Insulin Resistance Prediction From Wearables and Routine Blood Biomarkers

A. Ali Heydari; Ahmed A. Metwally; Alexandru Solot; Anthony Z Faranesh; Cathy Speed; Conor Heneghan; Daniel McDuff; David B. Savage; Javier L. Prieto; Menglian Zhou

arxiv: 2505.03784 · v1 · pith:EUGVMZCOnew · submitted 2025-04-30 · 💻 cs.LG

Insulin Resistance Prediction From Wearables and Routine Blood Biomarkers

Ahmed A. Metwally , A. Ali Heydari , Daniel McDuff , Alexandru Solot , Zeinab Esmaeilpour , Anthony Z Faranesh , Menglian Zhou , David B. Savage

show 4 more authors

Conor Heneghan Shwetak Patel Cathy Speed Javier L. Prieto

This is my paper

Pith reviewed 2026-05-25 08:16 UTC · model grok-4.3

classification 💻 cs.LG

keywords insulin resistancewearable devicesblood biomarkersmachine learningHOMA-IR predictiontype 2 diabetesdeep neural networks

0 comments

The pith

Wearable time series and blood biomarkers together predict insulin resistance better than either source alone.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper recruits over 1100 US participants remotely and trains deep neural networks to predict insulin resistance from wearable device data plus routine blood biomarkers. The combined models reach R2 of 0.5 and auROC of 0.80, outperforming single-source versions, with even higher sensitivity in obese and sedentary people. Results hold on an independent validation set of 72 participants and can feed into an LLM agent for interpretation. This matters because current insulin resistance tests are costly and limited, while early detection could support prevention of type 2 diabetes.

Core claim

Deep neural network models that combine wearable device time series with readily available blood biomarkers predict homeostatic model assessment for insulin resistance (HOMA-IR) with R2=0.5, auROC=0.80, sensitivity 76 percent and specificity 84 percent. The models reach 93 percent sensitivity and 95 percent adjusted specificity in obese and sedentary participants and reproduce their performance on an independent cohort of 72 participants.

What carries the argument

Deep neural network models that integrate wearable time series data with blood biomarkers to output a predicted HOMA-IR value.

If this is right

Prediction accuracy rises in the obese and sedentary subgroup that faces the highest risk of type 2 diabetes.
Performance reproduces on a held-out independent cohort of 72 participants.
Predicted HOMA-IR values can be passed to a large language model agent for contextual interpretation and personalized recommendations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Routine wearable and blood data could support population-scale screening without requiring clinic-based insulin clamps or fasting insulin assays.
The LLM integration points to a workflow where predicted values trigger automated but supervised health advice.

Load-bearing premise

Remotely collected wearable time series and blood biomarkers, when paired with measured HOMA-IR in the same people, supply a reliable ground truth that generalizes across populations.

What would settle it

A new cohort drawn from different regions or demographics where the combined model drops below auROC 0.70 or R2 0.3 would falsify the claimed generalizability.

read the original abstract

Insulin resistance, a precursor to type 2 diabetes, is characterized by impaired insulin action in tissues. Current methods for measuring insulin resistance, while effective, are expensive, inaccessible, not widely available and hinder opportunities for early intervention. In this study, we remotely recruited the largest dataset to date across the US to study insulin resistance (N=1,165 participants, with median BMI=28 kg/m2, age=45 years, HbA1c=5.4%), incorporating wearable device time series data and blood biomarkers, including the ground-truth measure of insulin resistance, homeostatic model assessment for insulin resistance (HOMA-IR). We developed deep neural network models to predict insulin resistance based on readily available digital and blood biomarkers. Our results show that our models can predict insulin resistance by combining both wearable data and readily available blood biomarkers better than either of the two data sources separately (R2=0.5, auROC=0.80, Sensitivity=76%, and specificity 84%). The model showed 93% sensitivity and 95% adjusted specificity in obese and sedentary participants, a subpopulation most vulnerable to developing type 2 diabetes and who could benefit most from early intervention. Rigorous evaluation of model performance, including interpretability, and robustness, facilitates generalizability across larger cohorts, which is demonstrated by reproducing the prediction performance on an independent validation cohort (N=72 participants). Additionally, we demonstrated how the predicted insulin resistance can be integrated into a large language model agent to help understand and contextualize HOMA-IR values, facilitating interpretation and safe personalized recommendations. This work offers the potential for early detection of people at risk of type 2 diabetes and thereby facilitate earlier implementation of preventative strategies.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper shows usable multimodal prediction of HOMA-IR from wearables plus blood markers on a sizable remote cohort, but remote label quality is the part that needs checking.

