Pith. sign in

REVIEW 3 major objections 8 minor 43 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

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

T0 review · glm-5.2

Step counts alone predict 21 health risks, beating richer sensor models

2026-07-09 01:23 UTC pith:LTS6VB2D

load-bearing objection Step-count-only foundation model with a real cohort-overlap problem between pre-training and primary evaluation the 3 major comments →

arxiv 2607.06954 v1 pith:LTS6VB2D submitted 2026-07-08 cs.LG

Physical activities enable scalable foundation modelling for broad-spectrum health prediction

Zhenghuang Wu , Yuyao Zhu , Songlin Xu This is my paper
classification cs.LG
keywords healthacrossfoundationdatamodelspredictionscalablestep
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper claims that minute-level step-count data, the simplest and most widely available output of any wearable device, is sufficient to build a foundation model for broad-spectrum health risk prediction. The authors train a 3.4-million-parameter model called StepFM on 141 million minute-level step observations from roughly 14,000 individuals, using a dual-stream architecture that simultaneously models hourly macro-patterns and intra-hour micro-morphology. The core argument is that the temporal dynamics and circadian structure embedded in step sequences already encode rich, generalizable behavioral phenotypes, and that a model pre-trained to predict future step tokens while aligning with hierarchical activity statistics can transfer to over 20 diverse health prediction tasks, spanning cardiovascular, metabolic, respiratory, neurological, and mental health domains. The paper reports that StepFM achieves a mean AUROC of 0.7318 across 21 health tasks, outperforming wearable foundation models that use richer sensor data such as IMU, PPG, and ECG signals, and that it generalizes across device types (ActiGraph, Fitbit, ScanWatch), geographical regions (US, Switzerland, UK), and disease categories unseen during pre-training (MS disability, anxiety). The authors further claim that the predictability of any given disease from step data is governed by its intrinsic behavioral correlation with physical activity rather than by model architecture, and that their approach preserves privacy and computational efficiency relative to raw-signal models.

Core claim

The central discovery is that a compact, low-dimensional signal, step counts per minute, carries enough behavioral and physiological information to support a single pre-trained model that predicts a broad spectrum of health risks, and that this model can outperform richer-sensor foundation models on the same tasks. The paper identifies a dual-stream encoding strategy, combining hourly macro tokens with minute-level micro-features via FiLM modulation and hierarchical phenotype alignment, as the mechanism that lets a small model extract transferable health representations from sparse step data. A secondary discovery is that the cross-disease performance profile is remarkably consistent across迥

What carries the argument

The model uses a log-scaled tokenizer that maps hourly step sums to 256 discrete tokens, allocating higher resolution to low-to-moderate activity ranges. A dual-stream Step-Mamba encoder processes hourly macro tokens through a Mamba state-space backbone while a 1D convolutional micro-stream extracts intra-hour activity morphology. FiLM modulation injects micro-stream features into the macro hidden states. Pre-training combines autoregressive next-token prediction with a hierarchical phenotype alignment loss that forces intermediate representations to encode hourly, daily, and weekly activity statistics. Temporal rhythm encoding via Fourier features for time-of-day and day-of-week is added to

Load-bearing premise

The primary evaluation uses 20 of 21 tasks drawn from the same NHANES 2011-2014 cohort that provided the pre-training data, meaning the model's representations may encode cohort-specific behavioral patterns rather than generalizable physiological signals. The external validation datasets are small (44 and 72 participants), limiting the strength of the generalization claim.

What would settle it

If a model trained on step counts from a genuinely independent cohort, with no participant overlap between pre-training and evaluation, failed to outperform hand-crafted step features or simple logistic regression on the same health tasks, the foundation-model claim would be weakened. Similarly, if the cross-disease performance profile shifted substantially when evaluated on a cohort with different lifestyle or cultural activity patterns, the claim of generalizable behavioral phenotypes would be undercut.

