REVIEW 3 major objections 5 minor 1 cited by
Beyond Sensor Data: Foundation Models of Behavioral Data from Wearables Improve Health Predictions
T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read A model trained purely on behavioral wearable data—activity, mobility, gait—matches or beats raw-sensor models on most of 57 health tasks; combining both embeddings is best almost everywhere.
desk verdict A strong, large-scale behavioral wearable foundation model; the static-task results are convincing, but the time-varying claims need within-person evaluation. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is WBM, a contrastively pre-trained encoder whose input pipeline turns a participant's week into a dense $168 \times 54$ matrix—168 hourly patches for 27 variables plus 27 missingness indicators—using Time Series Transformer (TST) tokenization with global-mean imputation, and whose backbone is a bi-directional Mamba-2 selective state-space model. The pre-training objective is a regularized InfoNCE contrastive loss with KoLeo spread regularization, with positive pairs formed by dropping a random 23.3 percent of tokens from the same participant's week. The machinery works by forcing the embedding to summarize the whole week's behavior in a fixed vector, so that downstream tasks reduce to linear probes on that vector; the authors selected this configuration by grid-searching three tokenizers and three backbones on the age-prediction task and confirming the winner on 47 survey outcomes.
What would settle it
Restrict the pregnancy evaluation to a narrow age band, for example only participants aged 25 to 34, or add age and sex as explicit regressors in the linear probe, and check whether WBM's 0.864 and the combined model's 0.921 AUROC survive; if the scores collapse toward baseline, the pregnancy result is demographic discrimination rather than detection of pregnancy-related change. A complementary check applies the same age-stratified probe to the infection and injury tasks, whose labels are not age-linked and should remain stable.
Extended reading notes
Core claim
The central discovery is that a foundation model pre-trained on irregularly sampled behavioral wearable data—weekly windows of hourly aggregates across 27 expert-derived health quantities—yields competitive or superior health detection compared with modeling low-level sensor data. On the authors' evaluation, WBM predicts age with MAE 3.67 and biological sex with AUROC 0.999, outperforms the simple baseline in 39 of 47 baseline disease and medication outcomes, and outperforms the PPG model on sleep tasks ($R^2$ 0.590 vs 0.110 for sleep duration) and other behavior-driven states such as infection and injury. Conversely, PPG remains superior where physiology alone suffices, most notably for diabetes classification. Concatenating WBM and PPG embeddings gives the best model in 42 of 47 baseline outcomes and in every time-varying task except diabetes, with pregnancy detection reaching AUROC 0.921. The paper reads this as evidence that behavioral and sensor modalities are complementary, with behavioral data contributing signal that raw sensors capture only sporadically or not at all.
Load-bearing premise
The results stand on the assumption that the downstream tasks measure what they claim to measure, rather than the demographic attributes WBM encodes with near-perfect accuracy; in particular, the pregnancy task labels weeks from 430 pregnancies as positive against non-pregnant women under 50 as negative controls, so its AUROC could reflect age and sex cues instead of behavioral and physiological changes from pregnancy.
Editorial extensions
If this is right
- Sleep quality can be estimated from behavioral data alone at useful accuracy ($R^2$ 0.590 for duration and 0.424 for efficiency), suggesting overnight sensor streams are not required for population-level sleep monitoring.
- Behavioral embeddings add signal to a dedicated PPG foundation model on most tasks—42 of 47 baseline outcomes and every time-varying task except diabetes—so fused wearable models are the natural deployment target.
- Behavior-dominated and physiology-dominated tasks separate cleanly: WBM wins on sleep, infection, and injury; PPG wins on diabetes and most baseline conditions; task designers can predict which modality will carry a given outcome.
- The winning configuration (TST tokenization with a Mamba-2 backbone) generalizes beyond age prediction, being best in 25 of 47 survey outcomes in the ablation, which argues against assuming transformer-based designs transfer to new wearable data types.