read the letter

The main result is that models using both wearable time series and routine blood biomarkers predict HOMA-IR better than either input alone, hitting R2 of 0.5, auROC 0.8, and solid sensitivity/specificity on a held-out portion of their 1165-person remote cohort, with stronger numbers in the obese sedentary subgroup and reproduction on 72 additional participants. They also add an LLM agent for turning the predictions into usable advice. That combination of scale and practical framing is what stands out. The remote recruitment across the US is larger than most prior digital health studies in this space, and the explicit comparison of single versus combined modalities gives a clear empirical signal that the fusion helps. The subgroup focus on high-risk participants is the part most likely to matter for real prevention work. The LLM integration is a straightforward way to address interpretability without overclaiming. The soft spot is exactly the one the stress-test note flags. HOMA-IR labels come from remote blood draws, and without reported checks on fasting compliance, phlebotomy timing, or cold-chain handling, label noise is plausible. That noise would make the reported metrics look better than they might in cleaner settings and would limit how far the models travel. The abstract is also thin on architecture, splits, feature processing, and missing-data rules, so the technical soundness is hard to judge from the summary alone. If the full paper supplies those details plus ablation results, the concern shrinks; otherwise it stays central. This is for groups working on wearable biomarkers and early diabetes risk tools. A reader who needs a large real-world dataset or wants to see multimodal fusion in practice will get value. It has enough dataset size and a testable claim to deserve referee time, even if the methods section will need tightening.

Referee Report

3 major / 1 minor

Summary. The manuscript presents a remotely recruited US cohort of N=1,165 participants (median BMI 28, age 45, HbA1c 5.4%) with paired wearable time-series and blood biomarkers including HOMA-IR as ground truth. Deep neural network models are trained to predict insulin resistance, claiming superior performance from combined wearable + biomarker inputs (R²=0.5, auROC=0.80, sensitivity 76%, specificity 84%) versus either source alone, with further gains in obese/sedentary subgroups (93% sensitivity, 95% adjusted specificity), reproduction on an independent N=72 validation cohort, and integration of predictions into an LLM agent for contextualized recommendations.

Significance. If the performance metrics are reproducible under full methodological disclosure, the work could enable scalable early detection of type-2-diabetes risk via consumer wearables and routine labs, addressing accessibility barriers of current IR assays. The large remote dataset size and explicit held-out validation cohort constitute concrete strengths that would support generalizability claims if label quality and modeling details are clarified.

major comments (3)

[Abstract/Methods] Abstract and Methods: Performance metrics (R²=0.5, auROC=0.80) are stated without any description of model architecture, training/validation splits, feature processing pipelines, handling of missing wearable data, or statistical testing procedures. This omission is load-bearing because it prevents assessment of whether the reported superiority of the combined model is supported by the experimental design.
[Methods (data collection)] Methods (data collection): No protocol details are supplied on fasting verification, phlebotomy timing, assay methods, or quality-control metrics for the remote blood draws used to compute HOMA-IR labels. This is load-bearing because systematic deviations in fasting compliance or sample integrity would inject unquantified label noise that could inflate the reported R² and auROC values and undermine generalizability.
[Results] Results (validation cohort): Reproduction of performance on the independent N=72 cohort is asserted without reporting cohort selection criteria, demographic or clinical matching to the training set, or confirmation that identical preprocessing and modeling choices were applied. This detail is required to substantiate the generalizability claim.

minor comments (1)

[Abstract] Abstract: Median cohort statistics (BMI, age, HbA1c) are reported without interquartile ranges or other dispersion measures that would aid interpretation of population coverage.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which identify areas requiring greater methodological transparency. We will revise the manuscript to incorporate the requested details on model implementation, data collection protocols, and validation cohort characteristics. Point-by-point responses follow.

read point-by-point responses

Referee: [Abstract/Methods] Abstract and Methods: Performance metrics (R²=0.5, auROC=0.80) are stated without any description of model architecture, training/validation splits, feature processing pipelines, handling of missing wearable data, or statistical testing procedures. This omission is load-bearing because it prevents assessment of whether the reported superiority of the combined model is supported by the experimental design.

