Pith. sign in

REVIEW 3 major objections 5 minor 125 references

This paper releases OpenMHC, a dataset of 67 million hours of minute-level wearable data from 11,894 consenting participants, and claims — for the first time at open scale — that wearable foundation models can be pretrained, compared, and r

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 09:58 UTC pith:3LLQRTZT

load-bearing objection OpenMHC is the real thing — a genuinely large open wearable dataset with a careful benchmark — but the full data isn't out yet and the non-wear heuristic deserves scrutiny before the numbers are taken as ground truth. the 3 major comments →

arxiv 2607.16235 v2 pith:3LLQRTZT submitted 2026-06-25 cs.LG cs.AI

OpenMHC: Accelerating the Science of Wearable Foundation Models

Narayan Schuetz , Yuze Bai , Lianggang Pan , Edgar Eggert , Favour Nerrise , Juan Delgado-SanMartin , Max Rosenblattl , Milana Gurbanova
show 11 more authors
Mohammad Asadi Anders Johnson Paul Schmiedmayer Dennis Wang Allan Lawrie Daniel Seung Kim Xin Liu Akshay Paruchuri Ehsan Adeli Euan Ashley Kelly W. Zhang
This is my paper
classification cs.LG cs.AI
keywords wearable healthfoundation modelsopen datasetbenchmarkimputationforecastingHealthKitself-supervised learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper releases OpenMHC, a dataset of 67 million hours of minute-level wearable data from 11,894 consenting participants in the My Heart Counts study, alongside 169 linked health and lifestyle variables. Its central claim is that this scale, combined with open model implementations, lets researchers train and fairly compare wearable foundation models on real-world, openly accessible data for the first time. To support that claim, the authors build a benchmark with three tracks: predicting 32 self-reported health outcomes, imputing masked sensor channels, and forecasting the next 24 hours. On this benchmark, a reimplementation of the masked-autoencoder LSM-2 leads prediction and imputation, a fine-tuned Chronos-2 leads forecasting, and a hand-crafted XGBoost remains the closest competitor to both. The authors are explicit that these are detection and characterization tasks at the time of survey, not prospective diagnosis.

Core claim

The central discovery is the resource itself plus a set of empirical findings obtained on it. OpenMHC is claimed to be the largest and most comprehensive open-access wearable health dataset to date: 67 million wear-hours, 19 minute-level channels covering activity, physiology, sleep, and workouts, and 169 linked self-reported and sparse device variables. On the accompanying benchmark, LSM-2 (a reimplementation of Google's masked sensor autoencoder) achieves the best aggregate prediction skill score (+15.1% relative to a linear reference), LSM-2-Sparse (the same encoder paired with a sparse cross-day decoder over a 7-day context) leads imputation (+64.7% relative to last-observation-carried-f

What carries the argument

The load-bearing objects are daily matrices d ∈ R^{19×1440} (19 channels at minute resolution) with explicit NaN semantics; a non-wear heuristic that flags runs of at least 30 minutes where all seven continuous channels are simultaneously zero or NaN; a day-level wear-time filter retaining days with at least 12 hours of estimated wear; and the skill score, a geometric mean of per-task error ratios against a fixed reference (Linear for prediction, LOCF for imputation, Seasonal Naive for forecasting). These heuristics convert raw HealthKit streams into benchmark-ready tensors, and the skill score converts heterogeneous task metrics into a single comparable number. The pretraining machinery is

Load-bearing premise

Every benchmark number rests on the hand-set rule that a run of at least 30 minutes with all seven continuous channels zero or NaN means the device was not worn, plus the 12-hour wear-day filter that discards about half the cohort; if sedentary wear is misread as removal, or the filter drops a different population, the scores inherit that error.

What would settle it

A validation study in which participants log actual device-off times, compared against the 30-minute zero/NaN rule, would directly estimate the non-wear misclassification rate; alternatively, rerunning the three benchmark tracks under a 6-hour or no wear-time filter and checking whether LSM-2 and Chronos-2 leads persist would reveal whether the leaderboard is an artifact of the 12-hour filter.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Researchers can pretrain and evaluate wearable foundation models on real-world consumer data without negotiating access to proprietary datasets.
  • The open reimplementations of LSM-2 and WBM, alongside classical baselines, allow apples-to-apples comparisons that were previously impossible for the broader community.
  • The reported results imply that masked reconstruction is better suited to sparse, irregular wearable data than contrastive weekly pretraining, guiding future architecture choices.
  • Adding longitudinal context helps: the 7-day sparse cross-day decoder improves imputation beyond single-day models, pointing to personal history as a key frontier.
  • All prediction tasks are framed as detection at survey time, not early diagnosis; the paper explicitly states it does not support adjudicated incident-event prediction.
  • If these benchmark numbers stand, the field gains a shared testbed for scaling-law studies, cross-device transfer, and fairness evaluations in wearable health.