Reading between the lines
- Because WBM encodes age (MAE 3.67) and sex (AUROC 0.999) almost perfectly, every downstream result on a demographically skewed label should be re-read with confound control before being taken at face value; the pregnancy task is the clearest case, but the same logic applies to other age- and sex-correlated outcomes.
- Some of WBM's advantage over raw-sensor models is inherited from expert knowledge baked into its inputs: quantities like VO2max and gait steadiness are outputs of validated clinical algorithms, so a fair division of credit would compare WBM against a model trained on the raw sensor streams that produce these quantities.
- The winning design—simple dense tokenization beating sophisticated irregular-time mechanisms—suggests that for heavily aggregated human-scale data, the representation objective and imputation scheme matter more than the model's handling of irregularity; a testable extension would ablate the token-drop augmentation against model capacity.
- If the complementarity holds, health applications could route by task type: physiology-heavy screens from PPG, behavior-heavy monitoring (sleep, recovery, injury) from behavioral models, and combined models where both matter, which would inform on-device compute budgets.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper introduces WBM, a foundation model pretrained with a regularized contrastive objective on 2.5B hours of hourly-aggregated behavioral wearable data from 161,855 participants of the Apple Heart and Movement Study. Weekly windows are tokenized into dense 168 x 54 matrices (TST tokenization), encoded by a bidirectional Mamba-2 backbone, and pooled to a single embedding per week. The model is evaluated on 57 health-related tasks: age/sex prediction, 47 baseline disease/medication survey outcomes, and time-varying tasks covering sleep regression, pregnancy, respiratory infection, injury, and HbA1c-based diabetes status. The main claims are that behavioral data alone yields strong health predictions, that WBM outperforms a simple statistical baseline and a raw PPG foundation model on behavior-driven tasks such as sleep and injury, and that concatenating WBM and PPG embeddings gives the best performance on most tasks. The paper also claims that WBM captures meaningful within-person variation over time, supported by the intra-subject task results.
Significance. If the claims hold, this is a substantial contribution: it is the largest-scale foundation model for behavioral wearable data that I am aware of, with a systematic and transparent comparison of tokenizers and backbones, participant-level data splits, bootstrap confidence intervals, and reconstruction sanity checks. The finding that WBM complements a PPG foundation model across many tasks is practically important and is supported by a large battery of evaluations. The authors are also careful to document the cohort's demographic limitations and to report subgroup results. However, the paper's central novelty for health monitoring is the claimed ability to track within-person, time-varying health states, and that claim rests on task constructions that currently permit a between-person shortcut. This undermines a headline contribution and requires a reanalysis before the results can be accepted as stated.
major comments (3)
- [Section 5.1 / Appendix A.6.1 / Table 2] The pregnancy, infection, and injury tasks are framed as intra-subject (time-varying) detection, but their positive/negative construction allows a linear probe to achieve high AUROC without any temporal discrimination. For pregnancy, positive weeks come from 385 ever-pregnant participants, while negative weeks are dominated by 24,225 never-pregnant female controls; for infection and injury, the negative pools similarly include 83K and 75.7K never-affected controls. A constant per-person score that separates ever-affected from never-affected participants would therefore yield high AUROC, so the reported values (pregnancy AUROC 0.864, infection 0.749, injury 0.680) do not demonstrate sensitivity to the transient health state. The abstract and Section 5.1 explicitly claim 'track[ing] meaningful within-person variation over time', so this confound is load-bearing. The fix is to evaluate within affected participants only, classifying their positive weeks versus their own negative weeks with participant-level clustering, or to use matched/paired analyses; bootstrap intervals over weeks should also account for this clustering.
- [Appendix A.6.2 / Table 25] The age-stratified pregnancy results do not resolve the participant-identity confound. Stratifying by age band still compares ever-pregnant women against never-pregnant women, who can differ on stable attributes such as parity, fertility history, lifestyle, or socioeconomic factors; the reported AUROC within each band can therefore still be driven by between-person discrimination rather than gestational state. A within-person test, restricted to the 385 ever-pregnant participants and comparing their pregnant versus non-pregnant weeks, is necessary to support the claim that WBM detects pregnancy-related changes.
