Pith. sign in

REVIEW 3 major objections 7 minor 2 cited by

A single model pretrained on a trillion minutes of wearable data predicts heart, metabolic, sleep, and mental-health outcomes across 35 tasks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 13:21 UTC pith:VAOIPHZR

load-bearing objection The largest wearable foundation model to date, with unusually thorough evaluation, but the central generalization claim needs a de-duplication check between pretraining and downstream cohorts before it can be taken as established. the 3 major comments →

arxiv 2605.22759 v3 pith:VAOIPHZR submitted 2026-05-21 cs.AI

Towards a General Intelligence and Interface for Wearable Health Data

Girish Narayanswamy , Maxwell A. Xu , A. Ali Heydari , Samy Abdel-Ghaffar , Marius Guerard , Kara Vaillancourt , Zhihan Zhang , Jake Garrison
show 32 more authors
Levi Albuquerque Dimitris Spathis Hong Yu Hamid Palangi Xuhai "Orson" Xu David G.T. Barrett Joseph Breda Jed McGiffin Yubin Kim Yuwei Zhang Naghmeh Rezaei Samuel Solomon Karan Ahuja Tim Althoff Jake Sunshine Ming-Zher Poh Benjamin Yetton Ari Winbush Nicholas B. Allen James M. Rehg Isaac Galatzer-Levy Yun Liu John Hernandez Anupam Pathak Conor Heneghan Yuzhe Yang Ahmed A. Metwally Pushmeet Kohli Mark Malhotra Shwetak Patel Xin Liu Daniel McDuff
This is my paper
classification cs.AI
keywords foundation modelwearable sensorsself-supervised learninghealth predictionscaling lawsmasked autoencoderLLM agentsdigital health
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.

The paper sets out to show that one sensor foundation model, trained without labels on a massive corpus of wearable data, can serve as a general-purpose representation of sensed human physiology. It introduces SensorFM, pretrained on over one trillion minutes of data from five million people and five sensor modalities, and evaluates it on 35 clinical and behavioral prediction tasks. The paper reports that SensorFM beats a supervised engineered-feature baseline on 34 of 35 tasks, that joint scaling of data and model capacity yields near-linear downstream gains, and that its generative capabilities preserve daily health metrics even when an hour of data is missing. It also reports that feeding SensorFM predictions to a personal health agent produces clinician ratings statistically indistinguishable from those given ground-truth labels. If these claims hold, one adaptable model could replace many task-specific wearable algorithms, easing label scarcity and enabling more grounded consumer health AI.

Core claim

SensorFM learns a generalizable representation of sensed human physiology through masked reconstruction pretraining on an unprecedented volume of unlabeled wearable data. Linear probes on the frozen embeddings outperform supervised models trained on engineered features on 34 of 35 discriminative tasks spanning cardiovascular, metabolic, sleep, mental health, lifestyle, and demographic outcomes. The paper further claims that scaling both model capacity and pretraining data volume by orders of magnitude produces near-linear improvements in downstream performance, and that the resulting generative modeling accurately infills missing sensor segments, keeping daily step, sleep, and exercise estim

What carries the argument

SensorFM is a masked-autoencoder-style Vision Transformer trained on 24-hour windows of 34 one-minute aggregate features derived from five sensor modalities: PPG, accelerometer, electrodermal activity, skin temperature, and altimeter. The central mechanism is adaptive/inherited masking (AIM): the model learns to reconstruct artificially masked patches while treating real-world missingness as an inherited mask, so it handles fragmented wearable data natively. After pretraining, the frozen encoder's embeddings are reduced with PCA and adapted via linear probes or via an LLM-driven 'classroom' of agents that writes and refines task-specific prediction heads. The scaling analysis varies model si

Load-bearing premise

The downstream evaluation people are assumed to be different from the people whose data was used for pretraining, but the paper provides no de-duplication evidence, so the reported transfer gains could partly reflect the model having seen those individuals before.

What would settle it

Train SensorFM on the current pretraining cohort, then evaluate only on downstream participants whose de-identified IDs can be provably excluded from that cohort; if the average ROC AUC and Pearson correlation advantages over the engineered-feature baseline shrink materially, the central transfer claim would be undermined.