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

If this is right

  • If step counts alone can match or exceed richer sensor data for health prediction, clinical and consumer health monitoring could shift toward minimal-data pipelines that are cheaper, more private, and deployable on any phone or pedometer.
  • The finding that disease predictability is bounded by the behavioral correlation between a condition and physical activity suggests a principled way to triage which health outcomes are suitable for step-based screening versus which require additional modalities.
  • The data efficiency results, 95% of full-shot performance with 30% of labels, imply that pre-trained step models could enable screening for rare conditions where annotated clinical data is scarce.
  • The cross-device and cross-region transfer results, if confirmed on larger external cohorts, would support the feasibility of globally deployable health risk models trained on a single population's step data.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 8 minor

Summary. The paper proposes StepFM, a foundation model pre-trained solely on minute-level step-count data from NHANES 2011–2014 (141.12M observations, ~14,000 individuals) and evaluated across 21 health risk prediction tasks. The model uses a dual-stream Mamba architecture with log-scaled tokenization, temporal rhythm encoding, a micro-stream FiLM module, and hierarchical activity phenotype alignment. StepFM achieves a mean AUROC of 0.7318 on the primary 21-task benchmark, outperforming wearable foundation models (NormWear: 0.7079) and a traditional ML baseline (0.7065). External validation is conducted on NHANES 2005–2006 (different cohort, same device family), BarKA-MS (Swiss, Fitbit, MS-related endpoints), and RESILIENT (UK, ScanWatch, psychological/cognitive endpoints). The central claim is that compact, privacy-preserving step-count data can serve as a viable substrate for broad-spectrum health foundation models.

Significance. The paper addresses a practically important question: whether low-dimensional step-count data alone can support a generalizable health foundation model, as opposed to high-frequency raw sensor signals. The dual-stream architecture (macro hourly tokens with micro intra-hour FiLM modulation) is a reasonable design for the multi-scale structure of step data. The ablation study (Table 2) provides incremental evidence for each component. The cross-device, cross-region, and novel-disease evaluations on three external datasets are a strength of the experimental design. The layer-wise representation analysis (Figure 3) and scaling experiments (Figure 2) add useful diagnostic depth. The finding that a hand-crafted feature baseline (Trad. ML, 0.7065) is competitive with several foundation models is an honest and informative result that contextualizes the gains.

major comments (3)
  1. §Methods, Dataset (Pre-Training Datasets) and §Methods, Dataset (Fine-Tuning Datasets): The pre-training corpus is NHANES 2011–2014 step data from ~14,000 individuals, and the primary downstream evaluation (Table 1, 20 of 21 tasks) uses health labels from the same NHANES 2011–2014 participants linked by SEQN. The paper states pre-training uses 'unlabeled' data, which is technically accurate (no health labels are exposed during pre-training), but the model has observed the exact input step-count sequences of every downstream test individual during self-supervised pre-training. With 3.4M parameters and 168 hourly tokens per participant, the model has sufficient capacity to encode participant-specific behavioral signatures. The headline AUROC of 0.7318 (Table 1) may therefore partly reflect individual-specific memorization rather than generalizable representations. This is the central load-
  2. §Methods, Dataset (Pre-Training Datasets) and §Methods, Dataset (Fine-Tuning Datasets) [continued]: bearing concern for the paper's central claim. The external validations partially address this: NHANES 2005–2006 (Table 3) shows a drop to 0.6842 mean AUROC (a 4.8-point decrease), which is consistent with partial cohort overfitting. BarKA-MS (44 participants) and RESILIENT (72 participants) are too small to rule out chance variation. The paper should either (a) split the NHANES 2011–2014 cohort so that pre-training and downstream evaluation use disjoint participants, or (b) explicitly acknowledge this shared-cohort design as a limitation and reframe the headline result accordingly, emphasizing the external validation results as the primary evidence for generalization. As currently presented, the headline claim of generalizable representations rests primarily on a shared-cohort evaluation.
  3. Tables 4, 5, and 6: The model is referred to as 'Prancer' in these tables, while it is called 'StepFM' everywhere else in the paper (title, abstract, Table 1, Table 2, main text). This appears to be an inconsistency from an earlier version of the manuscript. The name should be corrected throughout.
minor comments (8)
  1. §Results (Comparison with Existing Foundation Models): The paper states StepFM 'ranks first in 20 out of 21 tasks.' In Table 1, NormWear outperforms StepFM on Angina (0.7565 vs. 0.7524), so StepFM ranks first in 20 of 21 tasks. This is correct, but the text should clarify that the one exception is Angina, for consistency.
  2. Table 5 (BarKA-MS): The F1 scores for all models on both tasks are very high (0.83–0.89), which is unusual for a health prediction dataset with 44 participants. The paper attributes this to class imbalance with a dominant positive class, but the prevalence rates for the two tasks are not reported. Adding prevalence information would help readers interpret these F1 values.
  3. Table 6 (RESILIENT): StepFM loses to SSCP on 3 of 8 tasks (Essential Hypertension, Geriatric Depression, Sleep Disturbance). The text states StepFM 'outperforms SSCP at 0.6673 and 0.6166, respectively,' which is accurate for the mean, but the per-task losses should be acknowledged in the generalization narrative.
  4. §Methods (Implementation Details): The text states '6 Mamba layers,' but the FiLM description says modulation is applied 'before Mamba layers 0, 2 and 4.' If layers are 0-indexed, this implies layers 0–5 (6 layers), which is consistent, but the 0-indexing convention should be clarified.
  5. §Methods (Hierarchical Activity Phenotype Alignment): The specific phenotype targets (e.g., 'log step volume, active/sedentary ratios, and step entropy' for hourly; 'cross-day consistency, variability, and total activity burden' for weekly) are described qualitatively but not formally defined. Providing the exact formulas or a table of all phenotype targets would aid reproducibility.
  6. §Methods (Data Preparation): The choice of S_max = 6000 as the physiological saturation threshold is stated without justification. A brief note on why this value was chosen (e.g., based on the distribution of hourly step counts in the dataset) would be helpful.
  7. Figure 1: The disease labels on the x-axis are difficult to read in the current layout. Consider rotating labels or using a numbered legend with a separate key to improve legibility in the final version.
  8. The abstract states 'more than 20 health risk prediction tasks,' which is accurate (21 tasks), but the specific number 21 should be stated in the abstract for precision, or the phrase 'more than 20' should be used consistently throughout.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive review. The referee raises three points: (1) a shared-cohort concern regarding pre-training and downstream evaluation on the same NHANES 2011–2014 participants, (2) a request to either implement disjoint splits or reframe the headline claims, and (3) a naming inconsistency ('Prancer' vs. 'StepFM') in Tables 4–6. We address each below.

