be overdue for metabolic screening - Prioritize outreach for an HbA1c or fasting-glucose test - Close preventive-care gaps across a connected population - Route patients into established clinical pathways - Add longitudinal context between office visits - Support conversations about sleep, activity, weight, and cardiovascular health The proposed workflow is deliberately simple: 1. A patient’s connected wearable data show a pattern associated with elevated metabolic risk. 2. The system surfaces that indication for outreach or care-pathway routing. 3. Conventional laboratory testing and clinician assessment confirm whether action is needed. This preserves the scalability of consumer-generated data while keeping diagnosis and treatment within established clinical practice. ## Constraints of this study The findings are promising, but several boundaries matter: - *It is not a diagnostic test* An AUC in the 0.70-0.75 range supports ranking and prioritization. It does not supplant diagnosing diabetes, ruling it out, or making an autonomous treatment decision. - *It is retrospective* The wearable window occurred before the first recorded lab, which reduces post-diagnosis leakage, but the model was still developed and evaluated retrospectively. Prospective validation is required to understand real-world performance. - *It does not prove detection six months in advance* The model used up to 180 days of history immediately preceding the lab result. It did not test performance at fixed lead times such as 30, 90, or 180 days before the test. - *It has not isolated non-glucose wearable performance* The source data include glucose where available. Although the most prominent findings include heart rate, HRV, sleep, activity, weight, temperature, and blood-oxygen features, the model still requires data from continuous-glucose monitors. - *The cohort is not fully representative of the population as a whole* The broader analysis cohort was 59% male, had a median age of approximately 42, and was concentrated in wellness- and longevity-oriented organizations. Generalization to a broad health-system population remains unproven. - *Device-derived measures are non-diagnostic* Respiratory rate, weight, blood oxygen, energy expenditure, and other fields may be estimated differently by different devices. They are not interchangeable with clinical diagnostic tests. - *Some subgroup findings were underpowered* The female metabolic-dysfunction model had only 3.9% outcome prevalence and achieved an AUC of 0.581, close to chance. Chronic kidney disease was too rare in the cohort to model reliably. ## How the analysis works For every person, we identified the first recorded laboratory date and used device data only from the preceding 180 days. People needed at least 14 days of device coverage to qualify. Each numeric device measure was summarized as its average, variability, and linear trend, producing roughly 220 candidate device features after coverage filtering. Age was included as a contextual feature. Laboratory measurements were excluded from the prediction inputs. Models were trained separately for the pooled population, men, and women using gradient-boosted decision trees. Performance was evaluated with five-fold cross-validation, and feature attribution was examined to understand which signals contributed most strongly. <empty-block/>  The same framework was applied to lipid abnormalities and a broader composite metabolic-dysfunction outcome. Recurring signals included body weight, nocturnal heart rate, sleep regularity, movement, and cardiorespiratory-fitness proxies. <empty-block/>  ## Summary questions - *Can consumer wearable data detect diabetes?* Not by itself. In this analysis, wearable-derived signals helped prioritize people whose first recorded laboratory result indicated glycemic dysregulation. Laboratory testing and clinician assessment remain necessary for diagnosis. - *What did the model actually predict?* It ranked the likelihood that a person’s first recorded HbA1c or fasting-glucose-based status would fall in the dysregulated range. It did not predict the future onset date of diabetes or conversion from healthy to pre-diabetic. - *How well did it work?* The pooled glycemic model achieved an AUC of 0.726. Performance was 0.750 among women and 0.703 among men. That is useful for population prioritization, but not for ruling disease in or out. - *Were laboratory values used to generate the risk score?* No laboratory measurements were included as prediction inputs. They were used to determine the outcome against which the model was evaluated. The model used wearable-derived signals together with age. - *Does the result apply across consumer wearables?* It supports the value of harmonizing common signals from a mixed real-world wearable dataset. It does not establish that every brand performs equally. Brand-level and unseen-device validation are still needed. - *Can providers use the signal instead of ordering an HbA1c test?* No. The intended use is to help identify who may benefit from outreach and appropriate laboratory testing. ## References 1. Zhou W, et al. High-Resolution Digital Phenotypes From Consumer Wearables and Their Applications in Machine Learning of Cardiometabolic Risk Markers. Journal of Medical Internet Research. 2022;24. [https://doi.org/10.2196/34669](https://doi.org/10.2196/34669) 2. Patel MS, et al. Predicting changes in glycemic control among adults with prediabetes from activity patterns collected by wearable devices. NPJ Digital Medicine. 2021;4. [https://doi.org/10.1038/s41746-021-00541-1](https://doi.org/10.1038/s41746-021-00541-1) 3. Mun S, et al. Assessment of heart rate measurements by commercial wearable fitness trackers for early identification of metabolic syndrome risk. Scientific Reports. 2024;14. [https://doi.org/10.1038/s41598-024-74619-7](https://doi.org/10.1038/s41598-024-74619-7) 4. Kim J-K, Mun S, Lee S. Detection and Analysis of Circadian Biomarkers for Metabolic Syndrome Using Wearable Data: Cross-Sectional Study. JMIR Medical Informatics. 2024. [https://doi.org/10.2196/69328](https://doi.org/10.2196/69328) 5. Kingsnorth A, et al. Interchangeability of Research and Commercial Wearable Device Data for Assessing Associations With Cardiometabolic Risk Markers. Journal for the Measurement of Physical Behaviour. 2023. [https://doi.org/10.1123/jmpb.2022-0050](https://doi.org/10.1123/jmpb.2022-0050) 6. Ahmed A, et al. The Effectiveness of Wearable Devices Using Artificial Intelligence for Blood Glucose Level Forecasting or Prediction: Systematic Review. Journal of Medical Internet Research. 2023;25. [https://doi.org/10.2196/40259](https://doi.org/10.2196/40259)