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

If this is right

  • New health prediction tasks may require far fewer labeled examples, since SensorFM's few-shot performance improves with pretraining scale.
  • Missing wearable data, such as a one-hour device off-wrist period, can be imputed well enough that daily step, sleep-stage, and exercise estimates stay within about 1% of true values.
  • LLM-driven agentic search over prediction heads improves on simple linear probes on most tasks, and stronger language models yield better solutions, suggesting a scalable path for adapting embeddings.
  • A personal health agent given SensorFM predictions can produce responses that clinicians rate as comparable to responses given ground-truth measurements, supporting the use of model predictions as context in consumer health AI.
  • Because performance has not saturated, further increases in data volume and model capacity may continue to improve wearable health predictions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the downstream evaluation participants are not truly disjoint from the pretraining cohort, the reported transfer gains could partly reflect the model memorizing individuals rather than learning general physiology; a secure-hash de-duplication or temporal holdout would test this directly.
  • The finding that demographic features add less value at larger scales suggests the model implicitly encodes age, sex, and body composition; this raises questions about fairness, privacy, and what sensor data reveals beyond the targeted health outcome.
  • The generative infilling results point to a practical way to correct daily wearable summaries during device non-wear, but the demonstrated benefit is limited to a single contiguous one-hour gap and may not extend to longer or more complex missingness patterns.
  • The agentic classroom's success depends on the language model used and the scoring function; a testable extension is to apply the same search procedure to a held-out set of new tasks and measure whether validation-time gains persist out-of-fold.

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 / 7 minor

Summary. The paper proposes SensorFM, a masked-autoencoder foundation model pretrained on a large corpus of minute-level wearable sensor data from Fitbit/Pixel users. The authors report scaling laws for model and data size, evaluate frozen embeddings via linear probes and LLM-agent-searched heads on 35 health-related prediction tasks from three prospective studies (N=13,985), demonstrate generative imputation/forecasting and label-efficient adaptation, and integrate predictions into a clinician-evaluated Personal Health Agent. The central claim is that self-supervised pretraining at population scale yields a universal, transferable representation of sensed physiology.

Significance. If the claims hold, this is an important contribution to wearable health AI: it is the largest pretraining corpus reported in this domain, with a broad evaluation suite and a serious attempt at downstream clinical validation. The paper's strengths include detailed data curation and modeling methods, 5-fold person-independent evaluation with transformed confidence intervals, multiple model scales, and a blinded clinician study with explicit statistical comparisons. The agentic classroom and personal health agent evaluations are novel. However, the headline scale numbers are internally inconsistent, and the independence of downstream data from pretraining is not demonstrated; these issues need to be resolved before the generalization claims can be accepted.

major comments (3)
  1. [Abstract; M.2.1; §2.1] The pretraining corpus is described as 'over one trillion minutes' and 'over 2 billion hours' (M.2.1). 2 billion hours = 120 billion minutes, not 1 trillion; 1 trillion minutes ≈ 16.7 billion hours. With 5M participants and up to 10 weeks each, the maximum is ~350M person-days (8.4B hours, 504B minutes). M.3.2 later reports 175,062,146 day-long samples (~252B minutes). The figures are internally inconsistent by factors of 2–8. Because the novelty and scaling claims rest on corpus size, the authors must reconcile these numbers and specify whether 'minutes' count all timestamps, valid timestamps, or sensor-feature-minutes.
  2. [M.2.1; M.2.2; M.2.2.4] The downstream cohorts are described as 'unique' and 'independent' (Figure 1, M.2.2.4), but no de-duplication against the pretraining cohort is reported. The pretraining window (Sept 2024–Sept 2025) overlaps with the metabolic study (M.2.2.1, which links up to 3 months of prior data and runs up to 70 days) and the mental health study (M.2.2.3, four weeks). Since the encoder is pretrained on unlabeled streams, overlap could allow person-specific memorization and inflate the linear probe results, including the 34/35 win rate. Person-independent 5-fold CV on downstream labels does not address this. Provide an intersection analysis (e.g., hashed participant IDs) or explicitly exclude downstream participants from pretraining, and report results on the non-overlapping subset.
  3. [§2.6; M.6.2] The claim that SensorFM predictions are statistically indistinguishable from ground truth (p=0.396) rests on an unbalanced comparison. Condition B only provides ground-truth metabolic markers; mental health and sleep ground-truth targets are absent, while Condition A provides SensorFM predictions for those domains. The non-significance may reflect missing information in B rather than equivalent accuracy. Restrict the comparison to targets available in both conditions, or collect/ impute the missing ground truth for B.
minor comments (7)
  1. [Figure ED.7; ED.8] The axis labels contain rendering artifacts ('/uni00000040...' sequences). Replace with proper feature names.
  2. [M.3.3] Clarify the masking schedule: are '50% temporal block masking' and '50% modality block masking' selected with probability 0.5 each, and how does the 80% random patch masking combine with them?
  3. [Table ED.2] Pretraining does not collect non-binary gender ('n.c.'), while downstream includes 2.8% non-binary participants. Note this mismatch when comparing distributions.
  4. [Abstract; M.2.1] The text alternates between 'five million participants' and 5,020,000; use a consistent figure.
  5. [§2.5; Table ED.12] The classroom results are reported with F1 for classification, whereas Figure 4 uses ROC AUC. Although Table ED.12 reports F1 for both, the main text should acknowledge the metric change when comparing across panels.
  6. [§2.2] The counts '24 of 30' and '22 of 30' should be reconciled with the 35-task taxonomy; specify which tasks are excluded from the demographic-only comparison.
  7. [M.6.4] The clinician panel has only 31 profiles and 4 raters, and the ICC for Justifiability is -0.088. The aggregate ratings for that dimension should be interpreted with caution.