read point-by-point responses

  1. Referee: The pre-training corpus (NHANES 2011–2014) and the primary downstream evaluation (Table 1, 20 of 21 tasks) use the same participants. The model has observed the exact input step-count sequences of every downstream test individual during self-supervised pre-training. With 3.4M parameters and 168 hourly tokens per participant, the model has sufficient capacity to encode participant-specific behavioral signatures. The headline AUROC of 0.7318 may partly reflect individual-specific memorization rather than generalizable representations.

    Authors: We agree this is a legitimate and important concern. The referee is correct that, although no health labels are exposed during pre-training, the input step-count sequences of downstream test individuals are seen during self-supervised pre-training. We acknowledge that a 3.4M-parameter model operating on 168 hourly tokens per participant has sufficient capacity to encode participant-specific behavioral signatures, and that the headline AUROC of 0.7318 on Table 1 may partly reflect this shared-cohort design rather than purely generalizable representations. We will implement a disjoint-participant split in which pre-training and downstream evaluation use non-overlapping subsets of NHANES 2011–2014, and report the resulting performance. We will also explicitly acknowledge the shared-cohort limitation in the revised manuscript and reframe the presentation so that the external validation results (NHANES 2005–2006, BarKA-MS, RESILIENT) are emphasized as the primary evidence for generalization. We note that the external validations already provide meaningful evidence against pure memorization: NHANES 2005–2006 uses a different cohort and device placement, BarKA-MS introduces an entirely unseen disease domain (MS-related disability and fatigue) on a different device (Fitbit) in a different country (Switzerland), and RESILIENT evaluates on yet another device (ScanWatch) and country (UK) with novel psychological and cognitive endpoints. StepFM outperforms all baselines on each of these external datasets, which would not be expected if performance were driven primarily by participant-specific memorization. Nevertheless, we agree that the disjoint-split experiment is the cleanest way to address this concern and will include it in the revision. revision: yes

  2. Referee: The paper should either (a) split the NHANES 2011–2014 cohort so that pre-training and downstream evaluation use disjoint participants, or (b) explicitly acknowledge this shared-cohort design as a limitation and reframe the headline result accordingly, emphasizing the external validation results as the primary evidence for generalization. As currently presented, the headline claim of generalizable representations rests primarily on a shared-cohort evaluation. The external validations partially address this: NHANES 2005–2006 shows a drop to 0.6842 mean AUROC (a 4.8-point decrease), which is consistent with partial cohort overfitting. BarKA-MS (44 participants) and RESILIENT (72 participants) are too small to rule out chance variation.

