Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

LSM-2: Learning from Incomplete Wearable Sensor Data

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A shared learnable mask token lets a masked autoencoder represent both real and artificial gaps in day-long wearable data, so LSM-2 with AIM is trained and evaluated directly on incomplete sensor records without imputation.

desk verdict Genuine extension of MAE for incomplete wearables, but the missing participant-disjoint split on the Metabolics data undermines the headline claims as written. read the letter →

arxiv 2506.05321 v1 pith:IAE5AHEK submitted 2025-06-05 cs.LG

Maxwell A. Xu , Girish Narayanswamy , Kumar Ayush , Dimitris Spathis , Shun Liao , Shyam A. Tailor , Ahmed Metwally , A. Ali Heydari
show 17 more authors
Yuwei Zhang Jake Garrison Samy Abdel-Ghaffar Xuhai Xu Ken Gu Jacob Sunshine Ming-Zher Poh Yun Liu Tim Althoff Shrikanth Narayanan Pushmeet Kohli Mark Malhotra Shwetak Patel Yuzhe Yang James M. Rehg Xin Liu Daniel McDuff
This is my paper · ORCID
classification cs.LG
keywords self-supervisedlearningwearablesensorsmissingdatamaskedautoencoderfoundationmodelhealthtimeseriessensormissingnessrepresentation
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

Day-long wearable sensor recordings are almost never complete: battery drain, device removal, and signal noise leave gaps in every record. Prior self-supervised models either worked on short windows where missingness could be filtered out, or imputed the gaps before training, risking biases. LSM-2 with AIM claims a third route: train an autoencoder directly on the incomplete data by representing both real ('inherited') gaps and artificially masked patches with the same learnable mask token. Pre-trained on 40 million hours of data from 60,440 people, the model is reported to outperform prior self-supervised baselines on classification, regression, and generative tasks, to scale better with data and compute, and to degrade gracefully when entire sensors or time windows are removed. If correct, this would make day-long wearable data usable for health prediction without preprocessing that fabricates or discards measurements.

What carries the argument

The load-bearing object is the AIM mask token used inside a hybrid masking strategy. In the unified mask $M = M_{\mathrm{inherited}} \lor M_{\mathrm{artificial}}$, one learnable token stands for 'missing' whether the gap came from the raw signal or from the artificial pre-training mask; dropout removal drops a fixed number of tokens for efficiency, while an attention mask suppresses the variable remaining inherited-mask tokens so the encoder output represents only observed data. Pre-training uses a mixture of artificial masks—80% random imputation, 50% temporal slices, and 50% signal slices—chosen to mirror the real missingness modes of sensor noise, device-off periods, and sensor-off periods. The input is a 1440-minute day of 26 minutely features tokenized into 3744 patches, giving the model day-long context on which to learn reconstruction.

What would settle it

Evaluate LSM-2 on day-long data from a wearable device family that was absent from pre-training, holding prediction targets fixed (for instance, hypertension F1 and reconstruction MSE under 80% random imputation). If performance on unseen-device missingness drops substantially more than on held-out samples from seen devices with matched overall missingness, the learned mask token does not transfer and the robustness claim would need to be qualified.

Watch

Extended reading notes

Core claim

The paper's central claim is that a masked autoencoder can learn a general-purpose representation directly from fragmented multimodal wearable data, with no imputation step, by making the masking procedure aware of real-world missingness. The method, Adaptive and Inherited Masking (AIM), sets the model's mask to the union of an inherited mask (the gaps already present in the raw data) and an artificial mask (randomly selected observed patches that serve as the reconstruction target). A shared learnable mask token represents both kinds of missingness, and an adaptive attention mechanism ensures that only non-missing tokens contribute to the pooled encoder output. Because inherited masks vary in size, AIM combines dropout removal with attention masking to keep long-sequence training efficient. On 40 million hours of day-long minutely aggregated data, LSM-2 with AIM is reported to achieve the best performance among the evaluated self-supervised methods on classification, regression, and generative tasks, to scale better than the imputation-based LSM-1, and to show smaller performance drops under targeted sensor and temporal-window removal; where sensitivity remains, it follows clinical expectations such as nighttime biosignals mattering more for hypertension and anxiety.

Load-bearing premise

The approach assumes that the mask token learned on the specific wearable devices used in pre-training transfers to other devices and missingness regimes; the appendix notes that only a small subset of such devices was tested and that device-specific missingness distribution shift remains an open question.

