Pith. sign in

REVIEW 3 major objections 5 minor 52 references

Foundation models for movement data: Are they ready for prime-time?

T0 review · 3 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read Four accelerometer foundation models show no consistent advantage over supervised baselines on standard activity recognition in a 19-task benchmark.

desk verdict A genuinely useful accelerometer-FM benchmark, but the window-level bootstrap under the central 'no consistent advantage' claim is not subject-level and needs a major statistical fix. read the letter →

arxiv 2608.13316 v1 pith:GKRAJ43J submitted 2026-08-13 eess.SP cs.LG

Alexander Bräuer , Benjamin Cauchi , Nils Strodthoff This is my paper · ORCID
classification eess.SPcs.LG
keywords foundationmodelsaccelerometerdatahumanactivityrecognitionwearablesensorsbenchmarkingself-supervisedlearningstressdetectionsleepstaging
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 asks whether pretrained foundation models for accelerometer data are ready to replace or beat ordinary supervised models in health-monitoring and activity-recognition settings. To answer, it runs a like-for-like benchmark of four open-source movement foundation models against four supervised baselines across 19 tasks in five categories, using three evaluation modes and statistical significance testing. The central finding is task-dependent: on standard human activity recognition the foundation models show no consistent advantage, while they lead on fall and stress detection and on robustness to sensor placement; as frozen feature extractors, UniMTS is strongest. The authors conclude that for well-defined activity classes, cheap supervised baselines remain sufficient, and that foundation models are currently most useful in specific niches rather than as general replacements. The work matters because practitioners need evidence about when paying the pretraining and compute cost is worth it.

What carries the argument

The central mechanism is a standardized benchmark protocol that makes heterogeneous models comparable: all models are fed triaxial acceleration windows, adapted through wrappers to their native input lengths and sampling rates, and evaluated in three modes—linear probing (frozen features plus a linear head), frozen probing (frozen features plus a query-attention head), and full fine-tuning under a single uniform training recipe (AdamW, constant learning rate 0.0003, 100 epochs, early stopping). Statistical significance is decided by pairwise bootstrapping on the test set, so rank differences in the radar plots are not treated as noise. The representational analysis rests on cluster-based concept alignment and centered kernel alignment (CKA), which measure how similar different layers are within each model and how well discovered concepts line up with activity labels.

What would settle it

Re-run the 19-task benchmark with per-model hyperparameter search (for example, learning rates from 1e-4 to 1e-3) and give UniMTS and NormWear their native modalities (gyroscope and multimodal physiological signals); if UniMTS's fine-tuning rank moves from last to first, the no-consistent-advantage claim would need to be revised.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the promise of accelerometer foundation models as general-purpose feature extractors is not yet realized: across 19 tasks, no model, foundation or supervised, dominates, and on standard human activity recognition the four foundation models are statistically tied with or beaten by supervised baselines as often as they win. The exceptions are specific: Oxford SSL and ElderNet lead on fall and stress detection and across most sensor positions under fine-tuning; UniMTS provides the best frozen representations, the only foundation model that surpasses supervised baselines without fine-tuning; NormWear is strongest for demographics when paired with a non-linear head but costs roughly two orders of magnitude more than the efficient foundation models. Sleep staging stays near chance for every model, which the authors read as a modality limitation rather than a model-choice problem. Internally, all foundation models show high representational similarity across layers, suggesting their representations are less specialized than they could be.

Load-bearing premise

The rankings assume that one fixed fine-tuning recipe (learning rate 0.0003, AdamW, 100 epochs, early stopping) and a single input modality (triaxial acceleration) are fair to all eight models; if a model needs different optimization settings or its native sensors, its rank could change.

Editorial extensions

If this is right

  • For standard, well-defined activity classes, supervised CNN baselines such as Inception1D and XResNet1D match foundation-model performance at a fraction of the compute, so defaulting to a foundation model for human activity recognition is not currently justified.
  • When labeled data are available and the task is fall or stress detection, Oxford SSL and ElderNet are the strongest choices, with efficiency on par with lightweight baselines.
  • When labels are scarce and the model is used frozen, UniMTS is the best ready-to-use extractor overall, and NormWear is best for demographic attributes, but both need a non-linear read-out head to reach their potential.
  • For deployments where sensor position cannot be controlled, foundation models, led by ElderNet, outperform supervised baselines across six of seven body locations.
  • Sleep staging should not currently be attempted from accelerometer data alone with any of these models; all stay near chance, suggesting the bottleneck is the modality rather than the model choice.