- [Section 5.2 / Table 2] The complementarity claim for WBM + PPG is partly justified by the pregnancy task, which the authors describe as a 'clear example of the complementary nature' of the two modalities. Because the pregnancy evaluation currently confounds participant identity with gestational state, the reported WBM + PPG AUROC of 0.921 does not provide valid evidence for complementarity in detecting pregnancy. The complementarity claim should be re-established on a within-person pregnancy evaluation, or the pregnancy example should be removed from the highlighted results until such an analysis is provided.
minor comments (5)
- [Appendix A.6.1] The sentence 'Since substantial prior work has shown that wearables can detect influenza-like illness and COVID-19 (Nestor et al., 2023), we we also use this as a downstream task' contains a duplicated 'we'.
- [Appendix A.1.2] The Active Smoker sample size is reported as 'N = 23,8345', which appears to be a typo; it should likely be 23,845 or 23,834.
- [Section 4.2] The phrase 'performative foundation models from wearable signals' should probably be 'performant foundation models' or 'performance-oriented foundation models'; 'performative' has a different meaning in machine learning.
- [Table 5 caption] The caption contains 'the tudy' which should be 'the study'.
- [Section 5.1 / Appendix A.6.1] The sleep regression tasks are genuinely within-person and are a strong piece of evidence for the time-varying claim; I would encourage the authors to emphasize these results and to consider whether the discrete time-varying tasks can be reformulated in a fully within-person design.
Circularity Check
No significant circularity: the WBM training and evaluation pipeline is self-contained, with self-citations only to non-load-bearing prior baselines.
full rationale
The paper's derivation chain has no equation-level reduction of a predicted quantity to its inputs. WBM is pretrained with a self-supervised InfoNCE plus KoLeo objective on unlabeled weekly behavioral data, so downstream health labels enter only at the linear-probing stage; no downstream performance is used to define the pretraining loss, tokenizer, or backbone. Architecture selection uses validation age-prediction MAE, which is a standard model-selection criterion rather than a fitted parameter renamed as a prediction. The PPG baseline is cited from prior ICLR 2024 work by overlapping authors, but it is used as an external input method and not as the justification for WBM's own results, so the reported WBM+PPG gains are empirical and falsifiable rather than derived from the citation. The paper's explicit caveat about age/sex label leakage in demographic tasks and the potential confound in the pregnancy evaluation (never-pregnant controls as negatives) are validity limitations, not circular reductions: the reported AUROCs are not equal by construction to any input feature or fitted parameter. Overall, the derivation is self-contained against external benchmarks, and the mild self-references do not carry the argument.
Assumptions & free parameters
free parameters (1)
- Model hyperparameters (KoLeo weight 0.21, dropout 2.7%, token drop 23.3%, hidden dim 256, 24 layers) =
0.21, 2.7%, 23.3%, 256, 24
assumptions (4)
- domain assumption AHMS cohort is sufficiently representative of the target population for the claimed health predictions.
- domain assumption The 27 HealthKit quantities are accurate, validated measures of behavior and physiology.
- domain assumption Self-reported survey and medication labels are accurate enough to serve as ground truth.
- domain assumption Contrastive pretraining with same-subject positive pairs transfers to downstream health detection tasks via linear probing.
Cite this review
Pith. "Pith review of Beyond Sensor Data: Foundation Models of Behavioral Data from Wearables Improve Health Predictions." pith.science (2026). https://pith.science/paper/MXDVOTWU
@misc{pith2026250700191,
author = {Pith},
title = {Pith review of: Beyond Sensor Data: Foundation Models of Behavioral Data from Wearables Improve Health Predictions},
year = {2026},
howpublished = {\url{https://pith.science/paper/MXDVOTWU}},
note = {Machine review of arXiv:2507.00191}
}
read the original abstract
Wearable devices record physiological and behavioral signals that can improve health predictions. While foundation models are increasingly used for such predictions, they have been primarily applied to low-level sensor data, despite behavioral data often being more informative due to their alignment with physiologically relevant timescales and quantities. We develop foundation models of such behavioral signals using over 2.5B hours of wearable data from 162K individuals, systematically optimizing architectures and tokenization strategies for this unique dataset. Evaluated on 57 health-related tasks, our model shows strong performance across diverse real-world applications including individual-level classification and time-varying health state prediction. The model excels in behavior-driven tasks like sleep prediction, and improves further when combined with representations of raw sensor data. These results underscore the importance of tailoring foundation model design to wearables and demonstrate the potential to enable new health applications.
