Pith. sign in

REVIEW 4 major objections 5 minor 61 references

Inertia-1: An Open Exploration of Wearable Motion Foundation Models

T0 review · 4 major / 5 minor · reviewed 2026-07-11 · grok-4.5

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

Pith's one-line read Wearable motion foundation models work when data fidelity, sensing setup, objective, and scale are chosen together—not by any single recipe.

desk verdict Solid open systems paper: matched SSL bake-off plus sensing/scale ablations on 15 datasets; disease “recipes” rest on medication/survey proxies and wrist-centric pretraining, so treat clinical transfer as provisional. read the letter →

arxiv 2607.06617 v1 pith:SCASNV3M submitted 2026-07-07 cs.LG cs.AI

Zongzhe Xu , Aakarsh Anand , Sarah Jiang , Chuntung Zhuang , Zitao Shuai , Sriram Sankararaman , Yuzhe Yang This is my paper
classification cs.LGcs.AI
keywords wearablemotionfoundationmodelsself-supervisedlearningaccelerometryhumanactivityrecognitionfreezingofgaitdiseasepredictionsensingconfigurations
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that wearable accelerometer streams are a natural foundation-model substrate for behavior and health, but the field has been too fragmented to know which design choices actually transfer. The authors build a controlled open testbed—Inertia-1—over more than 18 million hours of motion data and more than a thousand trained models, systematically varying sensor modality, body placement, sampling rate, window length, architecture, model size, pretraining objective, and data scale. Across fifteen datasets covering short activity recognition, freezing-of-gait detection, and longer-horizon disease prediction, self-supervised pretraining consistently beats training from scratch, yet no single objective wins everywhere. What matters more is keeping full triaxial signals, matching temporal resolution and window length to task grain, scaling diverse unlabeled data rather than model size alone, and fusing complementary sensors or placements. The result is both practical recipes and an open cookbook for building motion representations that generalize across tasks and sensing conditions.

What carries the argument

Inertia-1: a controlled full-lifecycle framework that holds training and evaluation protocol fixed while sweeping data axes (modality, placement, rate, window, domain), model axes (architecture and size), and training axes (ten representative pretraining objectives and data scale) across 15 datasets and 18.2M hours of motion.

What would settle it

Hold the same models and protocol fixed but retrain and re-evaluate using high-frequency clinical triaxial cohorts with clinician-confirmed disease labels (not medication proxies) and non-wrist primary placements; if the reported ranking of objectives, the triaxial advantage, and the data-scale-over-model-size pattern reverse or collapse, the central claim fails.

Watch

Extended reading notes

Core claim

Self-supervised pretraining on large unlabeled accelerometer cohorts consistently improves transfer over supervised-from-scratch baselines across human activity recognition, freezing-of-gait detection, and disease prediction, but no single objective dominates; transferable performance depends on the joint choice of data fidelity (triaxial over magnitude summaries), sensing setup, objective, and data diversity more than on increasing model size alone.

Load-bearing premise

That wrist-centric pretraining on large cohort accelerometry (with fixed windows, rates, and survey- or medication-derived disease labels) fairly isolates design choices and represents the diversity of real deployment without major domain shift or label artifacts.

Editorial extensions

If this is right

  • Default to triaxial waveforms rather than vector-magnitude summaries when training general-purpose motion representations.
  • Prefer expanding diverse unlabeled pretraining data over simply enlarging encoder capacity under a fixed corpus.
  • Match sampling rate and window length to task timescale: short windows and modest rates can suffice for many activity labels, while disease signals often need higher temporal fidelity.
  • Wrist-accelerometer pretraining can seed transfer to other placements and modalities, and multi-stream fusion further improves cluster separation and accuracy.
  • Practitioners can use the open recipes and grid rather than re-discovering sensing and objective choices in isolation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If disease labels built from medications and surveys are noisy, the large reported SSL gains on disease tasks may partly reflect correlation with lifestyle confounders rather than pure biomechanical signatures.
  • The finding that model size saturates under fixed data suggests wearable motion may currently be data-diversity-limited more than architecture-limited, similar to other biosignal foundation efforts.
  • Extending the same controlled grid to continuous regression targets (energy expenditure, step quality, longitudinal trajectories) would test whether the cookbook generalizes beyond classification.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. Inertia-1 presents a large-scale, open empirical study of wearable motion foundation models. Using ~18.2M hours of accelerometer data (primarily NHANES, with UK Biobank for scaling) and a controlled grid over sensing choices (sampling rate, window length, axes, domain, placement, modality), model size, and 10 pretraining objectives, the authors evaluate transfer on 15 datasets spanning HAR, freezing-of-gait, and disease prediction. The central claims are that self-supervised pretraining consistently beats supervised-from-scratch baselines, that no single objective dominates, that data fidelity and diversity matter more than model size alone, and that coordinated design yields practical “recipes,” with code and a public cookbook released.