Reading between the lines

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

  • The uniformly high inter-layer similarity in all four foundation models implies that future pretraining objectives should push layers to specialize; this is a concrete, model-agnostic direction the paper leaves implicit.
  • UniMTS's dramatic gap between best frozen probing and last-place fine-tuning suggests its representations are fragile under full fine-tuning, so a two-stage fine-tuning strategy is a direct, testable remedy.
  • If concept-based activity profiles prove valid against video annotations, foundation-model representations could move beyond fixed activity labels toward continuous behavior characterization, a research direction the paper identifies but does not develop.
  • Giving NormWear and UniMTS their native modalities (multimodal physiological signals for NormWear, gyroscope data for UniMTS) could change the rankings; the acceleration-only design is the conservative worst case for those two models.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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. This paper benchmarks four open-source accelerometer foundation models (NormWear, UniMTS, Oxford SSL, ElderNet) against four supervised baselines (TinyHAR, Inception1D, XResNet1D, S4) on 19 tasks spanning human activity recognition, fall detection, sleep staging, stress detection, and demographics, using 10 public datasets. Each model is evaluated in linear, frozen, and fine-tuned modes, with additional analyses of sensor-position robustness, computational efficiency, layerwise representational similarity (CKA/CBA), and concept discovery. The central claims are that foundation models show no consistent advantage over supervised baselines on standard HAR tasks, that selected FMs lead on fall and stress detection, that frozen FM representations are strongest in specific categories (notably UniMTS), and that supervised baselines remain sufficient and more efficient for well-defined activity classes. The paper concludes with scenario-based deployment recommendations and suggests activity-profile inference as a promising research direction.

Significance. If the conclusions are sustained, this would be a valuable benchmark: it is a like-for-like comparison of multiple open accelerometer FMs with strong supervised baselines across a broad task set, with publicly available code and detailed efficiency and representation analyses. The central 'no consistent advantage' claim and the deployment recommendations are directly relevant to practitioners. The main weaknesses are statistical: the ranking procedure ignores subject-level clustering, which undermines the significance-based claims, and some headline statements overstate the equivalence evidence. Because the benchmark infrastructure and the bulk of the descriptive results remain useful, the issues are fixable within a revision rather than fatal.

major comments (3)
  1. [Section II-C, 'Metrics and statistical significance'] The pairwise bootstrap resamples test-set windows rather than subjects. With per-dataset test sets of roughly 1–4 subjects (e.g., HAR70+ has 18 subjects and an 8:1:1 split, leaving about 2 test subjects), window-level resampling ignores between-subject variability and yields artificially narrow confidence intervals for AUROC and MAE differences. Consequently, the statements that 'differences in rank correspond to statistically significantly different performance' and the many 'statistically tied' entries in Tables III, V, VI, VII, and VIII are not established at the subject level. This directly affects the central conclusion of 'no consistent advantage' for HAR and specific rank claims such as 'Oxford SSL and ElderNet which both rank first on four datasets.' Please replace the window bootstrap with a subject-level or cluster bootstrap, apply a correction for the multiple pairwise comparisons, and either provide subject-level confidence intervals or remove significance-based language from the summaries.
  2. [Section III-C, 'Comparative assessment and recommendations for practitioners'] The statement that 'supervised baselines such as Inception1D and XResNet1D are sufficient' is an equivalence or non-inferiority claim, but the paper only reports failure to reject the null of no difference, and that failure is based on the invalid window-level bootstrap. With very small test sets, absence of evidence has little evidential weight. Please support the 'sufficient' claim with a proper equivalence test (e.g., two one-sided tests) or rephrase to 'we did not find consistent evidence that FMs outperform supervised baselines on these tasks.'
  3. [Section II-C, 'Training', and Table VI] The uniform fine-tuning protocol (AdamW with constant learning rate 0.0003, 100 epochs, early stopping) is applied to all models. If UniMTS, or NormWear, requires different optimization settings, its last-place fine-tuning rank—and the resulting 'no consistent advantage' conclusion for HAR—could be artifacts of the protocol rather than model capability. The paper itself notes UniMTS's sensitivity to fine-tuning in Section III-B1, but no sensitivity analysis is provided. Please run a minimal hyperparameter sensitivity check (e.g., vary the learning rate for UniMTS and one strong baseline) or explicitly qualify all conclusions that depend on the fine-tuning rankings.
minor comments (5)
  1. [Figure 1 and Table VI] The demographic task counts are inconsistent: Figure 1 labels Age (n=5), Weight (n=4), and Height (n=3), which sum to 12, while Table VI lists only 4 age, 3 weight, and 2 height tasks (9 demographic tasks total) and the abstract states 19 tasks overall. Please clarify which counts are correct and align the figure labels with the tables.
  2. [Table IV] The header 'GPU Memory [MB]↑' uses an upward arrow as if higher memory is better, but the text and standard usage treat lower memory as better. The arrow direction should be inverted to '↓'.
  3. [Section III-A1] The sentence 'UniMTS ranks last across the HAR task' is not apparent from Table VI point estimates, where UniMTS is not the lowest-scoring model on several datasets (e.g., PAMAP2, HAR70+, USC-HAD). Please rephrase to refer to the aggregated rank rather than per-dataset AUROC, or provide the rank table that supports the statement.
  4. [Section III-B3 and Figure 4] The text states that 'NLMCD discovers 34 concepts' from ElderNet's last-layer representations, while the Figure 4 caption says 'based on 107 concepts discovered by NLMCD.' These numbers should be reconciled.
  5. [References] Some references are incomplete, for example reference [8] lacks full bibliographic details (journal volume, pages, year), and reference [27] is missing the year and venue. Please complete the bibliographic entries.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: benchmark claims are empirical and self-contained; self-citations are contextual, not load-bearing.