Editorial extensions

If this is right

  • Wearable foundation models can be pre-trained directly on raw, fragmented day-long recordings, making imputation an unnecessary preprocessing step for these models.
  • Full-day context becomes usable for self-supervised learning even though no day-long sample is complete, exposing models to circadian and daily-activity patterns that short windows cannot capture.
  • The same mask mechanism handles missingness at deployment, so one frozen model can serve data with very different sensor availability.
  • Targeted-missingness evaluation doubles as a sensitivity analysis: removing a sensor or time window degrades predictions roughly in proportion to that signal's clinical importance, as seen with nighttime biosignals for hypertension.
  • Because AIM only requires that data contain existing missingness, the same approach could extend to other missingness-afflicted domains such as tabular clinical records or environmental sensor arrays.

Reading between the lines

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

  • Probing the learned mask token's embedding across device families would test whether it encodes missingness semantics (charging gaps versus signal noise) or just acts as a learned constant; this needs no new training and is a natural next experiment.
  • The targeted-missingness results suggest a cheap attribution method for health models: mask each sensor or time window, measure the prediction drop, and use the resulting maps to identify which signals carry diagnostic information for a given outcome.
  • Because all baselines, including LSM-1, were trained on imputed data, the reported gains combine the AIM masking design with the removal of imputation artifacts; training an LSM-1-style model on non-imputed data would separate these two factors.
  • The reliance on minutely aggregated features leaves open whether AIM transfers to raw high-frequency sensor streams, which are the norm in much of the wearable-sensing literature; extending the method to raw waveforms would require new tokenization and storage-scale solutions.
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

3 major / 5 minor

Summary. The paper proposes LSM-2, a self-supervised foundation model for day-long wearable sensor data, trained with Adaptive and Inherited Masking (AIM). AIM unions an inherited mask derived from real missingness with an artificial mask, represents both with a shared learnable mask token, and combines dropout removal with attention masking to handle variable missingness rates efficiently. The model is pre-trained on 40M hours of private multimodal data and evaluated on generative reconstruction, classification (hypertension, anxiety, 20-class activity recognition), and regression (age, BMI), including targeted sensor- and time-window-removal robustness tests. The central claim is that LSM-2 with AIM achieves state-of-the-art performance across all three task families without imputation.

Significance. If the results hold, this is a useful advance: the method is well-specified, the hybrid dropout/attention masking addresses a real computational constraint, and Table 5's ablations support the specific contribution of inherited masking. The targeted missingness experiments produce physically interpretable patterns (e.g., nighttime biosignal importance for hypertension), and the paper is candid about the private data, the small device family, and the open question of missingness distribution shift. The main strengths are the clarity of the method, the scale of pre-training, and the breadth of downstream evaluation. The main unresolved risk is the train/test split of the Metabolics dataset, on which the classification and regression claims rest.

major comments (3)
  1. [Section 3, Table 8] The Metabolics downstream split is not stated to be participant-disjoint, and this is load-bearing for the classification and regression claims. The text reports 1,250 participants, while Table 8 lists 151,346 training and 90,186 test day-level samples, implying many days per participant; the labels (hypertension, anxiety, age, BMI) are person-level constants. In contrast, the Activity dataset section explicitly says the split is person-independent. If days from the same participant appear in both train and test, a linear probe can memorize person-specific patterns and inflate every headline metric. Please state the split granularity, provide a participant-disjoint evaluation, or apply a repeated-measures correction.
  2. [Section 4 and Appendix A.3] The artificial mask ratios (80% random, 50% temporal slice, 50% signal slice) were selected via ablation on the same generative reconstruction metrics used in Table 3, so the generative results are partly tuned rather than out-of-sample predictions. Appendix A.3 explicitly presents the ratio-selection tables on random imputation, temporal interpolation, and signal imputation MSE. This weakens the generative SOTA claim, although the classification and regression results do provide some independent grounding. Please clarify the model-selection procedure and, if possible, validate the chosen ratios on held-out generative tasks or on independent downstream tasks.
  3. [Tables 2-5 and Figures 5-6] No error bars, confidence intervals, or significance tests are reported anywhere. Several headline differences are small: Table 2 shows LSM-2 versus LSM-1 gains of +0.4 to +2.1 points, and Table 4 shows age correlation is -0.8% worse than LSM-1. Without variance estimates or multiple runs, these differences may be noise, and the claim of 'best performance across ... regression' is not supported for age. Report standard errors or confidence intervals over seeds or bootstrap resamples.
minor comments (5)
  1. [Section 3, Data Summary] The sentence '0% of samples over our entire dataset of 1.6 million instances of 1 day data' is self-contradictory; it should read '0% of samples are complete' or '100% of samples contain missingness'.
  2. [Section 4, Adaptive design] The statement 'we set D = 0.5N' appears to conflict with the 80% random artificial masking strategy; please clarify whether D equals the artificial mask size for the chosen strategy, or is a fixed value independent of the strategy.
  3. [Contribution 2 and Table 4] The contributions list '3x regression' tasks, but Table 4 reports only Age and BMI; please identify the third regression target or correct the claim.
  4. [Appendix A.4.1 and A.6.2] There are several typos: 'frp' should be 'from', 'amd' should be 'and', 'Coeffecient' should be 'Coefficient', and 'like are own' should be 'like our own'.
  5. [Figure 5] The scaling curves have no error bars and the axes are not clearly labeled; please specify the number of runs and define the plotted loss or metric precisely.