Significance. If the comparative findings hold under the stated protocols, this is a high-value contribution to wearable representation learning: it replaces fragmented single-axis studies with a unified testbed, reports extensive linear-probe and full-finetune metrics across many datasets, and releases open code and models. The controlled one-factor ablations (Figs. 5–6, 12), multi-stream fusion analysis (§4.4), and data-vs-parameter scaling sweeps (Figs. 3–4, 9–11) are the right experimental tools for the field. The open cookbook framing and breadth of coverage (Table 1, Table 7) are genuine strengths even if some clinical interpretations need tightening.

major comments (4)
  1. [Appendix A.4; Tables 5–6; §4.1–4.2] Appendix A.4 and Tables 5–6 / Fig. 4: Disease labels for NHANES Parkinson’s, depression, and diabetes are inferred from medications (anti-Parkinson’s agents, antidepressants, insulin) and survey reports, not clinical diagnoses. Table 6 treats disease prediction as the task family with the largest SSL gain (19.7 AUROC). Medication/survey proxies can encode treatment access, comorbidity, and reporting bias rather than motion-linked disease biology. The DP results remain useful as population-health transfer probes, but the manuscript currently presents them as “disease prediction” with clinical-sounding recipes. Please reframe DP claims, add an explicit limitations subsection on proxy validity, and (ideally) a sensitivity check or secondary label definition so the largest reported gains are not over-interpreted as clinical transfer.
  2. [§3.2; Table 7; Tables 3–5] §3.2 lists LSM among masked-reconstruction methods, and Table 7 includes LSM under pretraining objectives, but LSM does not appear in the main objective comparison tables (Tables 3–5, 10–14). Either report LSM under the same protocol or remove it from the claimed method suite so the “10 approaches” claim matches the evidence presented.
  3. [Abstract; §1; Tables 3–4] Abstract and §1 claim “state-of-the-art recipes” and broad generalization “across tasks and sensing conditions.” The evidence is strong for internal comparisons (SSL vs supervised ViT/CNN under a shared protocol; one-axis ablations). It is weaker for external SOTA: published per-dataset numbers from prior HAR/FoG work are not systematically tabulated against Inertia-1. Please either (i) add a compact external comparison table on a subset of standard benchmarks, or (ii) soften “state-of-the-art” language to “strong recipes under a unified protocol,” which the current experiments support.
  4. [§3.1–3.3; §4.2; Appendix A.2, D.3] §3.1–3.3 and §4.2: The default pretraining regime is wrist triaxial NHANES at 20 Hz / 30 s, while UK Biobank scaling uses 0.2 Hz Euclidean-norm (magnitude) 2 h windows (Appendix A.2, D.3). Multi-placement/modality transfer is studied in §4.4 and is a positive result, but the paper’s “full lifecycle” and “real-world sensing diversity” framing still rests heavily on wrist-ACC pretraining plus lower-fidelity cohort scale. Clarify which conclusions are conditional on wrist-centric high-resolution pretraining versus which survive magnitude/low-rate pretraining, so the cookbook does not over-generalize from the default axis.
minor comments (5)
  1. [§3; §4.4] In-text table/figure references are inconsistent (e.g., “figure 3 and Table 3.1” in §3 while the overview table is Table 2; “figure 7” vs “Fig.” elsewhere). Normalize numbering and capitalization.
  2. [Table 1; Appendix A.2] Table 1 footnote says “Only raw-waveform studies considered,” but UK Biobank scaling uses precomputed Euclidean-norm epochs. Make the comparison criteria explicit so readers do not misread coverage claims.
  3. [Appendix B; Table 13; §3.3; Appendix E] Several typos: “genertic” (Appendix B.1), “reesults” (Table 13 caption), “with with” (§3.3), “cleaniness” (Appendix E.1). A proofreading pass would help.
  4. [Abstract; §1; Table 8] Disease prediction is described as 7 tasks in places and 5+2 elsewhere; align the count in the abstract, §1, and Table 8.
  5. [§4; Appendix D] Error bars / multi-seed reporting are noted for disease tasks but not uniformly for HAR/FoG ablations. Even a short note on run-to-run variance for key ablations would strengthen confidence in small AUROC gaps.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: Inertia-1 reports comparative empirical measurements under shared protocols, not predictions forced by definition or self-citation.