full rationale

The paper's central claims are empirical benchmark results: it compares four open-source foundation models against four supervised baselines across 19 tasks using a fixed, openly described evaluation protocol. The 'no consistent advantage' finding on HAR is directly computed from test-set AUROC/MAE values and bootstrap-based rank ties, not derived from any fitted parameter, definition, or prior result that already contains the conclusion. No step in the derivation chain reduces to its own input by construction. The self-citations that appear in the manuscript are not load-bearing in the sense required for circularity. The NLMCD-ALIGN framework [23] is cited as the analysis tool for concept discovery and layerwise alignment; the authors of the present paper overlap with that citation, but the FM-versus-baseline rankings and the main task-performance conclusions are independent of whether one accepts NLMCD as a valid method. Likewise, [7] is cited as an example of recent benchmarking studies and [36], [43], [47] are cited as sources of baseline architectures or fine-tuning suggestions; none of these references supplies a premise that forces the paper's results. The paper is self-contained against external benchmark datasets and reports raw performance numbers, so there is no fitted input renamed as a prediction and no uniqueness theorem imported from the authors' own prior work. The only mild concern is methodological (e.g., the uniform fine-tuning protocol and window-level bootstrap ranks ignoring subject-level clustering), but that is a correctness/statistical issue, not circularity. Therefore the appropriate finding is no significant circularity, with a score of 1 reflecting the presence of non-load-bearing self-citations rather than any circular derivation.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The benchmark depends on hand-chosen training hyperparameters, the representativeness of the four included FMs, and the choice of accelerometer-only input. No new theoretical entities are introduced.

free parameters (4)
  • fine-tuning learning rate = 0.0003
    Hand-chosen uniform constant for all models; non-uniform tuning could change rankings.
  • fine-tuning weight decay = 0.001
    Uniform choice.
  • fine-tuning batch size = 128
    Uniform choice.
  • evaluation window length = 30 s
    Chosen for all tasks; affects segmentation of sleep and activity windows.
assumptions (4)
  • domain assumption The four open-source FMs (NormWear, UniMTS, Oxford SSL, ElderNet) are representative of current accelerometer foundation models.
    Section II-A notes SensorLM, LSM-2, and WBM were unavailable; if excluded models are substantially better, the conclusion 'not ready for prime-time' would not generalize.
  • domain assumption Triaxial acceleration is an adequate common input modality for like-for-like comparison.
    NormWear and UniMTS were pretrained on richer modalities (multimodal physiological signals and gyroscope data), so evaluating them on accelerometry alone may understate their potential (Section II-C and Section III-A.1).
  • domain assumption The soft multi-label target aggregation (per-class fraction of time in a window) preserves task information for all activities.
    Section II-C describes binarizing fractional targets for AUROC computation; this choice could affect benchmark results for windows with multiple simultaneous activities.
  • standard math Bootstrap significance testing with 1000 iterations on the test set yields reliable rank assignments.
    Pairwise bootstrapping assumes test-set independence; with small subject counts, bootstrap confidence intervals may be overconfident.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Foundation models for movement data: Are they ready for prime-time?." pith.science (2026). https://pith.science/paper/GKRAJ43J