Where Pith is reading between the lines

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

  • Inference: The non-wear heuristic and the 12-hour wear filter are hand-set and remove roughly half the cohort; if sedentary periods are misread as device removal, every predicted, imputed, and forecasted value inherits that error, so the reported leaderboard is only as strong as these heuristics.
  • Inference: Because the paper itself (Section A.1) acknowledges that all variables are self-reported and that outcomes are measured at the time of survey, the predictive track is best read as a probe of representation quality rather than a clinical-risk result; early-diagnosis claims would require adjudicated endpoints the study does not collect.
  • Inference: The demographic shifts from wear-time filtering (older participants, more hypertension) suggest that fairness scores computed after filtering may understate disparities in the full enrolled population.
  • Inference: The success of masks that mimic real-world missingness suggests a testable extension: training imputers on masks derived from actual device-off logs, rather than structural and semantic simulations, would tighten the link between benchmark performance and real-world deployment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces OpenMHC, an open-access wearable health dataset derived from the My Heart Counts study, comprising 11,894 consenting participants, 67M wear-hours, 19 minute-level sensor channels, and 169 linked variables. The authors also release open-source reimplementations of two wearable foundation models (WBM, LSM-2) and a benchmark with three tracks: 32 health/behavior prediction tasks, minute-level time-series imputation with six masking scenarios, and 24-hour forecasting on hourly data. Headline results are that LSM-2 leads the prediction track (skill score +15.1% vs Linear), LSM-2-Sparse leads imputation (+64.7% vs LOCF), and fine-tuned Chronos-2 leads forecasting (+37.6% vs Seasonal Naive). The paper is unusually careful in several design choices: participant-level train/val/test splits, participant-level bootstrap confidence intervals on all headline numbers, explicit fallback contracts for non-finite predictions, and an explicit disclaimer that the prediction tasks use both pre- and post-survey wearable data and are not early-diagnosis tasks.

Significance. If the dataset and benchmark hold up, this is a substantial community resource: it is the largest open-access consumer wearable dataset with linked health variables that the authors document, and it provides reproducible baselines and open implementations of previously proprietary models. The evaluation protocol is methodologically strong in its use of external reference models (Linear, LOCF, Seasonal Naive), participant-level bootstrapping, and explicit handling of missingness and non-finite predictions. The honest discussion of what the dataset cannot support (incident-event prediction, adjudicated outcomes) is a strength. However, the central artifact's validity rests on a non-wear heuristic that is not validated against any ground truth, and the full dataset is not yet publicly available, so the scale claims cannot currently be independently verified.

major comments (3)
  1. [Appendix D.1 / D.3.1; Table 11; statement in A.1] The non-wear heuristic is load-bearing and unvalidated. For the 41% of participants who are phone-only (Figure 5b), the watch channels are structurally absent, so the 'all 7 continuous channels zero/NaN' rule reduces to 'all three phone channels are zero' for 30 consecutive minutes. This will misclassify sedentary but device-carrying periods as non-wear. That binary non-wear vector then drives the 12-hour wear-time filter (D.3.1) that retains only 6,031 of 11,894 participants, and every prediction, imputation, and forecasting number in Tables 2–4 is computed on this filtered cohort. The sensitivity analysis in D.4 checks only demographic composition, not the correctness of the wear detector. The A.1 claim that 'data quality filtering introduces no meaningful bias' is contradicted by Table 11, which shows statistically significant shifts in age, BMI, sex, hypertension, and CVD prevalence.
  2. [Data Release / Section 3] The abstract and Section 3 state that OpenMHC is 'released,' but the Data Release paragraph says the full dataset 'will be released upon publication of the manuscript,' with only the MHC-XS 5% subsample currently available. The central claim of the paper is the scale and comprehensiveness of the released artifact, yet the full data are not accessible to reviewers or the community at the time of review. Please provide a concrete data-availability plan or a reviewer-accessible full-data snapshot, and clarify the exact terms under which the 67M wear-hours will be distributed.
  3. [Table 11 vs Figure 10 / Figure 8d] There is an internal numeric inconsistency in the retained cohort size after the 12-hour wear-time filter. Figure 8d reports 6,031 participants (50.7%) with at least one day meeting ≥12h wear time, and Table 11 reports 6,031 retained at the ≤50% non-wear threshold, but Figure 10's preprocessing flow shows n=5,941 after 'Wear-Time Threshold (≤50% non-wear)' filtering. This discrepancy needs to be resolved; if it reflects an additional low-variance filtering step, the text should say so explicitly, because all downstream task participant counts in Table 13 are derived from this cohort.
minor comments (5)
  1. [Appendix B.2] The fairness skill score is defined for per-task errors, but the exact per-task aggregation for the prediction track (where each participant has a single label) is not as explicitly specified as for the generative tracks (Appendix F.4). Please state whether the prediction-track disparity is computed on participant-level errors or on pooled task-level errors.
  2. [Section 3 / Figure 10] The participant flow diagram and the text should distinguish 'contributed passive HealthKit data' from 'retained days after quality filtering' more clearly, especially because Table 13 reports task-specific included participant counts that are much smaller than the 6,031/5,941 cohort.
  3. [References] Several references are incomplete or have placeholder formatting: 'Retamales et al.' lacks a year/venue, the 'Toye et al.' reference appears in the text but the bibliography entry is formatted inconsistently, and 'Shchur et al.' is listed as 'et al.' without full author list. Please run a reference check.
  4. [Appendix A.6 / Regulatory Context] The regulatory discussion is thoughtful, but it would be helpful to state explicitly that these are the authors' interpretations, not regulatory determinations (the text already says this in one place, but the section title and framing invite confusion).
  5. [Table 2 and Table 4] In Table 2, the confidence intervals for some fairness scores are very wide (e.g., XGBoost S_fair CI spans -30.3 to +17.2). Consider reporting the fairness score only at the aggregate level or drawing less strong conclusions about 'fairest model' from such noisy subgroup estimates.