full rationale

Inertia-1 is an empirical systems/benchmark paper. Its load-bearing claims (SSL pretraining beats supervised-from-scratch on HAR/FoG/DP; no single objective dominates; triaxial/data diversity matter more than model size; multi-stream fusion helps) are established by training >1,000 models and reporting AUROC/AUPRC/F1 on held-out subject splits across 15 public datasets (Tables 3–6, 10–16; Figs. 3–8). Those metrics are not algebraically or statistically forced by the pretraining loss, the NHANES/UKB corpus construction, or any fitted constant renamed as a prediction. Supervised baselines and alternate SSL objectives are trained under the same input/evaluation protocol, so relative gains are comparative measurements, not tautologies. Self-citations (e.g., RelCon, LSM, related biosignal FMs) appear as related work or method implementations, not as uniqueness theorems that forbid alternatives or as the sole support for the transfer findings. Disease labels inferred from medications/surveys (Appendix A.4) and wrist-centric defaults raise external validity/correctness questions, but they do not make the reported rankings circular by construction. Score 0 is appropriate: the derivation chain is experimental comparison against external benchmarks, not a closed definitional loop.

Assumptions & free parameters 7 free parameters · 5 assumptions · 1 invented entities

This is an empirical systems paper. Load-bearing content is experimental protocol and data construction rather than mathematical axioms. Free parameters are the many training and sensing defaults that define the ‘controlled’ grid. Domain assumptions include standard SSL transfer protocols and proxy disease labels. The main invented entity is the Inertia-1 framework/cookbook itself; it is operationalized by code and datasets rather than postulated physics.

free parameters (7)
  • Default window length (30 s) and sampling rate (20 Hz)
    Chosen as the default controlled setting for objective comparisons; other values are ablated but rankings and ‘recipes’ are anchored here (§4.1, Appendix A.4).
  • Patch length / stride (10) and mask ratio (0.4 for PatchTST)
    Architecture hyperparameters fixed for ART/PatchTST and related models; affect representation granularity (Appendix B.1).
  • Optimizer and schedule (AdamW lr 2e-4, weight decay 0.04, 170k steps / ~5 epochs, batch 1024)
    Shared training recipe that can favor some objectives over others if not equally tuned (Appendix B).
  • Model size presets (~5–8M / ~30M / ~100M; d/L/H/dff grids)
    Capacity definitions used to claim model scaling saturates under fixed data (Table 9, §4.2).
  • MIL bag construction (1024 windows / day for NHANES; 12×2h windows for UKB disease)
    Subject-level disease performance depends on this pooling design and sampling of times of day (Appendix A.4, D.3).
  • UK Biobank disease case-control ratio (~1:5) and 0.2 Hz / 2 h scaling protocol
    Scaling curves for osteoporosis/osteoarthritis are measured under this constructed evaluation protocol (Appendix D.3).
  • Augmentation strengths for DINO/SimCLR (jitter, scale, time mask, channel dropout)
    Generic augmentations adapted by hand; can under-serve motion-specific structure relative to domain methods (Appendix B.1).
assumptions (5)
  • domain assumption Self-supervised representations pretrained on large unlabeled free-living accelerometry transfer to labeled HAR, FoG, and disease tasks under linear probing / full finetuning.
    Core evaluation premise of §4; standard in SSL but not guaranteed across clinical labels and placements.
  • domain assumption NHANES medication/survey proxies (antidepressants, insulin, anti-Parkinson agents, etc.) are valid enough disease labels for ranking motion representations.
    Appendix A.4 constructs Parkinson’s, depression, diabetes labels this way; errors would reorder disease results.
  • domain assumption Varying one sensing axis at a time while holding others fixed isolates causal effects of sampling rate, window, axes, and domain.
    Stated design of §4.3; interactions among axes may remain.
  • domain assumption Standard SSL objectives (masked reconstruction, contrastive, autoregressive, DINO) and ResNet/Transformer/ViT backbones are adequate representatives of the method space.
    §3.2 selects 10 approaches as covering SOTA classes; omitted methods could change ‘no single objective dominates’.
  • standard math Subject-level train/val/test splits prevent leakage and support generalization claims.
    Appendix A.3; standard ML practice assumed correctly implemented.
invented entities (1)
  • Inertia-1 controlled framework / open cookbook independent evidence
    purpose: Unify data, model, and training axes for wearable motion foundation models and publish recipes plus code.
    The paper’s primary contribution is this integrated experimental object rather than a new physical entity; independent evidence is the public code/site and multi-dataset results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Inertia-1: An Open Exploration of Wearable Motion Foundation Models." pith.science (2026). https://pith.science/paper/SCASNV3M

@misc{pith2026260706617,
  author       = {Pith},
  title        = {Pith review of: Inertia-1: An Open Exploration of Wearable Motion Foundation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SCASNV3M}},
  note         = {Machine review of arXiv:2607.06617}
}
read the original abstract