Circularity Check

0 steps flagged

No circular dependency found: downstream predictions are linear probes on frozen embeddings, and the scaling and agent results are empirical rather than definitional.

full rationale

The paper's derivation chain is: (i) unlabeled sensor streams are pretrained with a masked-reconstruction objective; (ii) scaling model capacity and data volume improves reconstruction and downstream linear-probe performance; (iii) frozen embeddings transfer to 35 health tasks; (iv) LLM-agent-searched heads and a Personal Health Agent improve end-use. None of these steps is defined in terms of its own output. Downstream labels (age, BMI, diagnoses, screeners) are never inputs to pretraining, and the linear probes are trained out-of-fold on frozen embeddings with person-independent folds (M.3.4), so the 34/35 win rate is an empirical transfer result rather than a fitted variable renamed as a prediction. The scaling relationship in Figure 1.e is a fitted description of observed pretraining/posttraining scores, not a constraint imposed by construction; all model/data combinations are reported in Table ED.5. Self-citations to AIM masking (Xu et al. 2025) and prior wearable scaling (Narayanswamy et al. 2025) establish the architecture and motivation, but the central generalization claim is evaluated against external prospective studies (M.2.2) and supervised engineered-feature baselines (M.3.6), so they are not load-bearing circularity. The only substantial validity concern is that downstream cohorts are asserted to be 'independent' of the 5M pretraining cohort without showing de-duplication (Figure 1 caption; M.2.2.4); if participants overlapped, memorization could inflate transfer numbers. That is a data-contamination/leakage risk, not a circular-definition reduction: no prediction equals its training input by construction. Accordingly, the correct circularity score is 0, with the independence caveat noted for correctness review.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entity, force, or conserved quantity. Its scientific contribution is empirical scaling, so the ledger lists modeling choices and domain assumptions on which the empirical claims rest, with cohort-independence being the least verified and most load-bearing.

free parameters (3)
  • PCA-50 embedding reduction for linear probes = 50 principal components
    All downstream linear-probe results use embeddings reduced to 50 PCs; the choice is justified by label sparsity but is a hand-selected hyperparameter that affects every reported ROC/r value.
  • Pretraining artificial masking schedule = 80% random / 50% temporal block / 50% modality block
    The mixture of applied masks is chosen by the authors and shapes the generative and transfer behavior of the model.
  • Classroom search metric weights = balanced_accuracy 1.0, F1 1.0, precision 0.2, recall 0.2, accuracy 0.01, exec_time 0.2
    The LLM agents hill-climb on a composite score with hand-specified weights; this defines what 'better solution' means in the agentic head-search results and is not derived from any external objective.
axioms (5)
  • domain assumption Masked autoencoder reconstruction on one-minute aggregates is a sufficient pretraining objective for downstream health classification and regression.
    The entire approach assumes that filling in masked sensor patches produces embeddings that transfer to health labels. No proof or theory is given; it is an empirical claim validated only on the specific tasks and datasets used.
  • domain assumption Person-level labels can be predicted from mean/std aggregates of per-day embeddings.
    Section M.3.4 aggregates all non-masked tokens per person by mean and standard deviation before PCA and linear probing. This discards potentially predictive temporal structure within or across days.
  • domain assumption Self-reported diagnoses, medication use, and screener thresholds are valid health outcome labels.
    Many of the 35 tasks use self-report or binary thresholds of continuous screeners (e.g., PHQ-8>=10). The paper acknowledges this in the Limitations but the central transfer claims depend on these labels being meaningful.
  • domain assumption The downstream evaluation cohorts are disjoint from the pretraining cohort.
    The paper asserts independence but does not show de-duplication. Both populations are Fitbit/Pixel users recruited in similar periods, so participant overlap is plausible and would inflate transfer results.
  • domain assumption One-minute aggregated sensor features retain sufficient signal for the health outcomes studied.
    The model never sees raw waveforms (e.g., beat-to-beat variability, sub-second motion). The paper notes this limitation, but the claim of a 'universal representation' is bounded by this information loss.