Circularity Check

0 steps flagged

No significant circularity: benchmark results are scored on held-out participant-level splits against external references, and no central claim reduces to its own inputs.

full rationale

Walking the paper's derivation chain, I find no circular step. Track 1 (prediction) is scored against the Linear reference on a held-out participant-level test split (60/10/30), with skill scores defined as error ratios relative to that external reference (Eq. 1, Appendix B.1); the labels (Appendix E.1) are self-reported survey or HealthKit values, not functions of the evaluated models' outputs. Tracks 2A/2B are likewise anchored to external references (LOCF; Seasonal Naive) with pre-defined masks and held-out rolling evaluation, so the reported +15.1% (LSM-2), +64.7% (LSM-2-Sparse), and +37.6% (Chronos-2 FT) are empirical comparisons, not quantities produced by construction. The reimplemented WBM and LSM-2 are attributed to other groups (Erturk et al. 2025; Xu et al. 2025b), no uniqueness theorem is imported, and no ansatz is smuggled via self-citation. The paper's self-citations (McConnell 2017; Hershman 2019; Schmiedmayer 2026; Javed 2023; Delgado-SanMartin 2026; Merrill 2026; and Narayanswamy 2025/2026 via co-author overlap) appear in motivation, related work, or study-design context; none is load-bearing for the dataset-scale claims or benchmark findings, and the scaling-law motivation cites industry results on proprietary data — independent evidence, not OpenMHC feeding back onto itself. The fallback-to-Linear mechanism (Appendix E.2) is conservative (it forces error ratio 1 against the reference) and is reported with substitution rates. Flagged for the record: the non-wear heuristic, the 12-hour wear filter, and the claimed 'no meaningful bias' in the limitations (A.1) sit in tension with Table 11's statistically significant demographic shifts — but this is a measurement-assumption/validity concern about cohort selection, not a circular derivation, because the predictions are not equivalent to the heuristic by construction. Verdict: no significant circularity.

Axiom & Free-Parameter Ledger

9 free parameters · 4 axioms · 0 invented entities

The paper introduces no invented entities. Its constructive content consists of dataset curation, benchmark definitions, preprocessing thresholds, and tuned model hyperparameters. The free parameters above are all disclosed in the text or appendices — a genuine strength — but they are numerous and several materially shape the reported results: the non-wear heuristic (30 min) and the 12-hour day filter define which daily matrices exist at all, the 10% availability threshold defines which participants have which channels, and the mask parameters define imputation difficulty. The harness substitutes Linear/LOCF reference predictions for any NaN model output and reports the substitution rate (Appendix E.2 / F.3), which limits but does not eliminate the concern that fallback-heavy methods are partly scoring the reference baseline.