Authors: We agree that the current manuscript version does not provide these implementation details at the level required for independent assessment. The revised manuscript will expand the Methods section with: (i) full DNN architecture specifications (layers, units, activations, dropout, regularization); (ii) exact train/validation/test split ratios and any stratification used; (iii) complete feature processing pipeline; (iv) explicit handling of missing wearable time-series (e.g., imputation strategy or exclusion rules); and (v) statistical procedures, including any bootstrap or permutation tests used to compare model performances. These additions will allow direct evaluation of the combined-model superiority claim. revision: yes
Referee: [Methods (data collection)] Methods (data collection): No protocol details are supplied on fasting verification, phlebotomy timing, assay methods, or quality-control metrics for the remote blood draws used to compute HOMA-IR labels. This is load-bearing because systematic deviations in fasting compliance or sample integrity would inject unquantified label noise that could inflate the reported R² and auROC values and undermine generalizability.

Authors: We acknowledge that the manuscript currently omits these protocol specifics. In revision we will add: participant instructions for fasting duration and timing, phlebotomy window relative to the wearable data, laboratory assay methods for glucose and insulin, and any quality-control thresholds or rejection criteria applied by the central lab. If participant-level compliance logs exist they will be summarized; otherwise the limitation will be explicitly noted. revision: yes
Referee: [Results] Results (validation cohort): Reproduction of performance on the independent N=72 cohort is asserted without reporting cohort selection criteria, demographic or clinical matching to the training set, or confirmation that identical preprocessing and modeling choices were applied. This detail is required to substantiate the generalizability claim.

Authors: We agree that these details are necessary to support the generalizability statement. The revised Results and Methods sections will report: (i) explicit selection and inclusion/exclusion criteria for the N=72 cohort; (ii) a table or summary comparing key demographics and clinical variables (age, BMI, HbA1c, etc.) against the main cohort; and (iii) confirmation that the identical preprocessing pipeline, feature set, and trained model weights (or identical inference procedure) were applied without retraining. revision: yes

Circularity Check

0 steps flagged

Empirical supervised learning with held-out validation exhibits no circularity

full rationale

The manuscript trains DNNs to predict HOMA-IR from wearable time series plus routine blood biomarkers and reports performance on an independent validation cohort (N=72). No derivation chain, self-referential equations, or fitted parameters presented as predictions exist. Evaluation uses standard supervised-learning metrics on held-out data; the central claim therefore rests on external data rather than any reduction to its own inputs or self-citations.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Abstract-only review prevents full audit; the central claim rests on the validity of HOMA-IR as ground truth, the representativeness of the remote cohort, and standard supervised learning assumptions about i.i.d. data and appropriate model capacity.

free parameters (1)

Neural network architecture and hyperparameters
Chosen to optimize prediction on the collected data; exact values not stated in abstract.

axioms (1)

domain assumption HOMA-IR calculated from fasting insulin and glucose is a valid proxy for insulin resistance
Used as the supervised target variable throughout.

pith-pipeline@v0.9.0 · 5893 in / 1367 out tokens · 36642 ms · 2026-05-25T08:16:32.462742+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

13 extracted references · 5 canonical work pages

[1]

https://github.com/google-deepmind/onetwo

OneTwo, DeepMind. https://github.com/google-deepmind/onetwo. Accessed: 2024-9-6. XGBoost v2.1.1. https://xgboost.readthedocs.io/en/stable/python/. Accessed: 2024-9-6. Waist circumference is an independent predictor of insulin resistance in black and white youths.J. Pediatr., 148(2):188–194, Feb

2024
[2]

The association between the triglyceride to high-density lipoprotein cholesterol ratio with insulin resistance (HOMA-IR) in the general korean population: Based on the national health and nutrition examination survey in 2007–2009.Diabetes Res. Clin. Pract., 97(1):132–138, July

2007
[3]

Sleep apnea prediction, FDA filing

Accessed: 2024-9-5. Sleep apnea prediction, FDA filing. https://www.accessdata.fda.gov/cdrh_docs/reviews/ DEN230041.pdf),

2024
[4]

European Medicines Agency. Wegovy. https://www.ema.europa.eu/en/medicines/human/EPAR/ wegovy. Accessed: 2025-3-19. Fitbit. What are active zone minutes or active minutes on my fitbit device?https://support.google. com/fitbit/answer/14236509?hl=en#zippy=%2Chow-do-i-earn-active-zone-minutes , a. Ac- cessed: 2024-8-17. Fitbit. Get your heart pumping with fit...