@misc{pith2026260813316,
  author       = {Pith},
  title        = {Pith review of: Foundation models for movement data: Are they ready for prime-time?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GKRAJ43J}},
  note         = {Machine review of arXiv:2608.13316}
}
read the original abstract

Foundation models (FMs) trained on large-scale accelerometer data have been proposed as general-purpose feature extractors for health monitoring, but systematic evidence of their advantages is lacking. We present the first comprehensive evaluation of four open-source accelerometer FMs against supervised baselines covering 19 tasks across the domains of activity recognition including activities of daily living, clinical monitoring, and physiological inference. We find task-dependent performance results: supervised models remain competitive with FMs on human action recognition (HAR), with no consistent advantage for either, while selected FMs lead on fall and stress detection and are the most robust to sensor-placement variation. As frozen feature extractors, FMs are strongest for demographic inference, whereas sleep staging performance remains near chance level for all models. The internal FM representations show strong similarity across layers, highlighting potential for future FM improvements. Linear and frozen probing reveals that UniMTS provides the strongest representations and is the only FM that surpasses the supervised baselines without finetuning. Concept discovery analysis shows all models capture high-intensity activities clearly but struggle with sedentary, complex or ambiguous activities. We provide scenario-based deployment recommendations. Furthermore, we identify FM-derived activity profile inference-moving beyond fixed category classification-as a promising research direction.

Figures

Figures reproduced from arXiv: 2608.13316 by the authors.

Figure 1
Figure 1. Radar plots summarizing performance ranks for four FMs and four supervised baselines. For similar demographic tasks datasets were combined and the plots show the mean rank. Different ranks indicate statistically significant performance differences, equal ranks are assigned in case of statistically indistinguishable performances. Supervised models were trained only once from scratch but ranked in comparison to FMs fo… view at source ↗
Figure 2
Figure 2. Intra-model layerwise CBA on RealWorld for the finetuned FMs and baseline models. Each heatmap shows pairwise CBA scores between layers of the same model, lighter colors indicate higher concept similarities. L0 L1 L2 L3 L4 L5 L6 L7 L8 L9 L10 L11 Layer L0 L1 L2 L3 L4 L5 L6 L7 L8 L9 L10 L11 Layer NormWear L0 L1 L2 L3 L4 L5 L6 L7 L8 L9 Layer L0 L1 L2 L3 L4 L5 L6 L7 L8 L9 UniMTS L0 L1 L2 L3 L4 L5 L6 L7 L8 L9 Layer L0 L1… view at source ↗
Figure 3
Figure 3. Intra-model layerwise CBA on RealWorld for the FMs using their pretrained checkpoints. Each heatmap shows pairwise CBA scores between layers of the same model, lighter colors indicate higher concept similarities. investigated FMs (not shown).The public video recordings for RealWorld show that the climbing up and down activities consisted of walking bouts up-/downhill, up-/downstairs , turns, standing and mixed usage… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Concept-activity alignment for pretrained ElderNet on RealWorld, based on 107 concepts discovered by NLMCD from the last-layer representations. Left: UMAP of the discovered concepts. Center: concept activation proportions per ground-truth activity label (row-normalized…
Figure 5
Figure 5. Figure 5: Intra-model layerwise CKA of Foundation models on the RealWorld dataset using the pretrained checkpoints. Lighter colors indicate higher linear similarity between representations. Implementation of linear CKA followed [24]. This is the pretrained counterpart to the fin…
Figure 6
Figure 6. Figure 6: Intra-model CKA on RealWorld (wrist) for the four FMs and baseline models after finetuning. Lighter colors indicate higher linear similarity between the representations of layer pairs within the same model. Implementation of linear CKA followed [24] [PITH_FULL_IMAGE:f…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 36 canonical work pages