Wearable motion sensing provides a continuous and scalable window into human behavior and health, making it a natural fit for foundation models, yet its pretraining and scaling principles remain poorly understood. Prior work studies isolated design choices, such as sensor placement or sampling frequency, often under fixed settings and narrow downstream tasks that fail to capture real-world sensing diversity. We introduce Inertia-1, a fully open exploration of wearable motion foundation models. Using massive corpora of accelerometer data from global sources spanning more than 18.2M hours, we build a controlled framework for studying the full lifecycle of wearable motion foundation models, covering data choices such as sensor modality, device placement, sampling rate, window length; model choices such as architectures and model size; and training choices such as pretraining objective and data scale. Extensive evaluations across 15 datasets spanning human activity recognition, freezing-of-gait detection, and disease prediction reveal intriguing findings for building motion foundation models that generalize across tasks and sensing conditions. Collectively, Inertia-1 not only presents state-of-the-art recipes for diverse downstream tasks, but also serves as a comprehensive, practical, and open cookbook for wearable motion representation learning.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 61 canonical work pages

  1. [1]

    Miller, Saba Emrani, Udhyakumar Nal- lasamy, and Ian Shapiro

    Salar Abbaspourazad, Oussama Elachqar, Andrew C. Miller, Saba Emrani, Udhyakumar Nal- lasamy, and Ian Shapiro. Large-scale training of foundation models for wearable biosignals, 2024

    work page 2024

  2. [2]

    Adaptnet: Human activity recognition via bilateral domain adaptation using semi-supervised deep translation networks.IEEE Sensors Journal, 21(18):20398–20411, 2021

    Sungtae An, Alessio Medda, Michael N Sawka, Clayton J Hutto, Mindy L Millard-Stafford, Scott Appling, Kristine LS Richardson, and Omer T Inan. Adaptnet: Human activity recognition via bilateral domain adaptation using semi-supervised deep translation networks.IEEE Sensors Journal, 21(18):20398–20411, 2021

    work page 2021

  3. [3]

    Window size impact in human activity recognition.Sensors, 14(4):6474–6499, 2014

    Oresti Banos, Juan-Manuel Galvez, Miguel Damas, Hector Pomares, and Ignacio Rojas. Window size impact in human activity recognition.Sensors, 14(4):6474–6499, 2014

    work page 2014

  4. [4]

    Oresti Banos, Rafael Garcia, and Alejandro Saez. MHEALTH. UCI Machine Learning Repository,

    work page

  5. [5]

    DOI: https://doi.org/10.24432/C5TW22

    work page doi:10.24432/c5tw22

  6. [6]

    Akos Toth, Miguel Damas, Hector Pomares, and Ignacio Rojas

    Oresti Banos, M. Akos Toth, Miguel Damas, Hector Pomares, and Ignacio Rojas. Dealing with the effects of sensor displacement in wearable activity recognition.Sensors, 14(6):9995–10023, 2014

    work page 2014

  7. [7]

    Heterogeneity Activity Recognition

    Henrik Blunck, Sourav Bhattacharya, Thor Prentow, Mikkel Kjrgaard, and Anind Dey. Heterogeneity Activity Recognition. UCI Machine Learning Repository, 2015. DOI: https://doi.org/10.24432/C5689X

    work page doi:10.24432/c5689x 2015

  8. [8]

    Wear: An outdoor sports dataset for wearable and egocentric activity recognition, 2024

    Marius Bock, Hilde Kuehne, Kristof Van Laerhoven, and Michael Moeller. Wear: An outdoor sports dataset for wearable and egocentric activity recognition, 2024

    work page 2024

Show all 61 references
  1. [9]

    Emerging properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9650–9660, 2021

    2021

  2. [10]

    Domain adaptation for inertial measurement unit-based human activity recognition: A survey.arXiv preprint arXiv:2304.06489, 2023

    Avijoy Chakma, Abu Zaher Md Faridee, Indrajeet Ghosh, and Nirmalya Roy. Domain adaptation for inertial measurement unit-based human activity recognition: A survey.arXiv preprint arXiv:2304.06489, 2023

    2023 arXiv

  3. [11]

    Capture-24: A large dataset of wrist-worn activity tracker data collected in the wild for human activity recognition.Scientific Data, 11(1):1135, 2024

    Shing Chan, Yuan Hang, Catherine Tong, Aidan Acquah, Abram Schonfeldt, Jonathan Gershuny, and Aiden Doherty. Capture-24: A large dataset of wrist-worn activity tracker data collected in the wild for human activity recognition.Scientific Data, 11(1):1135, 2024

    2024

  4. [12]

    Capture-24: Activity tracker dataset for human activity recognition, 2021

    Chan Chang, S Walmsley, R Gershuny, J Harms, T Thomas, E Milton, K Kelly, P Foster, C Wong, A Gray, N Haque, S Hollowell, and S Doherty. Capture-24: Activity tracker dataset for human activity recognition, 2021

    2021

  5. [13]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. InProceedings of the 37th International Conference on Machine Learning (ICML), pages 1597–1607. PMLR, 2020. 12 Inertia-1: An Open Explorati...