    Authors: We accept this point and will implement option (a): a disjoint-participant split for NHANES 2011–2014, ensuring that no individual whose step-count sequences were used during pre-training appears in the downstream evaluation set. We will report the revised Table 1 results on this disjoint split. We will also implement option (b) in parallel: the revised manuscript will explicitly acknowledge the shared-cohort design as a limitation of the current Table 1 results and will reframe the external validation results as the primary evidence for generalization. Regarding the referee's observation about the 4.8-point AUROC decrease on NHANES 2005–2006: we agree this is consistent with some degree of cohort-specific overfitting, though we note that a portion of this decrease is also attributable to the device placement shift (wrist-to-waist ActiGraph) and the earlier cohort period (2005–2006 vs. 2011–2014). We also agree that BarKA-MS (44 participants) and RESILIENT (72 participants) are too small to rule out chance variation on their own; however, the consistency of StepFM's advantage across all three external datasets, each introducing different devices, regions, and disease domains, provides convergent evidence for generalization that goes beyond what any single small dataset could establish. We will add a discussion of these sample-size limitations in the revised manuscript. revision: yes

  3. Referee: The model is referred to as 'Prancer' in Tables 4, 5, and 6, while it is called 'StepFM' everywhere else in the paper. This appears to be an inconsistency from an earlier version of the manuscript. The name should be corrected throughout.

    Authors: The referee is correct. Tables 4, 5, and 6 retain the earlier internal name 'Prancer' instead of 'StepFM.' This is a remnant from an earlier version of the manuscript and will be corrected to 'StepFM' throughout all tables and captions in the revised version. We thank the referee for catching this inconsistency. revision: yes

Circularity Check

0 steps flagged

No circularity found: pre-training is self-supervised on step sequences, downstream health labels are never used as pre-training targets, and no self-citation chain bears the central claim.

full rationale

The paper's derivation chain is self-contained and does not reduce to its inputs by construction. Pre-training uses two objectives: (1) autoregressive next-token prediction on hourly step-count tokens, and (2) hierarchical phenotype alignment, where phenotype targets are step-derived statistics (log step volume, active/sedentary ratios, step entropy, daily/weekly aggregates) computed from the same step sequences — not health labels. The downstream health predictions come from a separate fine-tuning step using a frozen backbone + MLP head on questionnaire-derived disease labels. No health label appears in the pre-training loss, so the 'prediction' of health outcomes is not equivalent to a fitted health input by construction. The paper does not invoke any uniqueness theorem, does not rely on self-citation for its central premise (all cited work — Mamba, TimeSiam, NHANES, etc. — is by external authors), and does not rename a known empirical result as a derivation. The reader's concern about pre-training and primary evaluation sharing the same NHANES 2011–2014 participant pool is a legitimate generalization/data-leakage risk, but it is not circularity: the pre-training objective is genuinely unsupervised with respect to the downstream labels, and the paper provides external validation on three independent cohorts (NHANES 05–06, BarKA-MS, RESILIENT) with different devices, regions, and novel disease types. The fact that StepFM outperforms a traditional ML baseline that cannot memorize individual sequences (0.7318 vs 0.7065) on the shared cohort, and still leads on the external NHANES 05–06 cohort (0.6842 vs 0.6640), provides independent evidence that is not forced by construction.

Axiom & Free-Parameter Ledger

7 free parameters · 3 axioms · 2 invented entities

The model introduces several design choices (vocabulary size, saturation threshold, architecture dimensions) that are selected by domain knowledge rather than fitted to data. The phenotype alignment weight and FiLM modulation bound are not specified. The central domain assumption—that step counts encode sufficient signal for broad health prediction—is supported by epidemiological literature but not previously tested at foundation-model scale. No new physical entities or forces are postulated; the contributions are architectural and methodological.

