A sign is posted at Google headquarters on February 04, 2026 in Mountain View, California. Justin Sullivan/Getty Images

Google Research published SensorFM on Wednesday, a foundation model for wearable health that predicts cardiovascular risk, depression, anxiety, and 32 other conditions from consumer smartwatch signals — and does it as accurately as ground-truth clinical lab data when powering an AI health agent. That single finding upends an assumption that has organized digital health AI for a decade: that meaningful health prediction from wrist-worn sensors requires dedicated, laboriously labeled datasets for every clinical endpoint. SensorFM suggests it does not — and the trillion-minute training corpus behind the model means that any organization hoping to match it needs the kind of data ecosystem that very few institutions in the world actually own.

The preprint, by lead author Girish Narayanswamy and 39 co-authors from Google Research, Google DeepMind, and academia, appeared on arXiv in May 2026 and was accompanied by a Google Research Blog post from Senior Research Scientist Xin Liu and Staff Research Scientist Daniel McDuff.

Single-Task Models Cannot Scale to 35 Endpoints

For roughly a decade, wearable health AI has been built one outcome at a time. A research team targeting atrial fibrillation builds one supervised pipeline. A team targeting sleep apnea builds another. A team targeting depression builds a third. Each pipeline demands its own architecture, its own hyperparameter search, and most critically, its own labeled training data — confirmed diagnoses, lab results, validated questionnaires — which must be collected prospectively over months or years.

That approach "breaks down at thirty-five endpoints," as the SensorFM paper puts it. Labels are expensive, slow, and impossible to gather retrospectively from data already collected. The result is a field where AI models improve quickly within narrow lanes while the dream of a general-purpose physiological intelligence remains perpetually out of reach.

SensorFM applies the same solution that transformed natural language processing and computer vision: pretraining a large model on vast amounts of unlabeled data to learn a general-purpose representation, then adapting that representation for specific downstream tasks with minimal labeled examples. The novelty is making this work for wearable sensor data, which introduces obstacles that text and images do not share.

How SensorFM Processes 34 Features Across Five Sensors

The model's architecture is built around a Vision Transformer adapted for one-dimensional time series — a ViT-1D encoder trained with a masked autoencoder objective at a patch size of [20, 1]. Rather than processing raw sensor output directly, SensorFM ingests 34 aggregate features per minute derived from five sensor modalities: photoplethysmography (PPG) for heart rate and heart-rate variability; accelerometry for motion and step counts; electrodermal activity (EDA) for skin conductance; skin temperature; and altimetry. These 34 features are organized into seven categories and presented to the model across a 24-hour context window, giving SensorFM a full daily arc of physiological behavior — not isolated snapshots — to reason over.

The pretraining corpus was assembled from de-identified data from five million people who had opted in to having their data used for health and wellness research. Google sampled from Fitbit and Pixel Watch users across all 50 U.S. states and more than 100 countries, spanning more than 20 device models, with data collected between September 2024 and September 2025. The resulting dataset totals more than two billion hours — over one trillion minutes — of continuous, minute-resolution sensor signals. To Google's knowledge, it is the largest and most diverse wearable dataset used to train any model to date.

Adaptive and Inherited Masking: Learning From Data Other Systems Discard

Applying standard self-supervised learning to wearable data runs into an immediate problem: real wearable data is never truly complete. Sensors power-cycle. Batteries die. Users take their devices off to charge or shower. Conventional masked autoencoder approaches — which train by hiding portions of otherwise complete inputs and learning to reconstruct them — assume data that exists. When significant fractions of a training corpus are simply missing, those methods either impute the gaps (introducing bias) or discard incomplete windows (wasting data).

SensorFM inherits a solution from Google's earlier LSM-2 model: Adaptive and Inherited Masking (AIM). Rather than treating real-world gaps as a problem to paper over, AIM treats them as a signal. It combines the tokens inherited from genuine missingness with tokens artificially masked for the reconstruction objective, treating both equivalently. The model is trained to reconstruct both kinds of absent data, which means it learns not just what physiological signals look like, but what they look like when they're gone. The result is a representation that is missingness-aware by construction.

The practical payoff is striking. SensorFM-B, the largest model variant, improves random imputation by 74.8% against the best baseline, and improves sensor signal imputation by 83.7%. AIM is also a methodological contribution with potential beyond health: any field that deals with naturally fragmented time-series data — environmental monitoring, industrial sensors, climate systems — could benefit from similar approaches.

Scaling Pays Off: No Saturation Detected After 10,000x Data Range

Google ran scaling experiments across four orders of magnitude in both model size (roughly 100K to 100M parameters) and pretraining data volume (roughly 2 million to 2 billion sensor-hours). The results are among the most important findings in the paper: performance improves predictably across the entire range, and the improvement shows no sign of saturating.

SensorFM-B, the largest variant trained on the full five-million-person corpus, reduces reconstruction loss by 31% over the smallest variant. More critically, those pretraining gains carry over downstream: SensorFM-B improves average AUC by 9% on classification tasks and average Pearson correlation by 21% on regression tasks compared to the smallest variant. Scaling data and model capacity together — rather than either dimension alone — produces near-linear downstream gains.