    2020

  6. [14]

    China kadoorie biobank of 0.5 million people: survey methods, baseline characteristics and long-term follow-up.International journal of epidemiology, 40(6):1652–1666, 2011

    Zhengming Chen, Junshi Chen, Rory Collins, Yu Guo, Richard Peto, Fan Wu, and Liming Li. China kadoorie biobank of 0.5 million people: survey methods, baseline characteristics and long-term follow-up.International journal of epidemiology, 40(6):1652–1666, 2011

    2011

  7. [15]

    Empirical evaluation of gated recurrent neural networks on sequence modeling, 2014

    Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling, 2014

    2014

  8. [16]

    Anis Davoudi, Mamoun T Mardini, David Nelson, Fahd Albinali, Sanjay Ranka, Parisa Rashidi, and Todd M Manini. The effect of sensor placement and number on physical activity recognition and energy expenditure estimation in older adults: Validation study.JMIR Mhealth Uhealth, 9(...

    2021

  9. [17]

    Human activity recognition using inertial, physiological and environmental sensors: A comprehensive survey

    Florenc Demrozi, Graziano Pravadelli, Azra Bihorac, and Parisa Rashidi. Human activity recognition using inertial, physiological and environmental sensors: A comprehensive survey. IEEE Access, 8:210816–210836, 2020

    2020

  10. [18]

    Large scale population assessment of physical activity using wrist worn accelerometers: the uk biobank study.PloS one, 12(2):e0169649, 2017

    Aiden Doherty, Dan Jackson, Nils Hammerla, Thomas Plötz, Patrick Olivier, Malcolm H Granat, Tom White, Vincent T Van Hees, Michael I Trenell, Christoper G Owen, et al. Large scale population assessment of physical activity using wrist worn accelerometers: the uk biobank study....

    2017

  11. [19]

    Grady, Irfan Essa, Judy Hoffman, and Thomas Plötz

    Harish Haresamudram, Apoorva Beedu, Varun Agrawal, Patrick L. Grady, Irfan Essa, Judy Hoffman, and Thomas Plötz. Masked reconstruction based self-supervision for human activity recognition. InProceedings of the 2020 ACM International Symposium on Wearable Computers, ISWC ’20, ...

    2020

  12. [20]

    Assessing the state of self-supervised human activity recognition using wearables.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 6(3), 2022

    Harish Haresamudram, Irfan Essa, and Thomas Plotz. Assessing the state of self-supervised human activity recognition using wearables.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 6(3), 2022

    2022

  13. [21]

    Scaling laws in wearable human activity recognition.arXiv preprint arXiv:2502.03364, 2025

    Tom Hoddes, Alex Bijamov, Saket Joshi, Daniel Roggen, Ali Etemad, Robert Harle, and David Racz. Scaling laws in wearable human activity recognition.arXiv preprint arXiv:2502.03364, 2025

    2025 arXiv

  14. [22]

    Huang, Kebin Yan, and Jukka-Pekka Onnela

    Emily J. Huang, Kebin Yan, and Jukka-Pekka Onnela. Smartphone-based activity recognition using multistream movelets combining accelerometer and gyroscope data.Sensors, 22(7):2618, 2022

    2022

  15. [23]

    Tinghuai Huang, Meng Li, and Jianwei Huang. Recent trends in wearable device used to detect freezing of gait and falls in people with parkinson’s disease: A systematic review.Frontiers in Aging Neuroscience, Volume 15 - 2023, 2023

    2023

  16. [24]

    Lara and Miguel A

    Oscar D. Lara and Miguel A. Labrador. A survey on human activity recognition using wearable sensors.IEEE Communications Surveys & Tutorials, 15(3):1192–1209, 2013

    2013

  17. [25]

    Aguirre, Valdery Moura Junior, Jiarui Jin, Che Liu, Lanhai Zhong, Chenxi Sun, Gari Clifford, M

    Jun Li, Aaron D. Aguirre, Valdery Moura Junior, Jiarui Jin, Che Liu, Lanhai Zhong, Chenxi Sun, Gari Clifford, M. Brandon Westover, and Shenda Hong. An electrocardiogram foundation model built on over 10 million recordings.NEJM AI, 2(7):AIoa2401033, 2025