free parameters (9)
  • Non-wear detection run length = 30 minutes
    Appendix D.1: consecutive runs where all 7 continuous channels are simultaneously zero/NaN define non-wear; hand-set threshold that determines wear-time statistics, the 12-hour day filter, and coverage summaries.
  • Wear-time day retention threshold = ≥12 h/day (≤50% non-wear)
    Appendix D.3.1: days are dropped below 12 h estimated wear, retaining ~50.7% of participants for modeling; Table 11 reports statistically significant demographic shifts at this threshold. Benchmark default.
  • Per-channel stream availability threshold = 10% of days
    Appendix D.1 / Table 9: channels observed on <10% of a participant's days are labelled structurally unavailable (NaN fills all days), determining which participants can be scored on which channels.
  • Low-variance day filter thresholds = 1.0 (steps/distance/energy); 1e-4 (heart rate)
    Appendix D.3.1: days with near-constant but nonzero traces are removed as 'likely sensor artifacts/glitches'; thresholds chosen by hand.
  • Per-channel missing run length = 120 minutes
    Appendix D.1: runs of ≥120 min of zero/NaN per channel are stored as missing intervals, feeding missingness metadata used downstream.
  • Sleep-reliability threshold = 3 hours
    Appendix D.3.1: for sleep channels, zero-valued minutes are set to NaN when total detected sleep on the day is <3 h; hand-set.
  • Imputation mask design parameters = r=0.5, r=0.25; patch 30 min; blocks 30–60 min; HR>160 bpm, ≥5 min runs
    Appendix F.1: random-noise mask ratio, temporal-slice ratio and block sizes, and the intensity-failure trigger set the difficulty of the imputation track; benchmark-design choices.
  • Skill-score clipping bounds = ℓ=0.01, u=100
    Appendix B.1: error ratios are clipped to [0.01, 100] before geometric-mean aggregation; a wide clip chosen for numerical stability that can compress extreme wins.
  • WBM pretraining hyperparameters = drop_prob 0.223, λ_KoLeo 0.689, lr 1.3e-5 (Bayesian-optimized)
    Appendix E.2.5: selected via a two-stage Bayesian sweep on validation InfoNCE loss on a 1,783-participant proxy; other baselines were tuned on validation too. The comparisons reflect tuned implementations, so the scores are partly functions of these fitted choices.
axioms (4)
  • domain assumption HealthKit lacks explicit missingness; zeros in an available stream encode genuine inactivity, and simultaneous zero/NaN across all 7 continuous channels for ≥30 min means device non-wear
    Section 3 and Appendix D.1. The entire benchmarking pipeline — wear-time filtering, non-wear masks, and thus every daily matrix entering the three tracks — depends on this heuristic distinguishing 'inactive' from 'not wearing.' If it mislabels sedentary periods as non-wear, all reported skill scores inherit the distortion.
  • domain assumption Self-reported survey responses and HealthKit-derived values (e.g., BMI from self-reported height/weight) are treated as ground-truth labels for the 32 prediction tasks
    Appendix A.1 concedes 'all variables are self-reported, introducing a degree of label noise' and attributes failures to beat Linear on some medical targets partly to this noise. The benchmark scores are only as valid as these labels.
  • standard math Participant-level bootstrap (1,000 replicates) with percentile/BCa intervals gives valid uncertainty estimates for skill score, rank, and fairness aggregates
    Appendix B. Standard practice for benchmark uncertainty quantification; assumes participants are exchangeable units and that point estimates on the fixed test split are stable under resampling.
  • domain assumption The recording of broad vs. narrow consent in the study database is correct, so the 11,894 released users are exactly those who consented to open sharing
    Section 3 and Figure 4: 5,099 users were excluded for partner-only or no-sharing consent. A consent misclassification would change the cohort and breach the data-use terms; the paper provides no audit of this step.

pith-pipeline@v1.3.0-alltime-deepseek · 60191 in / 26254 out tokens · 248719 ms · 2026-08-02T09:58:26.099937+00:00 · methodology

0 comments
read the original abstract

Mobile and wearable devices offer an unprecedented opportunity for continuous, passive health monitoring and active health coaching. However, the largest wearable datasets are not publicly available for research, and leading wearable foundation models trained on such datasets are rarely open-weight or come with reproducible training code. To accelerate open science in wearable health, we release OpenMyHeartCounts (OpenMHC), the largest and most comprehensive open-access wearable health dataset to date, alongside open-source implementations of recent wearable foundation models. OpenMHC, derived from over a decade of data collected through the My Heart Counts study app, includes >60 million hours of wearable data across 19 sensor channels (e.g., step count, heart rate, sleep, workouts) and up to 169 linked variables, including health, lifestyle, mood, and behavior from 11,894 consenting participants. Furthermore, we introduce a unified, open benchmark that enables standardized comparison of wearable health models across three tracks: health and behavior downstream prediction, multivariate data imputation, and time-series forecasting. We benchmark classical methods alongside recent wearable and multivariate time series foundation models. By open-sourcing data, code, and model weights at this unprecedented scale, we aim to democratize wearable health AI research and enable the community to drive open progress in this domain.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

125 extracted references · 3 canonical work pages

  1. [1]

    Statistical science , volume=

    Bootstrap confidence intervals , author=. Statistical science , volume=. 1996 , publisher=

    1996

  2. [2]

    Journal of clinical epidemiology , volume=

    Clinical prediction models: diagnosis versus prognosis , author=. Journal of clinical epidemiology , volume=. 2021 , publisher=