Circularity Check

1 steps flagged · score 4.0 of 10

Generative SOTA partly reflects mask hyperparameters tuned on the same generative metrics; classification and regression results remain independent.

  1. fitted input called prediction [Section 4 (Unified Framework for Pre-training and Evaluation) and Appendix A.3 (Pre-training Masking % Ablation Experiment)]
    "The specific masking percentages were identified via an ablation study, reported within the Appendix. ... Below in Tables 9, 10, 11, we see that an 80% random imputation mask %, 50% temporal slice %, and a 50% sensor slice % produce a good mix of reconstruction results across small and large amounts of evaluation masking, for each generative task."

    The generative evaluation in Table 3 reports MSE for Random Imputation, Temporal Interpolation, and Signal Imputation. Appendix Tables 9-11 tune the pre-training random-imputation, temporal-slice, and sensor-slice percentages against exactly those same evaluation metrics: Random Imp. Eval Ratio, Temporal Interp. Eval Amount, and Sensor Imp. Eval Amount. The paper then presents the resulting Table 3 numbers as a state-of-the-art generative result. This is not an independent prediction: the mask mix is selected on the very metrics later reported as the model's generative performance. The classification and regression results are not selected on their own metrics and therefore retain independent evidentiary value.

full rationale

One genuinely circular step exists: the generative headline is partly a restatement of the hyperparameter selection criterion, since the 80/50/50 pre-training mask ratios were chosen by ablating on the same random-imputation, temporal-interpolation, and sensor-imputation metrics later reported as SOTA generative results. This affects only the generative claim family. The classification results (hypertension, anxiety, activity recognition) and regression results (age, BMI) are not tuned on their own metrics and independently support the central representation-learning claim. The paper's self-citations, primarily LSM-1 [42], are used as from-scratch baselines and scaling comparisons rather than as load-bearing external justification; no uniqueness theorem or ansatz is imported from prior work. The appendix's stated limitations, such as private data, device-specific missingness distribution shift, and minutely aggregated features, are acknowledged scope constraints rather than circular reasoning. The absence of an explicit participant-disjoint split for the Metabolics tasks is a potentially serious correctness and leakage concern, but it is not a circularity of the derivation chain, so it does not raise the circularity score.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the assumptions that masked reconstruction is a good SSL objective, that the 26 channels can be treated independently, and that the missingness seen in pre-training transfers to evaluation. The free parameters are the mask ratios and dropout amount, tuned via ablation on the same generative metrics.

free parameters (2)
  • Artificial mask ratios (random, temporal slice, signal slice) = 80%, 50%, 50%
    Chosen via ablation study (Appendix A.3) on the same generative evaluation metrics; affects all reported results.
  • Dropout removal ratio D = 0.5N
    Set to 0.5N to trade off efficiency and reconstruction; selected together with the mask ratios.
assumptions (3)
  • domain assumption Reconstruction of artificially masked tokens is a useful SSL objective for downstream tasks
    Standard MAE assumption, not proven for wearable data; the paper optimizes MSE on artificial masks.
  • domain assumption Features are independent channels
    Model treats 26 features as independent; appendix acknowledges correlations exist (A.6.3).
  • domain assumption Inherited missingness patterns in pretraining are representative of evaluation missingness
    The model is pre-trained on one set of devices and evaluated on similar devices; generalization across device-specific missingness patterns is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LSM-2: Learning from Incomplete Wearable Sensor Data." pith.science (2026). https://pith.science/paper/IAE5AHEK

@misc{pith2026250605321,
  author       = {Pith},
  title        = {Pith review of: LSM-2: Learning from Incomplete Wearable Sensor Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IAE5AHEK}},
  note         = {Machine review of arXiv:2506.05321}
}
read the original abstract