  1. [1]

    On the opportunities and risks of foundation models,

    R. Bommasaniet al., “On the opportunities and risks of foundation models,”arXiv preprint 2108.07258, 2022. [Online]. Available: https://arxiv.org/abs/2108.07258 AUTHORet al.: TITLE 9

    arXiv 2022

  2. [2]

    Emerging properties in self-supervised vision trans- formers,

    M. Caronet al., “Emerging properties in self-supervised vision trans- formers,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 9650–9660

    work page 2021

  3. [3]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” inPro- ceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), 2019, pp. 4171–4186

    2019

  4. [4]

    Towards a general-purpose foundation model for computational pathology,

    R. J. Chenet al., “Towards a general-purpose foundation model for computational pathology,”Nature medicine, vol. 30, no. 3, pp. 850–862, 2024

    work page 2024

  5. [5]

    Towards generalist foundation model for radiology by leveraging web-scale 2d&3d medical data,

    C. Wu, X. Zhang, Y . Zhang, H. Hui, Y . Wang, and W. Xie, “Towards generalist foundation model for radiology by leveraging web-scale 2d&3d medical data,”Nature Communications, vol. 16, no. 1, p. 7866, 2025

    2025

  6. [6]

    Neuro-gpt: Towards a foundation model for eeg,

    W. Cuiet al., “Neuro-gpt: Towards a foundation model for eeg,” in2024 IEEE International Symposium on Biomedical Imaging (ISBI). IEEE, 2024, pp. 1–5

    work page 2024

  7. [7]

    Benchmarking ECG foundational models: A reality check across clinical tasks,

    M. A. Al-Masud, J. L. Alcaraz, and N. Strodthoff, “Benchmarking ECG foundational models: A reality check across clinical tasks,” inThe Fourteenth International Conference on Learning Representations, 2026. [Online]. Available: https://openreview.net/forum?id=xXRqWpt3Xr

    work page 2026

  8. [8]

    Towards Customizable Foundation Models for Human Activity Recognition with Wearable Devices,

    M. Qiu, C. Weng, M. Fan, and K. Wu, “Towards Customizable Foundation Models for Human Activity Recognition with Wearable Devices,” vol. 9, no. 3, pp. 1–29. [Online]. Available: https: //dl.acm.org/doi/10.1145/3749479

    doi:10.1145/3749479

Show all 52 references
  1. [9]

    Collecting complex activity datasets in highly rich networked sensor environments,

    D. Roggenet al., “Collecting complex activity datasets in highly rich networked sensor environments,” in2010 Seventh International Conference on Networked Sensing Systems (INSS), pp. 233–240. [Online]. Available: https://ieeexplore.ieee.org/document/5573462/

  2. [10]

    Reiss and D

    A. Reiss and D. Stricker,Introducing a New Benchmarked Dataset for Activity Monitoring

  3. [11]

    On-body localization of wearable devices: An investigation of position-aware activity recognition,

    T. Sztyler and H. Stuckenschmidt, “On-body localization of wearable devices: An investigation of position-aware activity recognition,” in2016 IEEE International Conference on Pervasive Computing and Communications (PerCom), pp. 1–9. [Online]. Available: https: //ieeexplore.iee...

  4. [12]

    SisFall: A Fall and Movement Dataset,

    A. Sucerquia, J. D. L ´opez, and J. F. Vargas-Bonilla, “SisFall: A Fall and Movement Dataset,”Sensors (Basel, Switzerland), vol. 17, no. 1, p. 198, Jan. 2017

    2017

  5. [13]

    A large-scale open motion dataset (kfall) and benchmark algorithms for detecting pre- impact fall of the elderly using wearable inertial sensors,

    X. Yu, J. Jang, and S. Xiong, “A large-scale open motion dataset (kfall) and benchmark algorithms for detecting pre- impact fall of the elderly using wearable inertial sensors,” vol. 13. [Online]. Available: https://www.frontiersin.org/journals/ aging-neuroscience/articles/10....

  6. [14]

    A Machine Learning Model for Predicting Sleep and Wakefulness Based on Accelerometry, Skin Temperature and Contextual Information,

    A. Logacjov, E. Skarpsno, A. Kongsvold, K. Bach, and P. J. Mork, “A Machine Learning Model for Predicting Sleep and Wakefulness Based on Accelerometry, Skin Temperature and Contextual Information,” vol. V olume 16, pp. 699–710

  7. [15]

    Introducing WESAD, a Multimodal Dataset for Wearable Stress and Affect Detection,

    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. ACM, pp. 400–408. [Online]. Available: ...