Figures
Figures from the paper (1 more)
Forward citations
Cited by 1 Pith paper
-
Physical activities enable scalable foundation modelling for broad-spectrum health prediction
A 3.4M-parameter foundation model pre-trained on step-count data alone achieves best AUROC on 20 of 21 health risk prediction tasks across multiple devices, regions, and diseases.
Reference graph
Works this paper leans on
-
[1]
Large-scale training of foundation models for wearable biosignals
Abbaspourazad, S., Elachqar, O., Miller, A., Emrani, S., Nallasamy, U., and Shapiro, I. Large-scale training of foundation models for wearable biosignals. In Interna- tional Conference on Learning Representations, 2024a. Abbaspourazad, S., Mishra, A., Futoma, J., Miller, A. C., and Shapiro, I. Wearable accelerometer foundation mod- els for health via know...
-
[3]
List of all derived health and behavioral quantity variables used in modeling, along with the method of aggregation used to create hourly-level aggregates, a brief description, the native sampling rate, and summary statistics (what percentage of individual weeks and subjects have at least one measurement). Tuple. First, we consider mapping each tuple (ts,...
work page 2022
-
[4]
Dense. A potential limitation of tuple-based approaches is their poor scalablity and potential down-weighting of sparse but informative variables. Hence, we also consider the more traditional dense approach, where we convert data from tuples to a dense array of values. We create a matrix X s T ∈ R|T |×|V |, where |V | is the number of quantities or variab...
work page 2021
-
[6]
Ehrmamba: Towards generalizable and scalable foundation models for electronic health records
Fallahpour, A., Alinoori, M., Afkanpour, A., and Krish- nan, A. Ehrmamba: Towards generalizable and scalable foundation models for electronic health records. arXiv preprint arXiv:2405.14567,
-
[8]
Gu, A. and Dao, T. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752,
-
[9]
Bi-mamba4ts: Bidirectional mamba for time series forecasting
Liang, A., Jiang, X., Sun, Y ., and Lu, C. Bi-mamba4ts: Bidirectional mamba for time series forecasting. arXiv preprint arXiv:2404.15772,
-
[10]
Scaling wearable foundation models
Narayanswamy, G., Liu, X., Ayush, K., Yang, Y ., Xu, X., Liao, S., Garrison, J., Tailor, S., Sunshine, J., Liu, Y ., et al. Scaling wearable foundation models. arXiv preprint arXiv:2410.13638,
-
[11]
Ramachandran, P., Zoph, B., and Le, Q. V . Swish: a self-gated activation function. arXiv preprint arXiv:1710.05941, 7(1):5,
Show all 24 references
-
[14]
doi: 10.1038/ s41746-023-00851-6
ISSN 2398-6352. doi: 10.1038/ s41746-023-00851-6. URL https://www.nature. com/articles/s41746-023-00851-6 . Num- ber: 1 Publisher: Nature Publishing Group. Shaw, P., Uszkoreit, J., and Vaswani, A. Self-attention with relative position representations. arXiv preprint arXiv:1803.02155,
-
[16]
Unified training of universal time series fore- casting transformers
Woo, G., Liu, C., Kumar, A., Xiong, C., Savarese, S., and Sahoo, D. Unified training of universal time series fore- casting transformers. arXiv preprint arXiv:2402.02592,
-
[17]
Wornow, M., Bedi, S., Hernandez, M. A. F., Steinberg, E., Fries, J. A., R´e, C., Koyejo, S., and Shah, N. H. Context clues: Evaluating long context models for clinical predic- tion tasks on EHRs. arXiv preprint arXiv:2412.16178,
-
[18]
A., Narain, J., Darnell, G., Hallgrimsson, H., Jeong, H., Forde, D., Fineman, R., Raghuram, K
Xu, M. A., Narain, J., Darnell, G., Hallgrimsson, H., Jeong, H., Forde, D., Fineman, R., Raghuram, K. J., Rehg, J. M., and Ren, S. Relcon: Relative contrastive learning for a motion foundation model for wearable data. arXiv preprint arXiv:2411.18822,
-
[19]
Simts: rethinking contrastive representation learning for time series forecasting
Zheng, X., Chen, X., Sch ¨urch, M., Mollaysa, A., Allam, A., and Krauthammer, M. Simts: rethinking contrastive representation learning for time series forecasting. arXiv preprint arXiv:2303.18205,
-
[22]
Different colors refer to different variable types
Tuple tokenization framework for irregularly sampled wearable activity data. Different colors refer to different variable types. Each variable is represented by a separate token unlike traditional multivariate modeling approaches. The token representation for each variable is ...