Foundation models, a cornerstone of recent advancements in machine learning, have predominantly thrived on complete and well-structured data. Wearable sensor data frequently suffers from significant missingness, posing a substantial challenge for self-supervised learning (SSL) models that typically assume complete data inputs. This paper introduces the second generation of Large Sensor Model (LSM-2) with Adaptive and Inherited Masking (AIM), a novel SSL approach that learns robust representations directly from incomplete data without requiring explicit imputation. AIM's core novelty lies in its use of learnable mask tokens to model both existing ("inherited") and artificially introduced missingness, enabling it to robustly handle fragmented real-world data during inference. Pre-trained on an extensive dataset of 40M hours of day-long multimodal sensor data, our LSM-2 with AIM achieves the best performance across a diverse range of tasks, including classification, regression and generative modeling. Furthermore, LSM-2 with AIM exhibits superior scaling performance, and critically, maintains high performance even under targeted missingness scenarios, reflecting clinically coherent patterns, such as the diagnostic value of nighttime biosignals for hypertension prediction. This makes AIM a more reliable choice for real-world wearable data applications.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. TRACE-TS: Attribution-Grounded and Traceable Sensor-Language Reasoning for Human Activity Understanding

    cs.AI 2026-07 conditional novelty 6.0 of 10

    Attribution-guided DAG reasoning traces distilled into a compact sensor-conditioned LM give 84.43% average accuracy and 81.24% F1 across seven HAR datasets, with SNM as an LLM-judge diagnostic for trace fidelity.

  2. OpenMHC: Accelerating the Science of Wearable Foundation Models

    cs.LG 2026-06 conditional novelty 6.0 of 10

    OpenMHC contributes the largest open-access consumer wearable dataset to date (67M hours, 11,894 participants), a standardized three-track benchmark, and the first open implementations of Apple WBM and Google LSM-2.

Reference graph

Works this paper leans on

78 extracted references · 57 canonical work pages · cited by 2 Pith papers

  1. [1]

    Abbaspourazad, O

    S. Abbaspourazad, O. Elachqar, A. C. Miller, S. Emrani, U. Nallasamy, and I. Shapiro. Large-scale training of foundation models for wearable biosignals.arXiv preprint arXiv:2312.05409, 2023

    arXiv 2023

  2. [2]

    Abd-Alrazaq, R

    A. Abd-Alrazaq, R. AlSaad, S. Aziz, A. Ahmed, K. Denecke, M. Househ, F. Farooq, and J. Sheikh. Wearable artificial intelligence for anxiety and depression: scoping review.Journal of Medical Internet Research, 25:e42672, 2023

    work page 2023

  3. [3]

    Afdala, N

    A. Afdala, N. Nuryani, and A. S. Nugroho. Automatic detection of atrial fibrillation using basic shannon entropy of rr interval feature. InJournal of Physics: Conference Series, volume 795, page 012038. IOP Publishing, 2017

    work page 2017

  4. [4]

    Amiri and R

    M. Amiri and R. Jensen. Missing data imputation using fuzzy-rough methods.Neurocomputing, 205:152–164, 2016

    work page 2016

  5. [5]

    A. F. Ansari, L. Stella, C. Turkmen, X. Zhang, P. Mercado, H. Shen, O. Shchur, S. S. Rangapuram, S. P. Arango, S. Kapoor, et al. Chronos: Learning the language of time series.arXiv preprint arXiv:2403.07815, 2024. 11 LSM-2: Learning from Incomplete Wearable Sensor Data

    arXiv 2024

  6. [6]

    Assran, M

    M. Assran, M. Caron, I. Misra, P. Bojanowski, F. Bordes, P. Vincent, A. Joulin, M. Rabbat, and N. Ballas. Masked siamese networks for label-efficient learning. InEuropean conference on computer vision, pages 456–473. Springer, 2022

    work page 2022

  7. [7]

    Bähr, G.-C

    S. Bähr, G.-C. Haas, F. Keusch, F. Kreuter, and M. Trappmann. Missing data and other mea- surement quality issues in mobile geolocation sensor data.Social Science Computer Review, 40(1):212–235, 2022

    work page 2022

  8. [8]

    Beebe-Wang, S

    N. Beebe-Wang, S. Ebrahimi, J. Yoon, S. O. Arik, and T. Pfister. Paits: pretraining and augmenta- tion for irregularly-sampled time series.arXiv preprint arXiv:2308.13703, 2023

    arXiv 2023

Show all 78 references
  1. [9]

    Bleser, D

    G. Bleser, D. Steffen, A. Reiss, M. Weber, G. Hendeby, and L. Fradet. Personalized physical activity monitoring using wearable sensors.Smart health: Open problems and future challenges, pages 99–124, 2015

    2015

  2. [10]

    W. Cao, D. Wang, J. Li, H. Zhou, L. Li, and Y. Li. Brits: Bidirectional recurrent imputation for time series.Advances in neural information processing systems, 31, 2018

    2018

  3. [11]