  8. [16]

    Wearable technologies in clinical trials for drug development: Trends and emerging opportunities,

    Z. A. Fayad, R. P. Hirten, G. N. Nadkarni, and Y . S. Kim, “Wearable technologies in clinical trials for drug development: Trends and emerging opportunities,” vol. 25, no. 6, pp. 448–468. [Online]. Available: https://www.nature.com/articles/s41573-026-01403-9

  9. [17]

    Sensor-based digital health technologies to capture endpoints in recent clinical trials: A scoping review

    J. Garciaet al., “Sensor-based digital health technologies to capture endpoints in recent clinical trials: A scoping review.” [Online]. Available: https://www.nature.com/articles/s41746-026-02512-w

  10. [18]

    Digital measures of activity and motivation impact depression and anxiety in the real world,

    J. M. Beltr ´anet al., “Digital measures of activity and motivation impact depression and anxiety in the real world,”npj Digital Medicine, vol. 8, no. 1, p. 268, 2025

    2025

  11. [19]

    Self-supervised Learning for Accelerometer-based Human Activity Recognition: A Survey,

    A. Logacjov, “Self-supervised Learning for Accelerometer-based Human Activity Recognition: A Survey,” vol. 8, no. 4, pp. 149:1–149:42. [Online]. Available: https://dl.acm.org/doi/10.1145/3699767

    doi

  12. [20]

    SelfHAR: Improving Human Activity Recognition through Self-training with Unlabeled Data,

    C. I. Tang, I. Perez-Pozuelo, D. Spathis, S. Brage, N. Wareham, and C. Mascolo, “SelfHAR: Improving Human Activity Recognition through Self-training with Unlabeled Data,” vol. 5, no. 1, pp. 1–30. [Online]. Available: https://dl.acm.org/doi/10.1145/3448112

    doi

  13. [21]

    Bianet al.Foundation Models Defining A New Era In Sensor-based Human Activity Recognition: A Survey And Outlook

    S. Bianet al.Foundation Models Defining A New Era In Sensor-based Human Activity Recognition: A Survey And Outlook. [Online]. Available: http://arxiv.org/abs/2604.02711

    arXiv

  14. [22]

    Benchmarking foundation models as feature extrac- tors for weakly supervised computational pathology,

    P. Neidlingeret al., “Benchmarking foundation models as feature extrac- tors for weakly supervised computational pathology,”Nature biomedical engineering, pp. 1–11, 2025

    2025

  15. [23]

    Beyond scalars: Concept-based alignment analysis in vision transformers,

    J. Vielhaben, D. Bareeva, J. Berend, W. Samek, and N. Strodthoff, “Beyond scalars: Concept-based alignment analysis in vision transformers,” inAdvances in Neural Information Processing Systems, D. Belgraveet al., Eds., vol. 38. Curran Associates, Inc., 2025, pp. 71 379–71 414....

    2025

  16. [24]

    Similarity of neural network representations revisited,

    S. Kornblith, M. Norouzi, H. Lee, and G. E. Hinton, “Similarity of neural network representations revisited,” inInternational Conference on Machine Learning. PMLR, 2019, pp. 3519–3529. [Online]. Available: http://proceedings.mlr.press/v97/kornblith19a.html

    2019

  17. [25]

    SensorLM: Learning the Language of Wearable Sensors

    Y . Zhanget al., “SensorLM: Learning the Language of Wearable Sensors.” [Online]. Available: http://arxiv.org/abs/2506.09108

    arXiv

  18. [26]

    LSM-2: Learning from Incomplete Wearable Sensor Data,

    M. A. Xuet al., “LSM-2: Learning from Incomplete Wearable Sensor Data,”arXiv preprint 2506.05321

    arXiv

  19. [27]

    Beyond sensor data: Foundation models of behavioral data from wearables improve health predictions,

    E. Erturket al., “Beyond sensor data: Foundation models of behavioral data from wearables improve health predictions,” inForty-second Inter- national Conference on Machine Learning, 2025

    2025

  20. [28]