    2025

  18. [26]

    HEARTS: Benchmarking llm reasoning on health time series.arXiv preprint arXiv:2603.06638, 2026

    Sirui Li, Shuhan Xiao, Mihir Joshi, Ahmed Metwally, Daniel McDuff, Wei Wang, and Yuzhe Yang. HEARTS: Benchmarking llm reasoning on health time series.arXiv preprint arXiv:2603.06638, 2026. 13 Inertia-1: An Open Exploration of Wearable Motion Foundation Models

    2026 arXiv

  19. [27]

    Selfpab: Large-scale pre- training on accelerometer data for human activity recognition.Applied Intelligence, 54(6):4545– 4563, 2024

    Aleksej Logacjov, Sverre Herland, Astrid Ustad, and Kerstin Bach. Selfpab: Large-scale pre- training on accelerometer data for human activity recognition.Applied Intelligence, 54(6):4545– 4563, 2024

    2024

  20. [28]

    Aleksej Logacjov, Atle Kongsvold, Kerstin Bach, Hilde Bremseth Bårdstu, and Paul Jarle Mork. HARTH. UCI Machine Learning Repository. DOI: https://doi.org/10.24432/C5NC90

    doi

  21. [29]

    Aleksej Logacjov and Astrid Ustad. HAR70+. UCI Machine Learning Repository. DOI: https://doi.org/10.24432/C5CW3D

    doi

  22. [30]

    Beckman, Francis Ratsimbazafy, Kayla Marginean, Robert Carroll, Karthik Natarajan, Frank E

    Hiral Master, Jeffrey Annis, Shi Huang, Joshua A. Beckman, Francis Ratsimbazafy, Kayla Marginean, Robert Carroll, Karthik Natarajan, Frank E. Harrell, Dan M. Roden, Paul Harris, and Evan L. Brittain. Association of step counts over time with the risk of chronic disease in the ...

    2022

  23. [31]

    Insulin resistance prediction from wearables and routine blood biomarkers.Nature, pages 1–11, 2026

    Ahmed A Metwally, A Ali Heydari, Daniel McDuff, Alexandru Solot, Zeinab Esmaeilpour, An- thony Z Faranesh, Menglian Zhou, Girish Narayanswamy, Maxwell A Xu, Xin Liu, et al. Insulin resistance prediction from wearables and routine blood biomarkers.Nature, pages 1–11, 2026

    2026

  24. [32]

    Shenghuan Miao, Ling Chen, and Rong Hu. Spatial-temporal masked autoencoder for multi- device wearable human activity recognition.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 7(4):1–25, 2024

    2024

  25. [33]

    Scott Saponas, Andrew Guillory, and Ilya Kelner

    Dan Morris, T. Scott Saponas, Andrew Guillory, and Ilya Kelner. Recofit.Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, page 3225–3234, Apr 2014

    2014

  26. [34]

    Tailor, Jacob Sunshine, Yun Liu, Tim Althoff, Shrikanth Narayanan, Pushmeet Kohli, Jiening Zhan, Mark Malhotra, Shwetak Patel, Samy Abdel-Ghaffar, and Daniel McDuff

    Girish Narayanswamy, Xin Liu, Kumar Ayush, Yuzhe Yang, Xuhai Xu, Shun Liao, Jake Garri- son, Shyam A. Tailor, Jacob Sunshine, Yun Liu, Tim Althoff, Shrikanth Narayanan, Pushmeet Kohli, Jiening Zhan, Mark Malhotra, Shwetak Patel, Samy Abdel-Ghaffar, and Daniel McDuff. Scaling w...

    2025

  27. [35]

    National Health and Nutrition Examination Survey (NHANES) 2011-2014 Physical Activity Monitor Data

    National Center for Health Statistics. National Health and Nutrition Examination Survey (NHANES) 2011-2014 Physical Activity Monitor Data. Centers for Disease Control and Preven- tion, 2014

    2011

  28. [36]

    Sharing digital health data responsibly: Balancing open science with participant privacy.PLOS Digital Health, 5(4):e0001377, 2026

    Camille Nebeker, Shahin Samiei, and Santosh Kumar. Sharing digital health data responsibly: Balancing open science with participant privacy.PLOS Digital Health, 5(4):e0001377, 2026

    2026

  29. [37]

    A time series is worth 64 words: Long-term forecasting with transformers

    Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers. InInternational Conference on Learning Representations, 2023

    2023

  30. [38]

    Assessing inertial measurement unit locations for freezing of gait detection and patient preference.Journal of neuroengineering and rehabilitation, 19(1):20, 2022

    JohannaO’Day, MarissaLee, KirstenSeagers, ShannonHoffman, AvaJih-Schiff, ŁukaszKidziński, Scott Delp, and Helen Bronte-Stewart. Assessing inertial measurement unit locations for freezing of gait detection and patient preference.Journal of neuroengineering and rehabilitation, 1...