free parameters (7)
  • Vocabulary size V = 256
    Chosen for the log-scaled tokenizer; not fitted to data but selected by design.
  • Saturation threshold S_max = 6000
    Physiological saturation threshold for hourly step counts; chosen by domain knowledge.
  • Phenotype alignment weight λ_phenotype = not specified
    Controls contribution of phenotype alignment loss; value not stated in the paper.
  • FiLM modulation bound α = not specified
    Bounds the modulation strength in FiLM; value not stated.
  • Learning rate peak = 3e-4
    Standard hyperparameter, not a free parameter in the theoretical sense.
  • Number of Mamba layers = 6
    Architecture choice.
  • Hidden dimension = 256
    Architecture choice.
axioms (3)
  • domain assumption Step counts encode sufficient behavioral signal to predict a broad spectrum of health risks
    This is the central premise of the paper, supported by epidemiological literature (refs 8, 21-25, 30) but not previously demonstrated at foundation-model scale.
  • domain assumption Self-reported NHANES questionnaire labels are reliable ground truth for health conditions
    All 21 primary downstream tasks use self-reported health conditions from NHANES questionnaires. Self-reported labels have known noise and bias.
  • domain assumption Behavioral representations learned on a US population transfer to Swiss and UK populations
    External validation on BarKA-MS (Switzerland) and RESILIENT (UK) tests this assumption, but with very small samples (44 and 72 participants).
invented entities (2)
  • Activity phenotype alignment objective independent evidence
    purpose: Auxiliary pre-training loss aligning latent representations with hourly, daily, and weekly activity statistics
    The ablation (Table 2) shows it contributes +0.004 AUROC. The targets are computed from raw step data, not learned. This is a training objective, not a postulated physical entity.
  • Log-scaled step tokenizer independent evidence
    purpose: Discretizes continuous step counts into 256 vocabulary tokens with higher resolution for low-moderate activity
    Ablation shows tokenization + temporal encoding contributes +0.009 AUROC. The tokenizer formula is parameter-free given S_max.

pith-pipeline@v1.1.0-glm · 20653 in / 2740 out tokens · 486825 ms · 2026-07-09T01:23:11.984366+00:00 · methodology

0 comments
read the original abstract

Wearable and mobile sensing technologies have demonstrated strong potential for health inference; however, most sensor models are designed for specific disease types, limiting their transferability across different health risks. Wearable foundation models offer a more generalizable approach in diverse health risk types. Nevertheless, most existing methods rely on high-frequency raw sensor data, raising concerns about privacy, computational overhead, and scalability across devices and populations. In this paper, we propose StepFM, a foundation model built solely on step counter data for broad-spectrum health prediction. Leveraging the ubiquity and low-dimensional nature of step data, StepFM provides a practical, privacy-preserving, and computation-efficient alternative to traditional sensor-based models. We design a scalable pre-training framework that captures temporal dynamics and behavioral patterns from large-scale step sequences, enabling transfer across more than 20 health risk prediction tasks spanning diverse devices, new regions, and novel disease types. Extensive experiments demonstrate that StepFM achieves strong performance compared to existing methods while maintaining robustness across heterogeneous settings. Furthermore, our analysis reveals interpretable and generalizable relationships between physical activity patterns and various health risks, offering new insights into activity-based health modeling. Our work establishes step-based sensing as a viable foundation for scalable and real-world health monitoring.

Figures

Figures reproduced from arXiv: 2607.06954 by Songlin Xu, Yuyao Zhu, Zhenghuang Wu.

Figure 1
Figure 1. Figure 1: Performance trends across 21 diseases. (a) AUROC scores and (b) F1 scores for different models. The parallel oscillation across models indicates that disease predictability is primarily driven by the inherent correlation between the specific disease and physical activity patterns. A salient observation from these visualizations is the remarkable morphological consistency of the performance curves across al… view at source ↗
Figure 2
Figure 2. Figure 2: Scaling behaviors of StepFM. (a) Data Efficiency: performance under varying proportions of downstream labeled training data. (b) Temporal Receptive Field: performance under varying lengths of the input observation window. Why Does StepFM Work? The traditional machine learning baseline is built on 99 hand-crafted step-count features, including circadian rest-activity rhythm descriptors33, daily and weekly s… view at source ↗
Figure 3
Figure 3. Figure 3: Linear probing performance using representations extracted from different Mamba layers. The inverted-U trajectory indicates that intermediate layers retain optimal, generalizable physiological representations. inherently depends on the duration of the observation window. Short-term activity snapshots are easily distorted by transient behavioral anomalies, such as temporary illnesses, vacations, or atypical… view at source ↗
Figure 4
Figure 4. Figure 4: Self-supervised pre-training framework of StepFM. Raw minute-level step counts are first encoded into hourly macro tokens using log-scaled tokenization and temporal rhythm injection, while preserving local intra-hour dynamics via a micro CNN. These dual-stream representations are fed into a Mamba backbone, dynamically modulated by FiLM of micro stream, and jointly optimized through autoregressive next-toke… view at source ↗
Figure 5
Figure 5. Figure 5: Disease prevalence distribution across the NHANES 2011–2014 dataset. Each row corresponds to a health outcome, where the three color blocks denote the proportions of positive cases, negative cases, and missing, respectively. thyroid problem, kidney weak/failing); Respiratory & Systemic Conditions (emphysema, chronic bronchitis, arthritis, cancer, anemia treatment); and Neurological, Mental & Sensory (memor… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