    Generating Diverse and Natural 3D Human Motions from Text,

    C. Guoet al., “Generating Diverse and Natural 3D Human Motions from Text,” in2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5142–5151. [Online]. Available: https://ieeexplore.ieee.org/document/9880214

  21. [29]

    Large scale population assessment of physical activity using wrist worn accelerometers: The uk biobank study,

    A. Dohertyet al., “Large scale population assessment of physical activity using wrist worn accelerometers: The uk biobank study,”PLOS ONE, vol. 12, no. 2, p. e0169649, Feb. 2017. [Online]. Available: http://dx.doi.org/10.1371/journal.pone.0169649

    2017 doi

  22. [30]

    Overview and Findings from the Rush Memory and Aging Project,

    D. A. Bennett, J. A. Schneider, A. S. Buchman, L. L. Barnes, P. A. Boyle, and R. S. Wilson, “Overview and Findings from the Rush Memory and Aging Project,” vol. 9, no. 6, pp. 646–663. [Online]. Available: https://pmc.ncbi.nlm.nih.gov/articles/PMC3439198/

  23. [31]

    V-jepa: Latent video prediction for visual representation learning,

    A. Bardes, M. Mirza, B. Oreshkin, M. Auli, I. Misra, and Y . LeCun, “V-jepa: Latent video prediction for visual representation learning,” in International Conference on Learning Representations (ICLR), 2024. [Online]. Available: https://openreview.net/forum?id=WFYbBOEOtv

    2024

  24. [32]

    Ptb-xl, a large publicly available electrocardiography dataset,

    P. Wagneret al., “Ptb-xl, a large publicly available electrocardiography dataset,”Scientific data, vol. 7, no. 1, pp. 1–15, 2020. [Online]. Available: https://doi.org/10.1038/s41597-020-0495-6

    2020 doi

  25. [33]

    On the Stratification of Multi-label Data,

    K. Sechidis, G. Tsoumakas, and I. Vlahavas, “On the Stratification of Multi-label Data,” inMachine Learning and Knowledge Discovery in Databases, D. Gunopulos, T. Hofmann, D. Malerba, and M. Vazirgian- nis, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg, 2011, vol. 6913, ...

    2011

  26. [34]

    Umap: Uniform manifold approximation and projection for dimension reduction,

    L. McInnes, J. Healy, and J. Melville, “Umap: Uniform manifold approximation and projection for dimension reduction,”arXiv preprint arXiv:1802.03426, 2018

    2018 arXiv

  27. [35]

    hdbscan: Hierarchical density based clustering

    L. McInnes, J. Healy, S. Astelset al., “hdbscan: Hierarchical density based clustering.”J. Open Source Softw., vol. 2, no. 11, p. 205, 2017

    2017

  28. [36]

    Self-supervised representation learning from 12-lead ECG data,

    T. Mehari and N. Strodthoff, “Self-supervised representation learning from 12-lead ECG data,”Computers in Biology and Medicine, vol. 141, p. 105114, 2022. [Online]. Available: https://doi.org/10.1016/j. compbiomed.2021.105114

    2022

  29. [37]

    Zhanget al.UniMTS: Unified Pre-training for Motion Time Series

    X. Zhanget al.UniMTS: Unified Pre-training for Motion Time Series. [Online]. Available: http://arxiv.org/abs/2410.19818

    arXiv

  30. [38]

    Toward Foundation Model for Multivariate Wearable Sensing of Physiological Signals,

    Y . Luo, Y . Chen, A. Salekin, and T. Rahman, “Toward Foundation Model for Multivariate Wearable Sensing of Physiological Signals,” Dec. 2024

    2024

  31. [39]

    Self-supervised learning for human activity recognition using 700,000 person-days of wearable data,

    H. Yuanet al., “Self-supervised learning for human activity recognition using 700,000 person-days of wearable data,”npj Digital Medicine, vol. 7, no. 1, pp. 1–10, Apr. 2024

    2024

  32. [40]

    Assessing real-world gait with digital technology? validation, insights and recommendations from the mobilise-d consortium,

    M. E. Mic ´o-Amigoet al., “Assessing real-world gait with digital technology? validation, insights and recommendations from the mobilise-d consortium,”Journal of NeuroEngineering and Rehabilitation, vol. 20, no. 1, Jun. 2023. [Online]. Available: http://dx.doi.org/10.1186/s129...