pith-pipeline@v1.3.0-alltime-deepseek · 53704 in / 11403 out tokens · 140781 ms · 2026-08-02T13:21:42.343710+00:00 · methodology

0 comments
read the original abstract

While ubiquitous wearable sensors capture a wealth of behavioral and physiological information, effectively transforming these signals into personalized health insights is challenging. Specifically, converting low-level sensor data into representations capable of characterizing higher-level states is difficult due to high phenotypic diversity and variation in individual baseline health, physiology, and lifestyle factors. Moreover, collecting wearable data paired with health outcome annotations is laborious and expensive, and retrospective annotation remains practically unfeasible, contributing to a scarcity of data with high-quality labels. To overcome these limitations, we propose a foundation model for wearable health that is pretrained on more than one trillion minutes of unlabeled sensor signals drawn from a large cohort of five million participants. We demonstrate that the joint scaling of model capacity and pretraining data volume leads to systematic improvements in performance, as evaluated on a diverse set of 35 health prediction tasks, spanning cardiovascular, metabolic, sleep, and mental health, as well as lifestyle choices and demographic factors. We find that this population scale representation unlocks label-efficient few-shot learning and generative capabilities for robust daily metric estimation. To further leverage this learned representation, we deploy a classroom of LLM agents to autonomously search the space of downstream predictive heads built on the model embeddings, showing broad performance improvements that increase with LLM model capacity. Finally, we show how integrating these downstream predictors into a Personal Health Agent can support model responses that are more relevant, contextually aware, and safe, and we validate this via 1,860 ratings from a cohort of clinicians.

Review history (3 revisions) →

discussion (0)

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

Forward citations

Cited by 2 Pith papers

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

  1. Kepler-Encoder-v0.1: Towards a Multimodal Embedding Model for Robots

    cs.RO 2026-07 conditional novelty 6.0

    A self-supervised multimodal encoder trained with vision, proprioception, and force yields a vision-only latent that recovers end-effector state and force above vision baselines on RH20T, with modest absolute force accuracy.

  2. OpenMHC: Accelerating the Science of Wearable Foundation Models

    cs.LG 2026-06 conditional novelty 6.0

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

Reference graph

Works this paper leans on

22 extracted references · cited by 2 Pith papers

  1. [1]

    **large, uninterpretable embeddings (hundreds of features)** from a foundational model pre-trained on wearable data

  2. [2]

    ### Task Implement a Python function`fit_and_predict`that takes training features (X_train), training labels (y_train), and validation features (X_val)

    demographic information. ### Task Implement a Python function`fit_and_predict`that takes training features (X_train), training labels (y_train), and validation features (X_val). All these inputs are pandas DataFrames. The function should handle the entire training process internally and return the final predictions for the validation data. The function mu...

  3. [3]

    **Instructional Guidelines**

    **Short Length:** Keep entire response to 1 short paragraph with 2-4 sentences with the most relevant features. **Instructional Guidelines**

  4. [4]

    **Wearable Device Sensor Embeddings:** Latent features from a model encoding sensor data

  5. [5]

    **Demographics:** Age, Body Mass Index (BMI), Gender, etc. Total Samples: 1220, Features: 1540, Target:`hypertension_binary` #### Features (X) <STATISTICS OF FEATURE COLUMNS ON TRAIN SET> #### Target Variable (y) #####`hypertension_binary` <STATISTICS OF TARGET COLUMN ON TRAIN SET> ### Metrics: You will receive an **overall score** and scored metrics as f...