    2022

  31. [39]

    The all of us research program’s wearables dataset.Nature Medicine, pages 1–9, 2026

    Theresa Patten, Edward A Preble, Hiral Master, Jennifer Adjemian, Andrea Ramirez, James McClain, and Amy Rose Price. The all of us research program’s wearables dataset.Nature Medicine, pages 1–9, 2026

    2026

  32. [40]

    PAMAP2 Physical Activity Monitoring

    Attila Reiss. PAMAP2 Physical Activity Monitoring. UCI Machine Learning Repository, 2012. DOI: https://doi.org/10.24432/C5NW2H. 14 Inertia-1: An Open Exploration of Wearable Motion Foundation Models

    2012 doi

  33. [41]

    Caroline Ribeiro De Souza, Runfeng Miao, Júlia Ávila De Oliveira, Andrea Cristina De Lima- Pardini, Débora Fragoso De Campos, Carla Silva-Batista, Luis Teixeira, Solaiman Shokur, Bouri Mohamed, and Daniel Boari Coelho. A public data set of videos, inertial measurement unit, an...

    2022

  34. [42]

    Moreno Arostegui, Joan Cabestany, and Alejandro Rodriguez-Molinero

    Daniel Rodriguez-Martin, Albert Sama, Celia Perez-Lopez, Andreu Catala, Juan M. Moreno Arostegui, Joan Cabestany, and Alejandro Rodriguez-Molinero. Home detection of freezingofgaitusingsupportvectormachinesthroughasinglewaist-worntriaxialaccelerometer. PLOS ONE, 12(2):e0171764, 2017

    2017

  35. [43]

    OPPORTUNITY Activity Recognition

    Daniel Roggen, Alberto Calatroni, Long-Van Nguyen-Dinh, Ricardo Chavarriaga, and Hesam Sagha. OPPORTUNITY Activity Recognition. UCI Machine Learning Repository, 2010. DOI: https://doi.org/10.24432/C5M027

    2010 doi

  36. [44]

    Daphnet Freezing of Gait

    Daniel Roggen, Meir Plotnik, and Jeff Hausdorff. Daphnet Freezing of Gait. UCI Machine Learning Repository, 2010. DOI: https://doi.org/10.24432/C56K78

    2010 doi

  37. [45]

    Unsupervised domain adaptation in activity recognition: A gan-based approach.IEEE Access, 9:19421–19438, 2021

    Andrea Rosales Sanabria, Franco Zambonelli, and Juan Ye. Unsupervised domain adaptation in activity recognition: A gan-based approach.IEEE Access, 9:19421–19438, 2021

    2021

  38. [46]

    Jinjoo Shim, Elgar Fleisch, and Filipe Barata. Wearable-based accelerometer activity profile as digital biomarker of inflammation, biological age, and mortality using hierarchical clustering analysis in nhanes 2011–2014.Scientific Reports, 13(1):9326, Jun 2023

    2011

  39. [47]

    OSF: On pre-training and scaling of sleep foundation models.arXiv preprint arXiv:2603.00190, 2026

    Zitao Shuai, Zongzhe Xu, David Yang, Wei Wang, and Yuzhe Yang. OSF: On pre-training and scaling of sleep foundation models.arXiv preprint arXiv:2603.00190, 2026

    2026 arXiv

  40. [48]

    Dey, Tobias Sonne, and Mads Moller Jensen

    Allan Stisen, Henrik Blunck, Sourav Bhattacharya, Thor Siiger Prentow, Mikkel Baun Kjaergaard, Anind K. Dey, Tobias Sonne, and Mads Moller Jensen. Smart devices are different: Assessing and mitigating mobile sensing heterogeneities for activity recognition. InProceedings of th...

    2015

  41. [49]

    Chi Ian Tang, Ignacio Perez-Pozuelo, Dimitris Spathis, Soren Brage, Nick Wareham, and Cecilia Mascolo. Selfhar: Improving human activity recognition through self-training with unlabeled data.Proceedings of the ACM on interactive, mobile, wearable and ubiquitous technologies, 5...