One failure mode in the data tells the story cleanly. SensorFM-B trained on only 5,000 participants produces a validation loss of 1.082 — worse than every smaller variant at the same volume — because the large model overfit to a small sample. Maximum capacity requires proportional data. The diagonal of co-scaled data and capacity is where the gains live.

For anyone assessing the competitive implications of this work, that scaling curve matters enormously. It means more data keeps helping. Entities with larger ecosystems of consented sensor data have a permanent structural advantage that grows rather than plateaus.

One Representation, 35 Health Tasks: Results Across All Domains

To test the generality of SensorFM's learned representation, Google researchers evaluated it across 35 discriminative health prediction tasks drawn from three independent, IRB-approved prospective studies totaling 13,985 participants. The three studies cover different clinical domains: metabolic, cardiac, and respiratory health (1,655 participants); sleep (6,377 participants); and mental health, including depression and anxiety (5,953 participants). The 35 tasks span six categories: cardiovascular health, metabolic risk, mental health, sleep characteristics, demographic factors, and lifestyle traits.

The evaluation methodology is deliberately stringent. Rather than fine-tuning the entire model for each task, the researchers kept the SensorFM encoder frozen and trained only a lightweight linear classifier on top of the frozen embeddings. A linear probe can be beaten by any moderately competent classifier if the underlying embeddings are not genuinely informative; it can only win if the frozen representation already captures the relevant clinical signal.

SensorFM's linear probes beat the feature-engineered supervised baseline on 34 of the 35 tasks without any task-specific architecture. The gains are largest for mental health conditions — depression and anxiety — where population-level signals are typically buried under individual variation. The paper suggests that at scale, the model may be learning to represent individual physiological patterns in ways that allow it to pick up on person-specific depression signatures that conventional approaches miss.

Can an AI Agent Tutor Itself to Build Better Prediction Heads?

Training a linear probe on frozen embeddings is not the only way to adapt a foundation model for a new clinical endpoint. For production use, a more capable prediction head — one that captures nonlinear relationships in the embedding space — will typically outperform a linear classifier. But constructing those heads traditionally requires manual architecture selection and hyperparameter search across dozens of tasks.

Google's solution is an agentic "classroom" of collaborating large language models from the Gemini family. These agents iteratively generate, execute, score, and refine Python code to build prediction heads on top of the frozen SensorFM embeddings, competing and collaborating until a well-performing solution emerges. Across the experiments described in the paper, the classroom explored more than 30,000 candidate solutions.

Agent-designed heads beat the linear probe on 16 of 20 classification tasks and 12 of 15 regression tasks. Two patterns stand out: solution quality improves monotonically throughout the search, and it scales with the capability of the underlying language model. More capable Gemini variants produce better solutions, while collaboration between multiple agents helps less capable models close the gap — a result with implications for how automated machine learning pipelines may improve as the models powering them improve.

AI-Inferred Predictions Match Ground-Truth Clinical Data for Health Agent Use

The paper's most consequential experiment asks whether SensorFM's predictions are good enough to substitute for real clinical measurements when grounding a personal health agent — an AI system that generates individualized health summaries and coaching recommendations.

Google integrated SensorFM into its Personal Health Agent framework and compared three conditions for generating health summaries from 31 real participant profiles: summaries grounded in SensorFM's inferences; summaries grounded in actual ground-truth lab measurements and clinical data; and summaries generated from only basic demographics and daily wearable metrics, without either source of grounding.

A panel of four board-certified physicians, blinded to which condition generated each summary, rated the outputs across five dimensions: context, relevance, justifiability, personalization, and potential for harm. The evaluation produced 1,860 ratings across more than 40 hours of expert review.

The core finding: adding SensorFM predictions significantly improved clinician-rated summary quality on every dimension compared to the ungrounded baseline. More strikingly, there was no statistically significant difference between summaries grounded in SensorFM's inferences and summaries grounded in actual clinical ground-truth measurements. For the purpose of powering a health coaching agent, the model's predictions were as useful as having the real lab results.

The researchers' own caution deserves emphasis: SensorFM is appropriate for screening and risk stratification, not clinical diagnosis. The blinded physician evaluation assessed the quality of health summaries — not the accuracy of specific medical claims, and not the safety of clinical decisions made on the basis of those summaries. The gap between performing well in a controlled research evaluation and performing well in real-world deployment across diverse clinical populations is one that has undone many promising health AI systems before.

Data Moat: Why These Scaling Curves Are a Structural Fact, Not Just a Research Result

SensorFM is a research paper, and the caveats appropriate to research papers apply. The model has not been independently validated outside Google's own research environment. The evaluation populations, while large and multi-national, are Fitbit and Pixel Watch users — a population that is not representative of every demographic group, socioeconomic stratum, or disease burden globally. The paper's authors do not claim SensorFM is ready for clinical deployment.