2017
-
[23]
The relative performance of models is consistent with the age prediction performance, with TST with Mamba-2 performing best
The number of survey outcomes for which each embedding was the best performing model. The relative performance of models is consistent with the age prediction performance, with TST with Mamba-2 performing best. A.5.3. M ASKED AUTOENCODER RESULTS Due to the popularity of masked...
2024
-
[24]
positive
could be an interesting choice to overcome this problem; we leave this exploration for future work. A.6. Additional Downstream Task Details A.6.1. T ASK PREPROCESSING DETAILS Demographics: The first downstream task is to predict participant age and biological sex (male or fema...
2024
-
[2017]
R., Khorasani, A., Adamopoulos, G., Bhagwatkar, R., Biloˇs, M., Ghonia, H., Hassen, N
Rasul, K., Ashok, A., Williams, A. R., Khorasani, A., Adamopoulos, G., Bhagwatkar, R., Biloˇs, M., Ghonia, H., Hassen, N. V ., Schneider, A., et al. Lag-llama: Towards foundation models for time series forecasting. arXiv preprint arXiv:2310.08278,
-
[2018]
A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M
Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M. S., Bohg, J., Bosse- lut, A., Brunskill, E., et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258,
-
[2019]
doi: 10.1056/NEJMsr1809937
ISSN 0028-4793. doi: 10.1056/NEJMsr1809937. URL https://www.nejm. org/doi/10.1056/NEJMsr1809937. Publisher: Massachusetts Medical Society. Dufter, P., Schmitt, M., and Sch ¨utze, H. Position infor- mation in transformers: An overview. Computational Linguistics, 48(3):733–763,
-
[2020]
Is mamba effective for time series forecasting? arXiv preprint arXiv:2403.11144,
Wang, Z., Kong, F., Feng, S., Wang, M., Zhao, H., Wang, D., and Zhang, Y . Is mamba effective for time series forecasting? arXiv preprint arXiv:2403.11144,
-
[2021]
Spreading vectors for similarity search
Sablayrolles, A., Douze, M., Schmid, C., and J ´egou, H. Spreading vectors for similarity search. arXiv preprint arXiv:1806.03198,
-
[2022]
F., Stella, L., Turkmen, C., Zhang, X., Mercado, P., Shen, H., Shchur, O., Rangapuram, S
Ansari, A. F., Stella, L., Turkmen, C., Zhang, X., Mercado, P., Shen, H., Shchur, O., Rangapuram, S. S., Arango, S. P., Kapoor, S., et al. Chronos: Learning the language of time series. arXiv preprint arXiv:2403.07815,
-
[2023]
and Gu, A
Dao, T. and Gu, A. Transformers are SSMs: Generalized models and efficient algorithms through structured state space duality. arXiv preprint arXiv:2405.21060,
-
[2024]
and Mergenthaler-Canseco, M
Garza, A. and Mergenthaler-Canseco, M. Timegpt-1. arXiv preprint arXiv:2310.03589,
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.