work page arXiv 2025
[5]

W. T. Garvey, R. L. Batterham, M. Bhatta, S. Buscemi, L. N. Christensen, J. P. Frias, E. Jódar, K. Kandler, G. Rigas, T. A. Wadden, and S. Wharton. Two-year effects of semaglutide in adults with overweight or obesity: the STEP 5 trial.Nat. Med., 28(10):2083–2091, Oct

2083
[6]

S. S. Khan, J. Coresh, M. J. Pencina, C. E. Ndumele, J. Rangaswami, S. L. Chow, L. P. Palaniappan, L. S. Sperling, S. S. Virani, J. E. Ho, I. J. Neeland, K. R. Tuttle, R. Rajgopal Singh, M. S. V. Elkind, D. M. Lloyd-Jones, and American Heart Association. Novel prediction equations for absolute risk assessment of total cardiovascular disease incorporating ...

1982
[7]

doi: https://doi.org/10.1016/j.dsx.2016

ISSN 1871-4021. doi: https://doi.org/10.1016/j.dsx.2016. 03.002. URL https://www.sciencedirect.com/science/article/pii/S1871402116300273. A. Mahajan, C. N. Spracklen, W. Zhang, M. C. Y. Ng, L. E. Petty, H. Kitajima, G. Z. Yu, S. Rüeger, L. Speidel, Y. J. Kim, M. Horikoshi, J. M. Mercader, D. Taliun, S. Moon, S.-H. Kwak, N. R. Robertson, N. W. Rayner, M. L...

work page doi:10.1016/j.dsx.2016 2016
[8]

URLhttps://arxiv.org/abs/2503.23339. D. R. Matthews, J. P. Hosker, A. S. Rudenski, B. A. Naylor, D. F. Treacher, and R. C. Turner. Homeostasis model assessment: insulin resistance and𝛽-cell function from fasting plasma glucose and insulin concentrations in man.Diabetologia, 28(7):412–419, July

work page arXiv
[9]

URL https://doi.org/10.1038/ s41586-025-08869-4

doi: 10.1038/s41586-025-08869-4. URL https://doi.org/10.1038/ s41586-025-08869-4. T. McLaughlin, P. Schweitzer, S. Carter, C.-G. Yen, C. Lamendola, F. Abbasi, and G. Reaven. Persistence of improvement in insulin sensitivity following a dietary weight loss programme.Diabetes Obes. Metab., 10(12):1186–1194, Dec

work page doi:10.1038/s41586-025-08869-4
[10]

M. J. Niemann, L. A. Tucker, B. W. Bailey, and L. E. Davidson. Strength training and insulin resistance: The mediating role of body composition.J Diabetes Res, 2020:7694825, May

2020
[11]

Olson, B

K. Olson, B. Hendricks, and D. K. Murdock. The triglyceride to HDL ratio and its relationship to insulin resistance in pre- and postpubertal children: Observation from the wausau SCHOOL project. Cholesterol, 2012(1):794252, Jan

2012
[12]

H. Park, A. A. Metwally, A. Delfarah, Y. Wu, D. Perelman, M. Rodgar, C. Mayer, A. Celli, T. McLaughlin, E. Mignot, and M. Snyder. Lifestyle profiling using wearables and prediction of glucose metabolism in individuals with normoglycemia or prediabetes.medRxiv, page 2024.09.05.24312545, Sept

2024
[13]

Zhang, Y

Q. Zhang, Y. Wang, and Y. Wang. How mask matters: Towards theoretical understandings of masked autoencoders. Adv. Neural Inf. Process. Syst., abs/2210.08344, Oct

work page arXiv

[1] [1]

https://github.com/google-deepmind/onetwo

OneTwo, DeepMind. https://github.com/google-deepmind/onetwo. Accessed: 2024-9-6. XGBoost v2.1.1. https://xgboost.readthedocs.io/en/stable/python/. Accessed: 2024-9-6. Waist circumference is an independent predictor of insulin resistance in black and white youths.J. Pediatr., 148(2):188–194, Feb

2024

[2] [2]

The association between the triglyceride to high-density lipoprotein cholesterol ratio with insulin resistance (HOMA-IR) in the general korean population: Based on the national health and nutrition examination survey in 2007–2009.Diabetes Res. Clin. Pract., 97(1):132–138, July