43 extracted references · 43 canonical work pages · 6 internal anchors

  1. [1]

    ACM on Interactive, Mobile, Wearable Ubiquitous Technol.9, 1–26 (2025)

    Zhu, G.et al.Master: A multi-modal foundation model for human activity recognition.Proc. ACM on Interactive, Mobile, Wearable Ubiquitous Technol.9, 1–26 (2025)

    work page 2025

  2. [2]

    & Antonakopoulos, T

    Petropoulos, A., Sikeridis, D. & Antonakopoulos, T. Wearable smart health advisors: An imu-enabled posture monitor. IEEE Consumer Electron. Mag.9, 20–27 (2020)

    work page 2020

  3. [3]

    InInternational Conference on Wireless Mobile Communication and Healthcare, 67–74 (Springer, 2017)

    Kellner, K.-H.et al.Online monitoring of posture for preventive medicine using low-cost inertial sensors. InInternational Conference on Wireless Mobile Communication and Healthcare, 67–74 (Springer, 2017)

    work page 2017

  4. [4]

    In2024 9th International Conference on Smart and Sustainable Technologies (SpliTech), 1–6 (IEEE, 2024)

    Bisio, I.et al.Towards sensorized glasses: A smart wearable system for head movement monitoring. In2024 9th International Conference on Smart and Sustainable Technologies (SpliTech), 1–6 (IEEE, 2024)

    work page 2024

  5. [5]

    & Salim, F

    Li, Z., Deldari, S., Chen, L., Xue, H. & Salim, F. D. Sensorllm: Aligning large language models with motion sensors for human activity recognition. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 354–379 (2025)

    work page 2025

  6. [6]

    & Rahman, T

    Luo, Y ., Chen, Y ., Salekin, A. & Rahman, T. Toward foundation model for multivariate wearable sensing of physiological signals.ACM Transactions on Comput. for Healthc.(2025)

    work page 2025

  7. [7]

    & Chen, L

    Miao, S. & Chen, L. Wonderwall: A virtual-to-real foundation model for imu-based har.Proc. ACM on Interactive, Mobile, Wearable Ubiquitous Technol.10, 1–31 (2026)

    work page 2026

  8. [8]

    & Sandström, T

    Farooqi, N., Slinde, F., Carlsson, M., Håglin, L. & Sandström, T. Predicting energy requirement with pedometer-determined physical-activity level in women with chronic obstructive pulmonary disease.Int. J. Chronic Obstr. Pulm. Dis.1129–1137 (2015)

    work page 2015

  9. [9]

    Y ., Zhang, A., Oh, J

    Ruan, F. Y ., Zhang, A., Oh, J. Y ., Jin, S. & Jacobson, N. C. Foundation models for wearable movement data in mental health research.ArXivarXiv–2411 (2025)

    work page 2025

  10. [10]

    Buckinx, F.et al.Added value of a triaxial accelerometer assessing gait parameters to predict falls and mortality among nursing home residents: A two-year prospective study.Technol. Heal. Care23, 195–203 (2015). 11/18

    work page 2015

  11. [11]

    R., Daneshgar, S., Aragon, K

    Embry, K. R., Daneshgar, S., Aragon, K. & Jayaraman, A. Strategic user identification increases the impact of wearable airbags in high-fall risk populations with neurological disease. In2025 International Conference On Rehabilitation Robotics (ICORR), 121–127 (IEEE, 2025)

    work page 2025

  12. [12]

    & Naeem, M

    Nguyen, H., Zhou, F., Mirza, F. & Naeem, M. A. Fall detection using smartphones to enhance safety and security of older adults at home. In2018 Eleventh International Conference on Mobile Computing and Ubiquitous Network (ICMU), 1–2 (IEEE, 2018)

    work page 2018

  13. [13]

    An, S.et al.Mgait: Model-based gait analysis using wearable bend and inertial sensors.ACM Transactions on Internet Things3, 1–24 (2021)

    work page 2021

  14. [14]

    ACM on interactive, mobile, wearable ubiquitous technologies3, 1–23 (2019)

    Zhang, H.et al.Pdmove: Towards passive medication adherence monitoring of parkinson’s disease using smartphone-based gait assessment.Proc. ACM on interactive, mobile, wearable ubiquitous technologies3, 1–23 (2019)

    work page 2019

  15. [15]