What the scaling experiments do establish is something harder to qualify: the relationship between data volume and model capability has not plateaued in this domain. That means the relevant competitive question for wearable health AI is not which team has the best algorithm today — it is which institution has ongoing, consented access to the most longitudinal biometric data from the largest and most demographically diverse population. Google, through Fitbit and Pixel Watch, currently holds a position in that question that academic medical centers, competing wearable companies, and independent health AI researchers cannot easily replicate.

This is not a flaw in the SensorFM research. It is a structural consequence of where foundation model scaling logic leads when the limiting resource is not compute or architecture but rare, legally-consented, continuous physiological data from tens of millions of participants. Anyone evaluating SensorFM's significance for the wearable health field should read its scaling curves not only as evidence about this model's capabilities, but as a statement about what it will take to build a competitive one.

Privacy Framework for Fitbit Health Research Data

Fitbit's consent mechanism requires users to explicitly opt in to having their data used for health research. Google has committed publicly to not using Fitbit health data for advertising purposes — a condition attached to regulatory approval of its 2021 acquisition of Fitbit. Research data is de-identified before use.

Those commitments operate alongside a documented history of regulatory scrutiny. In 2023, European privacy advocacy group noyb filed GDPR complaints against Fitbit in Austria, the Netherlands, and Italy, alleging that the company's data transfer consent framework was coercive — that users faced losing all their tracked health history if they refused to consent to international data transfers, making consent neither free nor genuinely informed. Those complaints were referred to Ireland's Data Protection Commission under the GDPR's one-stop-shop mechanism; their resolution as of this writing remains pending.

The SensorFM pretraining corpus used data from the separate Fitbit Health Labs research consent framework rather than the general Fitbit data product pipeline. But the legal and regulatory context around biometric data collected from consumer devices — particularly health data — is actively contested across multiple jurisdictions, and SensorFM's scale puts that context in unusually sharp relief.

What This Means for Wearable Health AI Going Forward

SensorFM represents a genuine methodological advance on two fronts. AIM is a principled solution to the problem of learning from fragmented sensor data — a problem that is not unique to wearables and has applicability wherever time-series data arrives with real-world gaps. The demonstration that foundation model scaling logic applies to wearable physiological data, with clear performance gains across six clinical domains and no saturation, is a significant empirical result.

At the same time, the field now faces a structural question that SensorFM's own success makes sharper: if more data keeps helping, and the largest consented biometric datasets are concentrated in the ecosystems of a few companies, what does that imply for who gets to build the next generation of wearable health AI — and for how health data governance should be approached at a policy level? Those questions are not answered by a research paper, but they are made more urgent by one.

The full paper, "Towards a General Intelligence and Interface for Wearable Health Data" by Narayanswamy et al., is available on arXiv (2605.22759).

Frequently Asked Questions

What is SensorFM and how is it different from earlier health AI models?

SensorFM is a foundation model — a large AI model pretrained on unlabeled data that can be adapted to many downstream tasks — applied specifically to wearable sensor data. Most earlier wearable health AI models were built for a single clinical endpoint: one model for sleep apnea, another for atrial fibrillation, another for depression. Each required its own labeled training data and architecture. SensorFM learns a single general-purpose representation of human physiology from over one trillion minutes of continuous Fitbit and Pixel Watch signals, then adapts that representation to 35 different health prediction tasks using only a lightweight linear classifier — without task-specific architecture or large amounts of labeled data.

Can a smartwatch now actually predict depression or anxiety?

SensorFM's linear probes outperformed feature-engineered supervised baselines on mental health prediction tasks including depression and anxiety, as evaluated across 5,953 participants in an IRB-approved prospective study. The paper suggests that at scale, the model may be learning individual physiological patterns that carry signals for mental health conditions that population-level features typically obscure. However, SensorFM is designed for risk stratification and screening — not clinical diagnosis. These results were obtained in controlled research populations; independent validation in real-world clinical settings has not been conducted, and the model's outputs are not approved for diagnostic use.

Does Google use Fitbit health data to train AI, and is that legal?

SensorFM was pretrained on de-identified data from Fitbit and Pixel Watch users who had explicitly opted in to Google's health research consent program. Google has committed, as a condition of its 2021 Fitbit acquisition, not to use Fitbit health data for advertising. Separately, European privacy advocacy group noyb filed GDPR complaints against Fitbit in 2023, alleging that the company's consent framework for international data transfers was coercive and did not meet GDPR standards. Those complaints are under investigation by Ireland's Data Protection Commission; no ruling has been issued as of this writing.

Why does Google's scale advantage matter more than its algorithm advantage?

SensorFM's own scaling experiments show that performance improves predictably across four orders of magnitude in both model size and data volume, with no sign of saturation. That means more data keeps helping. Because the limiting resource for wearable health foundation models is not compute or algorithmic sophistication but continuous, consented biometric data from large, diverse populations, organizations with larger wearable ecosystems gain a structural advantage that compounds over time. Google's access to data from Fitbit and Pixel Watch users — across 100+ countries and 20+ device models — represents a training resource that academic labs and smaller competitors cannot replicate without similar scale of user adoption and consent infrastructure.