    Caron, H

    M. Caron, H. Touvron, I. Misra, H. Jégou, J. Mairal, P. Bojanowski, and A. Joulin. Emerging properties in self-supervised vision transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 9650–9660, 2021

    2021

  4. [12]

    Chang, C.-T

    L.-W. Chang, C.-T. Li, C.-P. Yang, and S.-d. Lin. Learning on missing tabular data: Attention with self-supervision, not imputation, is all you need.ACM Transactions on Intelligent Systems and Technology

  5. [13]

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton. A simple framework for contrastive learning of visual representations. InInternational conference on machine learning, pages 1597–1607. PmLR, 2020

    2020

  6. [14]

    H.-Y. S. Chien, H. Goh, C. M. Sandino, and J. Y. Cheng. Maeeg: Masked auto-encoder for eeg representation learning.arXiv preprint arXiv:2211.02625, 2022

    2022 arXiv

  7. [15]

    ChuDuc, K

    H. ChuDuc, K. NguyenPhan, and D. NguyenViet. A review of heart rate variability and its applications. APCBEE procedia, 7:80–85, 2013

    2013

  8. [16]

    Z. Dai, E. Getzen, and Q. Long. Sadi: Similarity-aware diffusion model-based imputation for incomplete temporal ehr data. InInternational Conference on Artificial Intelligence and Statistics, pages 4195–4203. PMLR, 2024

    2024

  9. [17]

    A. Das, W. Kong, R. Sen, and Y. Zhou. A decoder-only foundation model for time-series forecasting. InForty-first International Conference on Machine Learning, 2024

    2024

  10. [18]

    Decorte, S

    T. Decorte, S. Mortier, J. J. Lembrechts, F. J. Meysman, S. Latré, E. Mannens, and T. Verdonck. Missing value imputation of wireless sensor data for environmental monitoring.Sensors, 24(8):2416, 2024

    2024

  11. [19]

    C. M. DeGiorgio, P. Miller, S. Meymandi, A. Chin, J. Epps, S. Gordon, J. Gornbein, and R. M. Harper. Rmssd, a measure of vagus-mediated heart rate variability, is associated with risk factors for sudep: the sudep-7 inventory.Epilepsy & behavior, 19(1):78–81, 2010

    2010

  12. [20]

    J. Dong, H. Wu, H. Zhang, L. Zhang, J. Wang, and M. Long. Simmtm: A simple pre-training framework for masked time-series modeling.Advances in Neural Information Processing Systems, 36:29996–30025, 2023. 12 LSM-2: Learning from Incomplete Wearable Sensor Data

    2023

  13. [21]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    2010 arXiv

  14. [22]

    J. Du, M. Hu, and W. Zhang. Missing data problem in the monitoring system: A review.IEEE Sensors Journal, 20(23):13984–13998, 2020

    2020

  15. [23]

    T. Du, L. Melis, and T. Wang. Remasker: Imputing tabular data with masked autoencoding. arXiv preprint arXiv:2309.13793, 2023

    2023 arXiv

  16. [24]

    Ericsson, H

    L. Ericsson, H. Gouk, and T. M. Hospedales. How well do self-supervised models transfer? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5414–5423, 2021

    2021

  17. [25]

    J. Fan, J. Mei, Y. Yang, J. Lu, Q. Wang, X. Yang, G. Chen, R. Wang, Y. Han, R. Sheng, et al. Sleep-phasic heart rate variability predicts stress severity: Building a machine learning-based stress prediction model.Stress and Health, 40(4):e3386, 2024

    2024

  18. [26]

    E. Ford, P. Rooney, P. Hurley, S. Oliver, S. Bremner, and J. Cassell. Can the use of bayesian analysis methods correct for incompleteness in electronic health records diagnosis data? development of a novel method using simulated and real-life clinical data.Frontiers in Public ...

    2020

  19. [27]

    Haneuse, D

    S. Haneuse, D. Arterburn, and M. J. Daniels. Assessing missing data assumptions in ehr-based studies: a complex and underappreciated task.JAMA Network Open, 4(2):e210184–e210184, 2021

    2021

  20. [28]

    T. W. Hansen, Y. Li, J. Boggia, L. Thijs, T. Richart, and J. A. Staessen. Predictive role of the nighttime blood pressure.Hypertension, 57(1):3–10, 2011

    2011

  21. [30]

    K. He, X. Chen, S. Xie, Y. Li, P. Dollár, and R. Girshick. Masked autoencoders are scalable vision learners. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000–16009, 2022

    2022

  22. [31]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. InCVPR, 2016

    2016

  23. [32]