    Large-scale Training of Foundation Models for Wearable Biosignals

    Abbaspourazad, S.et al.Large-scale training of foundation models for wearable biosignals.arXiv preprint arXiv:2312.05409(2023). 16.Narayanswamy, G.et al.Scaling wearable foundation models.arXiv preprint arXiv:2410.13638(2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  16. [16]

    Gu, X.et al.Sensing cardiac health across scenarios and devices: A multi-modal foundation model pretrained on heterogeneous data from 1.7 million individuals.arXiv preprint arXiv:2507.01045(2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  17. [17]

    Beyond Sensor Data: Foundation Models of Behavioral Data from Wearables Improve Health Predictions

    Erturk, E.et al.Beyond sensor data: Foundation models of behavioral data from wearables improve health predictions. arXiv preprint arXiv:2507.00191(2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  18. [18]

    Yao, Q.et al.Watch the rhythm: Breaking privacy with accelerometer at the extremely-low sampling rate of 5hz. In Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security, CCS ’24, 1776–1790, DOI: 10.1145/3658644.3690370 (Association for Computing Machinery, New York, NY , USA, 2024)

    work page doi:10.1145/3658644.3690370 2024

  19. [19]

    Qiu, M., Weng, C., Fan, M. & Wu, K. Towards customizable foundation models for human activity recognition with wearable devices.Proc. ACM on Interactive, Mobile, Wearable Ubiquitous Technol.9, 1–29 (2025)

    work page 2025

  20. [20]

    & Fujimoto, S

    Fujino, Y ., Tokuda, F. & Fujimoto, S. Decreased step count prior to the first visit for mdd treatment: a retrospective, observational, longitudinal cohort study of continuously measured walking activity obtained from smartphones.Front. public health11, 1190464 (2023)

    work page 2023

  21. [21]

    W., Maxhuni, A., Kessing, L

    Hafiz, P., Miskowiak, K. W., Maxhuni, A., Kessing, L. V . & Bardram, J. E. Wearable computing technology for assessment of cognitive functioning of bipolar patients and healthy controls.Proc. ACM on Interactive, Mobile, Wearable Ubiquitous Technol.4, 1–22 (2020)

    work page 2020

  22. [22]

    R., Wyatt, H

    Bassett, D. R., Wyatt, H. R., Thompson, H., Peters, J. C. & Hill, J. O. Pedometer-measured physical activity and health behaviors in us adults: 2380board# 259 june 3 2: 00 pm-3: 30 pm.Medicine & Sci. Sports & Exerc.42, 606 (2010)

    work page 2010

  23. [23]

    Sports & Exerc.36, 1567–1573 (2004)

    Tudor-Locke, C.et al.Descriptive epidemiology of pedometer-determined physical activity.Medicine & Sci. Sports & Exerc.36, 1567–1573 (2004)

    work page 2004

  24. [24]

    & Walsh, D

    Eisenmann, J., Laurson, K., Wickel, E., Gentile, D. & Walsh, D. Utility of pedometer step recommendations for predicting overweight in children.Int. J. Obes.31, 1179–1182 (2007)

    work page 2007

  25. [25]

    Z.et al.Predicting depression treatment outcome using daily step count sensory data.ACM Transactions on Comput

    Hossain, M. Z.et al.Predicting depression treatment outcome using daily step count sensory data.ACM Transactions on Comput. for Healthc.(2025)

    work page 2025

  26. [26]

    ACM on Interactive, Mobile, Wearable Ubiquitous Technol.9, 1–35 (2025)

    Saha, M.et al.Pulse-ppg: An open-source field-trained ppg foundation model for wearable applications across lab and field settings.Proc. ACM on Interactive, Mobile, Wearable Ubiquitous Technol.9, 1–35 (2025)

    work page 2025

  27. [27]

    TimeSiam: A Pre-Training Framework for Siamese Time-Series Modeling

    Dong, J.et al.Timesiam: A pre-training framework for siamese time-series modeling.arXiv preprint arXiv:2402.02475 (2024). 29.LaValley, M. P. Logistic regression.Circulation117, 2395–2399 (2008)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  28. [28]

    S.et al.Systematic review of the prospective association of daily step counts with risk of mortality, cardiovascular disease, and dysglycemia.Int

    Hall, K. S.et al.Systematic review of the prospective association of daily step counts with risk of mortality, cardiovascular disease, and dysglycemia.Int. J. Behav. Nutr. Phys. Activity17, 78 (2020)

    work page 2020

  29. [29]