    2021

  3. [3]

    npj Cardiovascular Health , volume=

    Assessing the feasibility of using smartphone data to identify risk of idiopathic pulmonary arterial hypertension , author=. npj Cardiovascular Health , volume=. 2026 , publisher=

    2026

  4. [4]

    npj Digital Medicine , volume=

    Unsupervised machine learning to investigate trajectory patterns of COVID-19 symptoms and physical activity measured via the MyHeart Counts App and smart devices , author=. npj Digital Medicine , volume=. 2023 , publisher=

    2023

  5. [5]

    Proceedings of the Conference on Health, Inference, and Learning , pages =

    Homekit2020: A Benchmark for Time Series Classification on a Large Mobile Sensing Dataset with Laboratory Tested Ground Truth of Influenza Infections , author =. Proceedings of the Conference on Health, Inference, and Learning , pages =. 2023 , editor =

    2023

  6. [6]

    Computational Linguistics , volume=

    Probing classifiers: Promises, shortcomings, and advances , author=. Computational Linguistics , volume=

  7. [7]

    New England Journal of Medicine , volume=

    Large-scale assessment of a smartwatch to identify atrial fibrillation , author=. New England Journal of Medicine , volume=. 2019 , publisher=

    2019

  8. [8]

    Circulation , volume=

    Detection of atrial fibrillation in a large population using wearable devices: the fitbit heart study , author=. Circulation , volume=. 2022 , publisher=

    2022

  9. [9]

    Journal of the American College of Cardiology , volume=

    Accelerometer-measured sedentary behavior and risk of future cardiovascular disease , author=. Journal of the American College of Cardiology , volume=. 2025 , publisher=

    2025

  10. [10]

    Advances in neural information processing systems , volume=

    Superglue: A stickier benchmark for general-purpose language understanding systems , author=. Advances in neural information processing systems , volume=

  11. [11]

    Proceedings of the 2018 EMNLP workshop BlackboxNLP: Analyzing and interpreting neural networks for NLP , pages=

    GLUE: A multi-task benchmark and analysis platform for natural language understanding , author=. Proceedings of the 2018 EMNLP workshop BlackboxNLP: Analyzing and interpreting neural networks for NLP , pages=

    2018

  12. [12]

    2009 IEEE conference on computer vision and pattern recognition , pages=

    Imagenet: A large-scale hierarchical image database , author=. 2009 IEEE conference on computer vision and pattern recognition , pages=. 2009 , organization=

    2009

  13. [13]

    JAMA cardiology , volume=

    Feasibility of obtaining measures of lifestyle from a smartphone app: the MyHeart Counts Cardiovascular Health Study , author=. JAMA cardiology , volume=

  14. [14]

    The Lancet Digital Health , volume=

    The effect of digital physical activity interventions on daily step count: a randomised controlled crossover substudy of the MyHeart Counts Cardiovascular Health Study , author=. The Lancet Digital Health , volume=. 2019 , publisher=

    2019

  15. [15]

    Scientific data , volume=

    Physical activity, sleep and cardiovascular health data for 50,000 individuals from the MyHeart Counts Study , author=. Scientific data , volume=. 2019 , publisher=

    2019

  16. [16]

    npj Digital Medicine , year=

    Digital biomarkers for brain health: passive and continuous assessment from wearable sensors , author=. npj Digital Medicine , year=

  17. [17]

    NPJ Digital Medicine , volume=

    Circadian rhythm analysis using wearable-based accelerometry as a digital biomarker of aging and healthspan , author=. NPJ Digital Medicine , volume=. 2024 , publisher=

    2024

  18. [18]

    Scientific reports , volume=

    Examining health apps and wearable use in improving physical and mental well-being across US, China, and Singapore , author=. Scientific reports , volume=. 2024 , publisher=

    2024

  19. [19]

    Annals of behavioral medicine , pages=

    Just-in-time adaptive interventions (JITAIs) in mobile health: key components and design principles for ongoing health behavior support , author=. Annals of behavioral medicine , pages=. 2016 , publisher=

    2016

  20. [20]

    JMIR mHealth and uHealth , volume=

    Development and validation of multivariable prediction algorithms to estimate future walking behavior in adults: retrospective cohort study , author=. JMIR mHealth and uHealth , volume=. 2023 , publisher=

    2023

  21. [21]

    JMIR Research Protocols , volume=

    Advancing understanding of just-in-time states for supporting physical activity (Project JustWalk JITAI): protocol for a system ID study of just-in-time adaptive interventions , author=. JMIR Research Protocols , volume=. 2023 , publisher=

    2023

  22. [22]