    Hecht, S

    A. Hecht, S. Ma, J. Porszasz, R. Casaburi, C. C. R. Network, et al. Methodology for using long-term accelerometry monitoring to describe daily activity patterns in copd.COPD: Journal of Chronic Obstructive Pulmonary Disease, 6(2):121–129, 2009

    2009

  24. [33]

    Jeong, A

    H. Jeong, A. Roghanizad, H. Master, and et al. Data from the All of Us research program reinforces existence of activity inequality.npj Digital Medicine, 8(8), 2025

    2025

  25. [34]

    Representationlearningforwearable-basedapplications in the case of missing data.arXiv preprint arXiv:2401.05437, 2024

    J.Jungo, Y.Xiang, S.Gashi, andC.Holz. Representationlearningforwearable-basedapplications in the case of missing data.arXiv preprint arXiv:2401.05437, 2024

    2024 arXiv

  26. [35]

    Katori, S

    M. Katori, S. Shi, K. Ode, Y. Tomita, and H. Ueda. The 103,200-arm acceleration dataset in the uk biobank revealed a landscape of human sleep phenotypes.Proceedings National Academy of Science, U.S.A., 119(12), 2022. 13 LSM-2: Learning from Incomplete Wearable Sensor Data

    2022

  27. [36]

    Kinnunen, A

    H. Kinnunen, A. Rantanen, T. Kenttä, and H. Koskimäki. Feasible assessment of recovery and cardiovascular health: accuracy of nocturnal hr and hrv assessed via ring ppg in comparison to medical grade ecg.Physiological measurement, 41(4):04NT01, 2020

    2020

  28. [37]

    Z. Li, Z. Rao, L. Pan, P. Wang, and Z. Xu. Ti-mae: Self-supervised masked time series autoen- coders. arXiv preprint arXiv:2301.08871, 2023

    2023 arXiv

  29. [38]

    Z. Liu, A. Alavi, M. Li, and X. Zhang. Guidelines for augmentation selection in contrastive learning for time series classification.arXiv preprint arXiv:2407.09336, 2024

    2024 arXiv

  30. [39]

    McDermott, B

    M. McDermott, B. Nestor, E. Kim, W. Zhang, A. Goldenberg, P. Szolovits, and M. Ghassemi. A comprehensive ehr timeseries pre-training benchmark. InProceedings of the Conference on Health, Inference, and Learning, pages 257–278, 2021

    2021

  31. [40]

    Mekruksavanich, A

    S. Mekruksavanich, A. Jitpattanakul, K. Sitthithakerngkiet, P. Youplao, and P. Yupapin. Resnet- se: Channel attention-based deep residual network for complex activity recognition using wrist-worn wearable sensors.IEEE Access, 10:51142–51154, 2022

    2022

  32. [41]

    Y. Na, M. Park, Y. Tae, and S. Joo. Guiding masked representation learning to capture spatio- temporal relationship of electrocardiogram.arXiv preprint arXiv:2402.09450, 2024

    2024 arXiv

  33. [42]

    Narayanswamy, X

    G. Narayanswamy, X. Liu, K. Ayush, Y. Yang, X. Xu, S. Liao, J. Garrison, S. Tailor, J. Sunshine, Y. Liu, et al. Scaling wearable foundation models.arXiv preprint arXiv:2410.13638, 2024

    2024 arXiv

  34. [43]

    Bigsmall: Efficientmulti- task learning for disparate spatial and temporal physiological measurements

    G.Narayanswamy, Y.Liu, Y.Yang, C.Ma, X.Liu, D.McDuff, andS.Patel. Bigsmall: Efficientmulti- task learning for disparate spatial and temporal physiological measurements. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7914–7924, 2024

    2024

  35. [44]

    Atimeseriesisworth64words: Long-term forecasting with transformers.arXiv preprint arXiv:2211.14730, 2022

    Y.Nie, N.H.Nguyen, P.Sinthong, andJ.Kalagnanam. Atimeseriesisworth64words: Long-term forecasting with transformers.arXiv preprint arXiv:2211.14730, 2022

    2022 arXiv

  36. [45]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. Vo, M. Szafraniec, V. Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, et al. Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  37. [46]

    Y.-C. Pan, B. Goodwin, E. Sabelhaus, K. M. Peters, K. F. Bjornson, K. L. Pham, W. Walker, and K. M. Steele. Feasibility of using acceleration-derived jerk to quantify bimanual arm use.Journal of NeuroEngineering and Rehabilitation, 17:1–8, 2020

    2020

  38. [47]

    Pillai, D

    A. Pillai, D. Spathis, F. Kawsar, and M. Malekzadeh. Papagei: Open foundation models for optical physiological signals.International Conference on Learning Representations (ICLR), 2025

    2025

  39. [48]