2007

[3] [3]

Sleep apnea prediction, FDA filing

Accessed: 2024-9-5. Sleep apnea prediction, FDA filing. https://www.accessdata.fda.gov/cdrh_docs/reviews/ DEN230041.pdf),

2024

[4] [4]

European Medicines Agency. Wegovy. https://www.ema.europa.eu/en/medicines/human/EPAR/ wegovy. Accessed: 2025-3-19. Fitbit. What are active zone minutes or active minutes on my fitbit device?https://support.google. com/fitbit/answer/14236509?hl=en#zippy=%2Chow-do-i-earn-active-zone-minutes , a. Ac- cessed: 2024-8-17. Fitbit. Get your heart pumping with fit...

work page arXiv 2025

[5] [5]

W. T. Garvey, R. L. Batterham, M. Bhatta, S. Buscemi, L. N. Christensen, J. P. Frias, E. Jódar, K. Kandler, G. Rigas, T. A. Wadden, and S. Wharton. Two-year effects of semaglutide in adults with overweight or obesity: the STEP 5 trial.Nat. Med., 28(10):2083–2091, Oct

2083

[6] [6]

S. S. Khan, J. Coresh, M. J. Pencina, C. E. Ndumele, J. Rangaswami, S. L. Chow, L. P. Palaniappan, L. S. Sperling, S. S. Virani, J. E. Ho, I. J. Neeland, K. R. Tuttle, R. Rajgopal Singh, M. S. V. Elkind, D. M. Lloyd-Jones, and American Heart Association. Novel prediction equations for absolute risk assessment of total cardiovascular disease incorporating ...

1982

[7] [7]

doi: https://doi.org/10.1016/j.dsx.2016

ISSN 1871-4021. doi: https://doi.org/10.1016/j.dsx.2016. 03.002. URL https://www.sciencedirect.com/science/article/pii/S1871402116300273. A. Mahajan, C. N. Spracklen, W. Zhang, M. C. Y. Ng, L. E. Petty, H. Kitajima, G. Z. Yu, S. Rüeger, L. Speidel, Y. J. Kim, M. Horikoshi, J. M. Mercader, D. Taliun, S. Moon, S.-H. Kwak, N. R. Robertson, N. W. Rayner, M. L...

work page doi:10.1016/j.dsx.2016 2016

[8] [8]

URLhttps://arxiv.org/abs/2503.23339. D. R. Matthews, J. P. Hosker, A. S. Rudenski, B. A. Naylor, D. F. Treacher, and R. C. Turner. Homeostasis model assessment: insulin resistance and𝛽-cell function from fasting plasma glucose and insulin concentrations in man.Diabetologia, 28(7):412–419, July

work page arXiv

[9] [9]

URL https://doi.org/10.1038/ s41586-025-08869-4

doi: 10.1038/s41586-025-08869-4. URL https://doi.org/10.1038/ s41586-025-08869-4. T. McLaughlin, P. Schweitzer, S. Carter, C.-G. Yen, C. Lamendola, F. Abbasi, and G. Reaven. Persistence of improvement in insulin sensitivity following a dietary weight loss programme.Diabetes Obes. Metab., 10(12):1186–1194, Dec

work page doi:10.1038/s41586-025-08869-4

[10] [10]

M. J. Niemann, L. A. Tucker, B. W. Bailey, and L. E. Davidson. Strength training and insulin resistance: The mediating role of body composition.J Diabetes Res, 2020:7694825, May

2020

[11] [11]

Olson, B

K. Olson, B. Hendricks, and D. K. Murdock. The triglyceride to HDL ratio and its relationship to insulin resistance in pre- and postpubertal children: Observation from the wausau SCHOOL project. Cholesterol, 2012(1):794252, Jan

2012

[12] [12]

H. Park, A. A. Metwally, A. Delfarah, Y. Wu, D. Perelman, M. Rodgar, C. Mayer, A. Celli, T. McLaughlin, E. Mignot, and M. Snyder. Lifestyle profiling using wearables and prediction of glucose metabolism in individuals with normoglycemia or prediabetes.medRxiv, page 2024.09.05.24312545, Sept

2024

[13] [13]

Zhang, Y

Q. Zhang, Y. Wang, and Y. Wang. How mask matters: Towards theoretical understandings of masked autoencoders. Adv. Neural Inf. Process. Syst., abs/2210.08344, Oct

work page arXiv