    2023 doi

  33. [41]

    Self-supervised learning of wrist-worn daily living accelerometer data improves the automated detection of gait in older adults,

    Y . E. Brandet al., “Self-supervised learning of wrist-worn daily living accelerometer data improves the automated detection of gait in older adults,”Scientific Reports, vol. 14, no. 1, p. 20854, Sep. 2024

    2024

  34. [42]

    Tinyhar: A lightweight deep learning model designed for human activity recognition,

    Y . Zhou, H. Zhao, Y . Huang, M. Hefenbrock, T. Riedel, and M. Beigl, “Tinyhar: A lightweight deep learning model designed for human activity recognition,” inInternational Symposium on Wearable Computers (ISWC’22) , Atlanta, GA and Cambridge, UK, September 11-15, 2022, p. 89. ...

    2022

  35. [43]

    Deep learning for ecg analysis: Benchmarks and insights from ptb-xl,

    N. Strodthoff, P. Wagner, T. Schaeffter, and W. Samek, “Deep learning for ecg analysis: Benchmarks and insights from ptb-xl,”IEEE journal 10 IEEE TRANSACTIONS AND JOURNALS TEMPLATE of biomedical and health informatics, vol. 25, no. 5, pp. 1519–1528, 2020

    2020

  36. [44]

    Machine-learning for photoplethysmography analysis: Benchmarking feature, image, and signal-based approaches,

    M. Moulaeifardet al., “Machine-learning for photoplethysmography analysis: Benchmarking feature, image, and signal-based approaches,” arXiv preprint arXiv:2502.19949, 2025

    2025

  37. [45]

    Inceptiontime: Finding alexnet for time series classification,

    H. Ismail Fawazet al., “Inceptiontime: Finding alexnet for time series classification,”Data Mining and Knowledge Discovery, vol. 34, no. 6, pp. 1936–1962, 2020

    1936

  38. [46]

    Efficiently modeling long sequences with structured state spaces,

    A. Gu, K. Goel, and C. Re, “Efficiently modeling long sequences with structured state spaces,” inInternational Conference on Learning Representations, 2021

    2021

  39. [47]

    S4sleep: Elucidating the design space of deep-learning-based sleep stage classification models,

    T. Wang and N. Strodthoff, “S4sleep: Elucidating the design space of deep-learning-based sleep stage classification models,”Computers in Biology and Medicine, vol. 187, p. 109735, 2025

    2025

  40. [48]

    Validation of an Activity Type Recognition Model Classifying Daily Physical Behavior in Older Adults: The HAR70+ Model,

    A. Ustadet al., “Validation of an Activity Type Recognition Model Classifying Daily Physical Behavior in Older Adults: The HAR70+ Model,” vol. 23, no. 5, p. 2368. [Online]. Available: https://www.mdpi.com/1424-8220/23/5/2368

  41. [49]

    HARTH: A Human Activity Recognition Dataset for Machine Learn- ing,

    A. Logacjov, K. Bach, A. Kongsvold, H. B. B ˚ardstu, and P. J. Mork, “HARTH: A Human Activity Recognition Dataset for Machine Learn- ing,”Sensors, vol. 21, no. 23, p. 7853, Nov. 2021

    2021

  42. [50]

    USC-HAD: A daily activity dataset for ubiquitous activity recognition using wearable sensors,

    M. Zhang and A. A. Sawchuk, “USC-HAD: A daily activity dataset for ubiquitous activity recognition using wearable sensors,” inProceedings of the 2012 ACM Conference on Ubiquitous Computing, 2012

    2012

  43. [51]

    Activity recognition using cell phone accelerometers,

    J. R. Kwapisz, G. M. Weiss, and S. A. Moore, “Activity recognition using cell phone accelerometers,” vol. 12, no. 2, pp. 74–82. [Online]. Available: https://dl.acm.org/doi/10.1145/1964897.1964918

  44. [52]

    WISDM Smartphone and Smartwatch Activity and Biometrics Dataset

    G. M. Weiss, “WISDM Smartphone and Smartwatch Activity and Biometrics Dataset.” APPENDIX A. Foundation Models We investigate four foundation models, which are described below and also summarized in Table I. UniMTSThe Unified pre-trained model for Motion Time Series (UniMTS) ad...

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.