    I. M. Pires, F. Hussain, N. M. Garcia, and E. Zdravevski. Improving human activity monitoring by imputation of missing sensory data: Experimental study.Future Internet, 12(9):155, 2020

    2020

  40. [49]

    Qin and Y

    R. Qin and Y. Wang. Imputegan: Generative adversarial network for multivariate time series imputation. Entropy, 25(1):137, 2023

    2023

  41. [50]

    M. M. Rahman, N. Ali, R. Bari, N. Saleheen, M. al’Absi, E. Ertin, A. Kennedy, K. L. Preston, and S. Kumar. mDebugger: Assessing and diagnosing the fidelity and yield of mobile sensor data. In Mobile Health: Sensors, Analytic Methods, and Applications, chapter 7, page 121–143. 2017

    2017

  42. [51]

    Rommel, J

    C. Rommel, J. Paillard, T. Moreau, and A. Gramfort. Data augmentation for learning predictive models on eeg: a systematic comparison.Journal of Neural Engineering, 19(6):066020, 2022. 14 LSM-2: Learning from Incomplete Wearable Sensor Data

    2022

  43. [52]

    M. Saha, M. A. Xu, W. Mao, S. Neupane, J. M. Rehg, and S. Kumar. Pulse-ppg: An open-source field-trained ppg foundation model for wearable applications across lab and field settings.arXiv preprint arXiv:2502.01108, 2025

    2025 arXiv

  44. [53]

    Schmidt, A

    P. Schmidt, A. Reiss, R. Duerichen, C. Marberger, and K. Van Laerhoven. Introducing wesad, a multimodal dataset for wearable stress and affect detection. InProceedings of the 20th ACM international conference on multimodal interaction, pages 400–408, 2018

    2018

  45. [54]

    Sevil, M

    M. Sevil, M. Rashid, M. R. Askari, Z. Maloney, I. Hajizadeh, and A. Cinar. Detection and characterization of physical activity and psychological stress from wristband data.Signals, 1(2):188–208, 2020

    2020

  46. [55]

    G. F. Silva, T. P. Fagundes, B. C. Teixeira, and A. D. Chiavegatto Filho. Machine learning for hypertension prediction: a systematic review.Current hypertension reports, 24(11):523–533, 2022

    2022

  47. [56]

    Silva, G

    I. Silva, G. Moody, D. J. Scott, L. A. Celi, and R. G. Mark. Predicting in-hospital mortality of icu patients: The physionet/computing in cardiology challenge 2012. In2012 computing in cardiology, pages 245–248. IEEE, 2012

    2012

  48. [57]

    Spathis, I

    D. Spathis, I. Perez-Pozuelo, S. Brage, N. J. Wareham, and C. Mascolo. Self-supervised transfer learning of physiological representations from free-living wearable data. InProceedings of the Conference on Health, Inference, and Learning, pages 69–78, 2021

    2021

  49. [58]

    Srimedha, R

    B. Srimedha, R. N. Raj, and V. Mayya. A comprehensive machine learning based pipeline for an accurate early prediction of sepsis in icu.Ieee Access, 10:105120–105132, 2022

    2022

  50. [59]

    C. I. Tang, I. Perez-Pozuelo, D. Spathis, and C. Mascolo. Exploring contrastive learning in human activity recognition for healthcare.arXiv preprint arXiv:2011.11542, 2020

    2011 arXiv

  51. [60]

    G. Team, R. Anil, S. Borgeaud, J.-B. Alayrac, J. Yu, R. Soricut, J. Schalkwyk, A. M. Dai, A. Hauth, K. Millican, et al. Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805, 2023

    2023 arXiv

  52. [61]

    Tonekaboni, D

    S. Tonekaboni, D. Eytan, and A. Goldenberg. Unsupervised representation learning for time series with temporal neighborhood coding.arXiv preprint arXiv:2106.00750, 2021

    2021 arXiv

  53. [62]

    Truslow, A

    J. Truslow, A. Spillane, H. Lin, K. Cyr, A. Ullal, E. Arnold, R. Huang, L. Rhodes, J. Block, J. Stark, et al. Understanding activity and physiology at scale: The apple heart & movement study.npj Digital Medicine, 7(1):242, 2024

    2024

  54. [63]

    mice: Multivariate imputationbychained equations in r.Journal of statistical software, 45:1–67, 2011

    S.VanBuurenand K.Groothuis-Oudshoorn. mice: Multivariate imputationbychained equations in r.Journal of statistical software, 45:1–67, 2011

    2011

  55. [64]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    A.Vaswani, N.Shazeer, N.Parmar, J.Uszkoreit, L.Jones, A.N.Gomez, Ł.Kaiser, andI.Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    2017

  56. [65]

    H. Wei, M. A. Xu, C. Samplawski, J. M. Rehg, S. Kumar, and B. M. Marlin. Temporally multi-scale sparse self-attention for physical activity data imputation.Proceedings of machine learning research, 248:137, 2024

    2024

  57. [66]

    J. M.-T. Wu, M.-H. Tsai, S.-H. Xiao, and Y.-P. Liaw. A deep neural network electrocardiogram analysis framework for left ventricular hypertrophy prediction.Journal of Ambient Intelligence and Humanized Computing, pages 1–17, 2020. 15 LSM-2: Learning from Incomplete Wearable Se...