    European Heart Journal-Digital Health , volume=

    Personalized digital behaviour interventions increase short-term physical activity: a randomized control crossover trial substudy of the MyHeart Counts Cardiovascular Health Study , author=. European Heart Journal-Digital Health , volume=. 2023 , publisher=

    2023

  23. [23]

    Neuropsychopharmacology , volume=

    Harnessing smartphone-based digital phenotyping to enhance behavioral and mental health , author=. Neuropsychopharmacology , volume=. 2016 , publisher=

    2016

  24. [24]

    medRxiv , year=

    11 million days of longitudinal wearable data reveal novel future health insights , author=. medRxiv , year=

  25. [25]

    arXiv preprint arXiv:2407.07874 , year=

    Toto: Time series optimized transformer for observability , author=. arXiv preprint arXiv:2407.07874 , year=

    Pith/arXiv arXiv

  26. [26]

    npj digital Medicine , volume=

    Verification, analytical validation, and clinical validation (V3): the foundation of determining fit-for-purpose for Biometric Monitoring Technologies (BioMeTs) , author=. npj digital Medicine , volume=. 2020 , publisher=

    2020

  27. [27]

    PLOS Digital Health , volume=

    Wearables research for continuous monitoring of patient outcomes: A scoping review , author=. PLOS Digital Health , volume=. 2025 , publisher=

    2025

  28. [28]

    Nature Medicine , volume=

    A personal health large language model for sleep and fitness coaching , author=. Nature Medicine , volume=. 2025 , publisher=

    2025

  29. [29]

    Mantena and Alexander Tolas and Samuel Montalvo and Mariana Ramirez-Posada and Jack W

    Paul Schmiedmayer and Anders Johnson and Narayan Schuetz and Lukas Kollmer and Paul Goldschmidt and Juan Delgado-SanMartin and Kelly W Zhang and Sriya D. Mantena and Alexander Tolas and Samuel Montalvo and Mariana Ramirez-Posada and Jack W. O'Sullivan and Marily Oppezzo and Abby C King and Fatima Rodriguez and Euan Ashley and Allan Lawrie and Daniel Seung...

    2026

  30. [30]

    Bloom: Designing for LLM-Augmented Behavior Change Interactions , year =

    J. Bloom: Designing for LLM-Augmented Behavior Change Interactions , year =. Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26) , date-modified =. doi:10.1145/3772318.3790506 , location =

    arXiv 2026

  31. [31]

    Annals of Behavioral Medicine , volume=

    Efficacy of contextually tailored suggestions for physical activity: a micro-randomized optimization trial of HeartSteps , author=. Annals of Behavioral Medicine , volume=. 2019 , publisher=

    2019

  32. [32]

    Proceedings of the ACM on interactive, mobile, wearable and ubiquitous technologies , volume=

    Personalized heartsteps: A reinforcement learning algorithm for optimizing physical activity , author=. Proceedings of the ACM on interactive, mobile, wearable and ubiquitous technologies , volume=. 2020 , publisher=

    2020

  33. [33]

    JMIR Research Protocols , volume=

    Just-in-Time Adaptive Intervention to Promote Walking Behavior and Reduce Stationary Time in Physically Inactive Adults: Protocol for the Walking With JITAIs Study , author=. JMIR Research Protocols , volume=. 2026 , publisher=

    2026

  34. [34]

    2025 , type =

    Worldwide Quarterly Wearable Device Tracker: Q1 2025 Release , institution =. 2025 , type =

    2025

  35. [35]

    2026 , note =

    Tajammul Pangarkar , title =. 2026 , note =

    2026

  36. [36]

    PloS one , volume=

    Large scale population assessment of physical activity using wrist worn accelerometers: the UK biobank study , author=. PloS one , volume=. 2017 , publisher=

    2017

  37. [37]

    npj Digital Medicine , volume=

    Understanding activity and physiology at scale: the Apple Heart & Movement Study , author=. npj Digital Medicine , volume=. 2024 , publisher=

    2024

  38. [38]

    2024 , note =

    Resources for Using. 2024 , note =

    2024

  39. [39]

    Journal of medical Internet research , volume=

    Analysis of physical activity using wearable health technology in US adults enrolled in the all of US research program: multiyear observational study , author=. Journal of medical Internet research , volume=. 2024 , publisher=

    2024

  40. [40]

    Scientific Reports , volume=

    Large-scale assessment of physical activity in a population using high-resolution hip-worn accelerometry: the German National Cohort (NAKO) , author=. Scientific Reports , volume=. 2024 , publisher=

    2024

  41. [41]

    International Journal of Behavioral Nutrition and Physical Activity , volume=

    Device-measured movement behaviours in over 20,000 China Kadoorie Biobank participants , author=. International Journal of Behavioral Nutrition and Physical Activity , volume=. 2023 , publisher=

    2023

  42. [42]