    2021

  42. [50]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need, 2017

    2017

  43. [51]

    Smooth-distill: A self-distillation framework for multitask learning with wearable sensor data.arXiv preprint arXiv:2507.00061, 2025

    Hoang-Dieu Vu, Duc-Nghia Tran, Quang-Tu Pham, Hieu H Pham, Nicolas Vuillerme, and Duc- Tan Tran. Smooth-distill: A self-distillation framework for multitask learning with wearable sensor data.arXiv preprint arXiv:2507.00061, 2025

    2025 arXiv

  44. [52]

    WISDMSmartphoneandSmartwatchActivityandBiometricsDataset

    GaryWeiss. WISDMSmartphoneandSmartwatchActivityandBiometricsDataset. UCIMachine Learning Repository, 2019. DOI: https://doi.org/10.24432/C5HK59

    2019 doi

  45. [53]

    Relcon: Rel- ative contrastive learning for a motion foundation model for wearable data.arXiv preprint arXiv:2411.18822, 2025

    Maxwell A Xu, Jaya Narain, Gregory Darnell, Haraldur Hallgrimsson, Hyewon Jeong, Darren Forde, Richard Fineman, Karthik J Raghuram, James M Rehg, and Shirley Ren. Relcon: Rel- ative contrastive learning for a motion foundation model for wearable data.arXiv preprint arXiv:2411....

    2025 arXiv

  46. [54]

    Xu, Girish Narayanswamy, Kumar Ayush, Dimitris Spathis, Shun Liao, Shyam A

    Maxwell A. Xu, Girish Narayanswamy, Kumar Ayush, Dimitris Spathis, Shun Liao, Shyam A. Tailor, Ahmed Metwally, A. Ali Heydari, Yuwei Zhang, Jake Garrison, Samy Abdel-Ghaffar, Xuhai Xu, Ken Gu, Jacob Sunshine, Ming-Zher Poh, Yun Liu, Tim Althoff, Shrikanth Narayanan, Pushmeet K...

    2025

  47. [55]

    SleepLM: Natural-language intelligence for human sleep.arXiv preprint arXiv:2602.23605, 2026

    Zongzhe Xu, Zitao Shuai, Eideen Mozaffari, Ravi S Aysola, Rajesh Kumar, and Yuzhe Yang. SleepLM: Natural-language intelligence for human sleep.arXiv preprint arXiv:2602.23605, 2026

    2026 arXiv

  48. [56]

    Effects of sampling frequency on human activity recognition with machine learning aiming at clinical applications.Sensors, 25(12):3780, 2025

    Takahiro Yamane, Moeka Kimura, and Mizuki Morita. Effects of sampling frequency on human activity recognition with machine learning aiming at clinical applications.Sensors, 25(12):3780, 2025

    2025

  49. [57]

    Simper: Simple self-supervised learning of periodic targets

    Yuzhe Yang, Xin Liu, Jiang Wu, Silviu Borac, Dina Katabi, Ming-Zher Poh, and Daniel McDuff. Simper: Simple self-supervised learning of periodic targets. InInternational Conference on Learning Representations (ICLR), 2023

    2023

  50. [58]

    Tarolli, Daniel Crepeau, Jan Bukartyk, Mithri R

    Yuzhe Yang, Yuan Yuan, Guo Zhang, Hao Wang, Ying-Cong Chen, Yingcheng Liu, Christopher G. Tarolli, Daniel Crepeau, Jan Bukartyk, Mithri R. Junna, et al. Artificial intelligence-enabled detection and assessment of parkinson’s disease using nocturnal breathing signals.Nature Med...

    2022

  51. [59]

    Creagh, Catherine Tong, Aidan Acquah, David A

    Hang Yuan, Shing Chan, Andrew P. Creagh, Catherine Tong, Aidan Acquah, David A. Clifton, and Aiden Doherty. Self-supervised learning for human activity recognition using 700,000 person-days of wearable data.npj Digital Medicine, 7(1):91, 2024

    2024

  52. [60]

    Sensorlm: Learning the language of wearable sensors.Neural Information Processing Systems (NeurIPS), 2025

    Yuwei Zhang, Kumar Ayush, Siyuan Qiao, A Ali Heydari, Girish Narayanswamy, Maxwell A Xu, Ahmed A Metwally, Shawn Xu, Jake Garrison, Xuhai Xu, et al. Sensorlm: Learning the language of wearable sensors.Neural Information Processing Systems (NeurIPS), 2025. 16 Inertia-1: An Open...

    2025

  53. [61]

    Finally, we include the raw AUROC and AUPRC across all axes ablations discussed in the main paper (Table 15, Table 16)

    tasks. Finally, we include the raw AUROC and AUPRC across all axes ablations discussed in the main paper (Table 15, Table 16). D. Supplementary Scaling Analyses D.1. Parameter scaling under fixed pretraining data We investigate scaling behavior by varying model capacity (Small...

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.