    & Chen, Z

    Yuan, L., Tian, Z., Jia, X. & Chen, Z. The association between daily step count, step frequency and the risk of chronic obstructive pulmonary disease: a cross-sectional study using nhanes data.Int. J. Chronic Obstr. Pulm. Dis.2325–2335 (2025)

    work page 2025

  30. [30]

    Care65, 455–463 (2020)

    Rice, H.et al.Reduced step count and clinical frailty in hospitalized adults with community-acquired pneumonia.Respir. Care65, 455–463 (2020). 12/18

    work page 2020

  31. [31]

    & Swaab, D

    Witting, W., Kwa, I., Eikelenboom, P., Mirmiran, M. & Swaab, D. F. Alterations in the circadian rest-activity rhythm in aging and alzheimer’s disease.Biol. psychiatry27, 563–572 (1990)

    work page 1990

  32. [32]

    & Bassett Jr, D

    Tudor-Locke, C. & Bassett Jr, D. R. How many steps/day are enough? preliminary pedometer indices for public health. Sports medicine34, 1–8 (2004)

    work page 2004

  33. [33]

    Doherty, A.et al.Large scale population assessment of physical activity using wrist worn accelerometers: the uk biobank study.PloS one12, e0169649 (2017)

    work page 2017

  34. [34]

    Open7, e2451208 (2024)

    Bizzozero-Peroni, B.et al.Daily step count and depression in adults: a systematic review and meta-analysis.JAMA Netw. Open7, e2451208 (2024)

    work page 2024

  35. [35]

    JAMA psychiatry79, 550–559 (2022)

    Pearce, M.et al.Association between physical activity and risk of depression: a systematic review and meta-analysis. JAMA psychiatry79, 550–559 (2022)

    work page 2022

  36. [36]

    P.et al.Physical activity in the united states measured by accelerometer.Medicine & Sci

    Troiano, R. P.et al.Physical activity in the united states measured by accelerometer.Medicine & Sci. Sports & Exerc.40, 181–188 (2008)

    work page 2008

  37. [37]

    & Sieber, C

    Baumer, A., von Wyl, V ., Gonzenbach, R., Haag, C. & Sieber, C. Barka-ms: Fitbit and fatigue data of people with multiple sclerosis during and after rehabilitation. Zenodo, DOI: 10.5281/zenodo.17485784 (2025)

    work page doi:10.5281/zenodo.17485784 2025

  38. [38]

    & Barnaghi, P

    Nilforooshan, R. & Barnaghi, P. The resilient dataset: Multimodal monitoring of ageing-related comorbidities and cognitive decline. Zenodo, DOI: 10.5281/zenodo.16755408 (2025)

    work page doi:10.5281/zenodo.16755408 2025

  39. [39]

    Bae, S., Dey, A. K. & Low, C. A. Using passively collected sedentary behavior to predict hospital readmission. In Proceedings of the 2016 ACM International Joint Conference on Pervasive and Ubiquitous Computing, 616–621 (2016)

    work page 2016

  40. [40]

    Time2Vec: Learning a Vector Representation of Time

    Low, C. A.et al.Fitbit step counts during inpatient recovery from cancer surgery as a predictor of readmission.Annals Behav. Medicine52, 88–92 (2018). 43.NHANES. Nhanes dataset (2026). 44.Vaswani, A.et al.Attention is all you need.Adv. neural information processing systems30(2017). 45.Kazemi, S. M.et al.Time2vec: Learning a vector representation of time.a...

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  41. [41]

    Mamba: Linear-Time Sequence Modeling with Selective State Spaces

    Gu, A. & Dao, T. Mamba: Linear-time sequence modeling with selective state spaces.arXiv preprint arXiv:2312.00752 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  42. [42]

    & Muschelli, J

    Koffman, L. & Muschelli, J. Minute level step counts and physical activity data from the National Health and Nutrition Examination Survey (NHANES) 2011-2014.PhysioNetDOI: 10.13026/9n0r-tv02 (2025). Version 1.0.1

    work page doi:10.13026/9n0r-tv02 2011

  43. [43]

    L., Kroenke, K., Williams, J

    Spitzer, R. L., Kroenke, K., Williams, J. B. & Löwe, B. A brief measure for assessing generalized anxiety disorder: the gad-7.Arch. internal medicine166, 1092–1097 (2006). Methods Framework Pipeline Step-count data contain health-relevant patterns at multiple temporal scales, from within-hour activity bursts to circadian regularity and cross-day variabili...

    work page 2006