    American heart journal , volume=

    Evaluation of Huawei smart wearables for detection of atrial fibrillation in patients following ischemic stroke: The Liverpool-Huawei stroke study , author=. American heart journal , volume=. 2023 , publisher=

    2023

  43. [43]

    American Heart Journal , volume=

    Rationale and design of a large population study to validate software for the assessment of atrial fibrillation from data acquired by a consumer tracker or smartwatch: the Fitbit heart study , author=. American Heart Journal , volume=. 2021 , publisher=

    2021

  44. [44]

    Scientific Data , year=

    A digital biomarker dataset from hematopoietic cell transplant caregivers and patients , author=. Scientific Data , year=

  45. [45]

    Medicine and science in sports and exercise , pages=

    Fitbit physical activity and sleep data in the all of us research program: data exploration and processing considerations for research , author=. Medicine and science in sports and exercise , pages=

  46. [46]

    PhysioNet , year=

    Multilevel monitoring of activity and sleep in healthy people , author=. PhysioNet , year=

  47. [47]

    , author=

    Lora: Low-rank adaptation of large language models. , author=. Iclr , volume=

  48. [48]

    Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies , volume=

    Globem: Cross-dataset generalization of longitudinal human behavior modeling , author=. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies , volume=. 2023 , publisher=

    2023

  49. [49]

    Scandinavian Journal of Medicine & Science in Sports , volume=

    Daily and peak monitor independent movement summary (MIMS) values associated with metabolic syndrome: NHANES 2011--12 and 2013--14 , author=. Scandinavian Journal of Medicine & Science in Sports , volume=. 2024 , publisher=

    2011

  50. [50]

    Journal of the American College of Cardiology , volume=

    Mobile photoplethysmographic technology to detect atrial fibrillation , author=. Journal of the American College of Cardiology , volume=. 2019 , publisher=

    2019

  51. [51]

    The Thirteenth International Conference on Learning Representations , year=

    Scaling Wearable Foundation Models , author=. The Thirteenth International Conference on Learning Representations , year=

  52. [52]

    arXiv preprint arXiv:2506.05321 , year=

    Lsm-2: Learning from incomplete wearable sensor data , author=. arXiv preprint arXiv:2506.05321 , year=

    Pith/arXiv arXiv

  53. [53]

    Journal of medical Internet research , volume=

    Challenges in participant engagement and retention using mobile health apps: literature review , author=. Journal of medical Internet research , volume=. 2022 , publisher=

    2022

  54. [54]

    Forty-second International Conference on Machine Learning , year=

    Beyond Sensor Data: Foundation Models of Behavioral Data from Wearables Improve Health Predictions , author=. Forty-second International Conference on Machine Learning , year=

  55. [55]

    PLoS medicine , volume=

    The rise of consumer health wearables: promises and barriers , author=. PLoS medicine , volume=. 2016 , publisher=

    2016

  56. [56]

    arXiv preprint arXiv:2203.15556 , volume=

    Training compute-optimal large language models , author=. arXiv preprint arXiv:2203.15556 , volume=

    Pith/arXiv arXiv

  57. [57]

    Journal of Machine Learning Research , volume =

    Random Search for Hyper-Parameter Optimization , author =. Journal of Machine Learning Research , volume =. 2012 , url =

    2012

  58. [58]

    Journal of Machine Learning Research , volume =

    Hyperband: A Novel Bandit-Based Approach to Hyperparameter Optimization , author =. Journal of Machine Learning Research , volume =. 2018 , url =

    2018

  59. [59]

    PLOS ONE , volume =

    The Precision-Recall Plot Is More Informative than the ROC Plot When Evaluating Binary Classifiers on Imbalanced Datasets , author =. PLOS ONE , volume =. 2015 , doi =

    2015

  60. [60]

    Monthly Weather Review , volume =

    Skill Scores Based on the Mean Square Error and Their Relationships to the Correlation Coefficient , author =. Monthly Weather Review , volume =. 1988 , doi =

    1988

  61. [61]

    Proceedings of the 35th Uncertainty in Artificial Intelligence Conference , series =

    A Flexible Framework for Multi-Objective Bayesian Optimization using Random Scalarizations , author =. Proceedings of the 35th Uncertainty in Artificial Intelligence Conference , series =. 2020 , publisher =

    2020

  62. [62]

    Proceedings of the 20th International Conference on Artificial Intelligence and Statistics , series =

    Fast Bayesian Optimization of Machine Learning Hyperparameters on Large Datasets , author =. Proceedings of the 20th International Conference on Artificial Intelligence and Statistics , series =. 2017 , publisher =

    2017

  63. [63]

    2018 , publisher =

    Falkner, Stefan and Klein, Aaron and Hutter, Frank , booktitle =. 2018 , publisher =