  6. [6]

    This DataFrame should have the same index as`X_val`and contain a single column named'predictions'

    **A pandas DataFrame** containing the predictions. This DataFrame should have the same index as`X_val`and contain a single column named'predictions'

  7. [7]

    Model: Random Forest (depth=5, n=100)

    A string (can be empty) containing any analysis, notes, feature importance, or other metadata you want to record. This is for`execution_feedback`. An example function is provided below: <EXAMPLE LINEAR PROBE CODE> ### Goal: Iteratively refine your pipeline to produce a **simple, robust model** that achieves the best`overall_score`for hypertension_binary p...

  8. [8]

    Do not include introductory filler

    **Direct Answer First:** Address the user's specific query clearly in the opening sentence. Do not include introductory filler

  9. [9]

    Do not do more

    **Follow-up Interpretation/Action:** Add 1-3 more sentences elaborating on the interpretation and action. Do not do more

  10. [11]

    Do not list out unrelated metrics (e.g., do not mention sleep or HRV if the query is strictly about blood sugar)

    **Ruthless Prioritization:** Focus EXCLUSIVELY on the data points most pertinent to the user's query. Do not list out unrelated metrics (e.g., do not mention sleep or HRV if the query is strictly about blood sugar). Eliminate all distracting filler data

  11. [12]

    **Precision:** If discussing demographics or wearable aggregate info, include exact numbers

  12. [13]

    **Protect AI Predictions:** NEVER output exact regression values or explicit boolean (true/false) flags from the AI Models

  13. [14]

    The model flags a potential trend to monitor

    **Appropriate Use of AI Predictions:** If AI Model Predictions are present, actively use them to drive your insights and to help paint a holistic picture. Interpret them qualitatively (e.g., "The model flags a potential trend to monitor ..." or "your predictive profile aligns with ..."). This should be done in a way that broadly explains what the predicti...

  14. [15]

    How can I improve my health?

    **Synthesis:** Don't just list facts. Explain the relationship between their metrics. For example, explicitly link how their specific lifestyle data (wearables) is influencing their physiological state or predictive risks for their specific age/demographic. --- **Current User Query:** "How can I improve my health?" Provide your response strictly adhering ...

    2022

  15. [16]

    What is your current employment status? □Full-time □Part-time □Contract / Temporary □Unemployed □Unable to work □Choose not to answer

    Currently Working. What is your current employment status? □Full-time □Part-time □Contract / Temporary □Unemployed □Unable to work □Choose not to answer

  16. [17]

    Disability. Do you identify as having a disability as defined under the Americans with Disabilities Act? The ADA defines a person with a disability as a person who has a physical or mental impairment that substantially limits one or more major life activity. □Yes □No □Prefer not to answer

  17. [18]

    Does your disability affect how you work? □Yes □No □Prefer not to answer

    Disability Affects Work. Does your disability affect how you work? □Yes □No □Prefer not to answer

  18. [19]

    Are you a smoker? □Yes □No

    Smoking. Are you a smoker? □Yes □No

  19. [20]

    Have you been diagnosed with any of the following? Select all that apply

    Diagnoses. Have you been diagnosed with any of the following? Select all that apply. □Diabetes □High blood pressure (hypertension) □High cholesterol (Hyperlipidemia or hypercholesterolemia) □Cardiovascular disease □Kidney condition □Respiratory condition (e.g. asthma, COPD, sleep apnea) 69 Towards a General Intelligence and Interface for Wearable Health D...

  20. [21]

    Diabetes Medication. Do you take any of the following diabetes medications? □Blood thinners □Beta blockers □Daily aspirin □Blood pressure medications □Statin or other cholesterol lowering medications □Heart medications □Antidepressant or antianxiety medications □Metformin or other oral diabetes drugs □Insulin □Hypothyroidism drugs □Hyperthyroidism drugs □...

  21. [22]

    Do you take any of the following medications? Select all that apply

    Medications. Do you take any of the following medications? Select all that apply. □Metformin (e.g. Glucophage) □Other oral diabetes medications □Insulin □I do not take any diabetes medication 70 Towards a General Intelligence and Interface for Wearable Health Data Survey ED.3: Patient Health Questionnaire (PHQ-8) Little interest or pleasure in doing thing...

  22. [2017]

    Head” Development To search the space of task-specific prediction

    to quantify the contribution of each embedding dimension to specific downstream tasks. As previously mentioned, we utilized a Principal Component Analysis (PCA) preprocessing step before the linear probing heads to reduce dimensionality to allow for more appropriate comparisons with the baseline models, as well as reducing collinearity of the raw embeddin...

    2026