    2020

  58. [67]

    M. Wu, H. Cao, H.-L. Nguyen, K. Surmacz, and C. Hargrove. Modeling perceived stress via hrv and accelerometer sensor streams. In2015 37th annual international conference of the IEEE engineering in medicine and biology society (EMBC), pages 1625–1628. IEEE, 2015

    2015

  59. [68]

    M. Xu, A. Moreno, S. Nagesh, V. Aydemir, D. Wetter, S. Kumar, and J. M. Rehg. Pulseimpute: A novel benchmark task for pulsative physiological signal imputation.Advances in Neural Information Processing Systems, 35:26874–26888, 2022

    2022

  60. [69]

    M. A. Xu, A. Moreno, H. Wei, B. M. Marlin, and J. M. Rehg. Rebar: Retrieval-based reconstruction for time-series contrastive learning.arXiv preprint arXiv:2311.00519, 2023

    2023 arXiv

  61. [70]

    M. A. Xu, J. Narain, G. Darnell, H. Hallgrimsson, H. Jeong, D. Forde, R. Fineman, K. J. Raghuram, J. M. Rehg, and S. Ren. Relcon: Relative contrastive learning for a motion foundation model for wearable data.arXiv preprint arXiv:2411.18822, 2024

    2024 arXiv

  62. [71]

    Yilmaz, X

    G. Yilmaz, X. Lyu, J. L. Ong, L. H. Ling, T. Penzel, B. T. Yeo, and M. W. Chee. Nocturnal blood pressure estimation from sleep plethysmography using machine learning.Sensors, 23(18):7931, 2023

    2023

  63. [72]

    J. Yoon, J. Jordon, and M. Schaar. Gain: Missing data imputation using generative adversarial nets. InInternational conference on machine learning, pages 5689–5698. PMLR, 2018

    2018

  64. [73]

    H. Yuan, S. Chan, A. P. Creagh, C. Tong, A. Acquah, D. A. Clifton, and A. Doherty. Self-supervised learning for human activity recognition using 700,000 person-days of wearable data.NPJ digital medicine, 7(1):91, 2024

    2024

  65. [74]

    Zhang, Z

    X. Zhang, Z. Zhao, T. Tsiligkaridis, and M. Zitnik. Self-supervised contrastive pre-training for time series via time-frequency consistency.Advances in neural information processing systems, 35:3988–4003, 2022

    2022

  66. [75]

    Y. Zhou, J. Shi, R. Stein, X. Liu, R. N. Baldassano, C. B. Forrest, Y. Chen, and J. Huang. Missing data matter: an empirical evaluation of the impacts of missing ehr data in comparative effec- tiveness research.Journal of the American Medical Informatics Association, 30(7):124...

    2023

  67. [76]

    Evaluate the feasibility of using the data provided by wrist-worn wearable devices to develop algorithms and scores to assess metabolic health

    T. Zielinski, A. M. Moore, E. Troup, K. J. Halliday, and A. J. Millar. Strengths and limitations of period estimation methods for circadian data.PloS one, 9(5):e96462, 2014. 16 Appendix — LSM-2: Learning from Incomplete Sensor Data Table of Contents A.1 Data Details 18 A.1.1 I...

    2014

  68. [77]

    This is useful for modeling sensor noise, in which random channels at random times will be missing

    Random Imputation Pre-training: Here we drop out a % of total tokens. This is useful for modeling sensor noise, in which random channels at random times will be missing

  69. [78]

    Temporal Slice Pre-training: Here we drop out a % of total temporal slices, across all sensor channels. This is useful for modeling device off, in which, for a given period of time, all 22 LSM-2: Learning from Incomplete Wearable Sensor Data Figure 8| Gallery of Data Examples ...

  70. [79]

    my resting heart-rate has been elevated ever since I fell sick

    Sensor Slice Pre-training: Here we drop out a % of total sensor slices, across all time points. This is useful for modeling sensor off, in which a given sensor channel is off because of a non-random missingness mechanism that tells the device to turn off the channel (i.e. to s...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.