    2018

  64. [64]

    The Thirteenth International Conference on Learning Representations , year=

    PaPaGei: Open Foundation Models for Optical Physiological Signals , author=. The Thirteenth International Conference on Learning Representations , year=

  65. [65]

    Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies , volume=

    Pulse-ppg: An open-source field-trained ppg foundation model for wearable applications across lab and field settings , author=. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies , volume=. 2025 , publisher=

    2025

  66. [66]

    The Twelfth International Conference on Learning Representations , year=

    Large-scale Training of Foundation Models for Wearable Biosignals , author=. The Twelfth International Conference on Learning Representations , year=

  67. [67]

    Proceedings of the 19th ACM Conference on Embedded Networked Sensor Systems , pages=

    Limu-bert: Unleashing the potential of unlabeled data for imu sensing applications , author=. Proceedings of the 19th ACM Conference on Embedded Networked Sensor Systems , pages=

  68. [68]

    Nature , pages=

    Insulin resistance prediction from wearables and routine blood biomarkers , author=. Nature , pages=. 2026 , publisher=

    2026

  69. [69]

    NPJ digital medicine , volume=

    Self-supervised learning for human activity recognition using 700,000 person-days of wearable data , author=. NPJ digital medicine , volume=. 2024 , publisher=

    2024

  70. [70]

    arXiv preprint arXiv:2412.11276 , year=

    Wearable accelerometer foundation models for health via knowledge distillation , author=. arXiv preprint arXiv:2412.11276 , year=

    Pith/arXiv arXiv

  71. [71]

    The Thirteenth International Conference on Learning Representations , year=

    RelCon: Relative Contrastive Learning for a Motion Foundation Model for Wearable Data , author=. The Thirteenth International Conference on Learning Representations , year=

  72. [72]

    , date =

    Tan, Chang Wei and Dempster, Angus and Bergmeir, Christoph and Webb, Geoffrey I. , date =. 2022 , issn =. doi:10.1007/s10618-022-00844-1 , abstract =

    doi:10.1007/s10618-022-00844-1 2022

  73. [73]

    , title =

    Snoek, Jasper and Larochelle, Hugo and Adams, Ryan P. , title =. Proceedings of the 26th International Conference on Neural Information Processing Systems - Volume 2 , pages =. 2012 , publisher =

    2012

  74. [74]

    2025 , url=

    Erik Xie and Wyatt Chang and Raquel Rodriguez Martinez and Brandon Ballinger , booktitle=. 2025 , url=

    2025

  75. [75]

    Assessing the feasibility of using smartphone data to identify risk of idiopathic pulmonary arterial hypertension , volume =

    Delgado-. Assessing the feasibility of using smartphone data to identify risk of idiopathic pulmonary arterial hypertension , volume =. doi:10.1038/s44325-026-00114-9 , abstract =

    work page doi:10.1038/s44325-026-00114-9

  76. [76]

    The Impact of Wearable Technologies in Health Research: Scoping Review

    Huhn, Sophie and Axt, Miriam and Gunga, Hanns-Christian and Maggioni, Martina Anna and Munga, Stephen and Obor, David and Si \'e , Ali and Boudo, Valentin and Bunker, Aditi and Sauerborn, Rainer and B \"a rnighausen, Till and Barteit, Sandra. The Impact of Wearable Technologies in Health Research: Scoping Review. JMIR Mhealth Uhealth. 2022. doi:10.2196/34384

    doi:10.2196/34384 2022

  77. [77]

    A Simple Approach to Ordinal Classification

    Frank, Eibe and Hall, Mark. A Simple Approach to Ordinal Classification. Machine Learning: ECML 2001. 2001

    2001

  78. [78]

    Systematic review and meta-analysis of performance of wearable artificial intelligence in detecting and predicting depression , volume =

    Abd-Alrazaq, Alaa and. Systematic review and meta-analysis of performance of wearable artificial intelligence in detecting and predicting depression , volume =. doi:10.1038/s41746-023-00828-5 , pages =

    work page doi:10.1038/s41746-023-00828-5

  79. [79]

    and Solaiman, Barry and Househ, Mowafa , year =

    Ahmed, Arfan and Aziz, Sarah and Alzubaidi, Mahmood and Schneider, Jens and Irshaidat, Sara and Serhan, Hashem Abu and Abd-alrazaq, Alaa A. and Solaiman, Barry and Househ, Mowafa , year =. Wearable devices for anxiety & depression: A scoping review , volume =. doi:https://doi.org/10.1016/j.cmpbup.2023.100095 , abstract =

    arXiv 2023

  80. [80]

    Sleep , year=

    Sleep stage prediction with raw acceleration and photoplethysmography heart rate data derived from a consumer wearable device , author=. Sleep , year=

Showing first 80 references.