Google Research Launches SensorFM, Challenging One-Task Wearable Health Models

Google Research has launched SensorFM after training it on more than one trillion minutes of wearable data from five million consenting participants. The model targets a costly weakness in digital health: most wearable algorithms remain built for one outcome, one dataset, and one device configuration.

The SensorFM research presents a different route. Google says one reusable physiological representation transferred across 35 prediction tasks covering cardiovascular, metabolic, sleep, mental health, demographic, and lifestyle outcomes.

Google Research launches SensorFM as a research system, not a clinical product. That distinction matters because its results come from controlled studies and retrospective evaluations. They do not establish safe performance inside hospitals, consumer coaching apps, or population screening programs.

The competitive reference point is Apple. Its researchers have trained foundation models on Apple Watch data, including behavioral signals gathered from a large consented cohort. Both companies now view wearable data as material for general models, rather than isolated feature pipelines.

The immediate contest is therefore not Fitbit against Apple Watch hardware. It is general-purpose physiological modeling against the established practice of developing a separate supervised model for every health endpoint.

Google Research Launches SensorFM at Population Scale

SensorFM changes the unit of development from an individual health prediction to a reusable representation of human physiology.

Google Research announced SensorFM on July 9, 2026. Senior research scientist Xin Liu and staff research scientist Daniel McDuff introduced the work through the Google Research blog.

The model was pretrained on de-identified data collected between September 2024 and September 2025. Five million participants had consented to health and wellness research involving their data.

The corpus spans more than 100 countries, all 50 US states, and over 20 Fitbit and Pixel Watch models. Google sampled several weeks of data from each participant.

The resulting dataset contains more than two billion hours, equivalent to over one trillion minutes, of minute-level sensor information. That official figure corrects summaries that have described the corpus as 100 trillion minutes.

SensorFM processes 34 aggregate features for each minute of a 24-hour window. Those features come from five sensor modalities: photoplethysmography, accelerometry, electrodermal activity, skin temperature, and altimetry.

Photoplethysmography, commonly called PPG, uses light to measure blood-volume changes near the skin. Wearables derive signals such as heart rate and blood-oxygen estimates from those changes.

Accelerometry captures motion. Electrodermal activity measures changes in skin conductance, while altimetry estimates elevation changes. Together, the inputs describe sleep, movement, temperature, cardiovascular activity, and related daily patterns.

This multimodal design matters because health signals rarely appear in isolation. A higher resting heart rate can have different meanings depending on sleep, movement, temperature, and the individual’s established baseline.

Google evaluated the learned representation across 35 tasks from three prospective studies approved by institutional review boards. Those studies included 13,985 participants in total.

The tasks covered six categories. They included cardiovascular health, metabolic risk, mental health, sleep, demographics, and lifestyle characteristics.

Researchers froze the main SensorFM encoder during a key evaluation. They then trained a lightweight linear predictor on its embeddings, which are compressed mathematical representations of each person’s sensor patterns.

Those linear predictors outperformed supervised baselines built with engineered features on 34 of 35 tasks, according to Google. SensorFM’s largest version recorded the strongest result on 33 tasks.

These results do not mean one model can diagnose 35 conditions. They show that the same representation retained useful information across many research targets without a specialized encoder for each target.

That is the central change. Developers can start with a shared physiological representation, then adapt it using a smaller task-specific prediction layer.

Why One Model for 35 Tasks Matters

SensorFM puts pressure on teams that still rebuild wearable health pipelines for every endpoint and every study.

Traditional wearable models depend heavily on labeled examples. Researchers first define an outcome, collect verified labels, engineer features, and train a supervised model for that specific problem.

That workflow can work well for narrow tasks. However, diagnoses, laboratory results, and validated questionnaires are expensive to collect and difficult to reconstruct after a study ends.

Wearable data presents an unusual imbalance. Devices produce continuous streams of measurements, but researchers obtain relatively few trustworthy labels explaining what those streams mean clinically.

SensorFM uses self-supervised learning to narrow that gap. This approach hides parts of an input and trains the model to reconstruct them without requiring a diagnosis for every recording.

The model therefore learns from the structure of the sensor data itself. Later, researchers can test whether that structure supports a specific prediction using a smaller labeled dataset.

Google reports that SensorFM became more competitive than demographic and feature-engineered baselines while using only a fraction of available labels. The company has not positioned that result as a replacement for clinical validation.

Label efficiency still changes the economics of early research. A team can investigate a new endpoint without first assembling a completely independent large-scale modeling pipeline.

The model’s scale also separates SensorFM from Google’s earlier wearable work. Its predecessor, LSM-2, trained on 40 million hours from more than 60,000 participants.

The earlier LSM-2 framework established a method for learning from incomplete sensor streams. SensorFM expands that method across far more participants, devices, locations, and downstream tasks.

Google tested model sizes ranging from 100,000 to 100 million parameters. It also varied pretraining data across roughly four orders of magnitude, from two million to two billion sensor-hours.

The largest version reduced reconstruction loss by 31 percent compared with the smallest version. Google also reports average gains of 9 percent on classification metrics and 21 percent on regression metrics.

Those percentages describe relative performance improvements reported by the research team. They do not represent clinical accuracy rates, disease-risk reductions, or improvements in patient outcomes.

More important, Google says performance increased when data volume and model capacity grew together. Its reported scaling curves had not reached an obvious plateau.

That result creates pressure on smaller research programs. If general physiological representations improve predictably with population-scale data, organizations with large wearable platforms gain a structural advantage.

Device manufacturers already control the continuous data needed for this approach. They also operate the consent systems, hardware fleets, and longitudinal services required to gather it.

Hospitals possess richer clinical labels, but their wearable data often remains fragmented across vendors and records systems. Startups may have focused datasets, yet few can match platform-scale longitudinal collection.

SensorFM therefore shifts competition toward access, consent, data quality, and adaptation infrastructure. Model architecture remains important, but the training corpus becomes a central strategic asset.

The Real Mechanism Is Learning From Missing Data

SensorFM’s most consequential feature is not its parameter count, but its ability to learn from the gaps that define real wearable use.

Wearable streams are rarely complete. People charge devices, loosen straps, disable sensors, activate battery-saving settings, and leave watches on nightstands.

Hardware configurations also differ. One device may record skin temperature or electrodermal activity, while another provides only motion and optical heart signals.

Conventional self-supervised systems often assume complete inputs. Researchers must then fill gaps with estimated values or remove incomplete samples before training.

Both options create problems. Imputation can introduce artificial patterns, while filtering can discard the users and situations that produce the messiest data.

SensorFM builds on Adaptive and Inherited Masking, known as AIM. The method treats naturally missing data and deliberately hidden training data as related parts of the reconstruction objective.

“Inherited” masks represent gaps already present in a recording. “Adaptive” masking lets the model handle the varying number and position of those gaps during training.

This design makes the learned representation aware of missingness. It does not require every sensor to produce an uninterrupted signal across the full day.

The mechanism originated in LSM-2. Google reported that none of the 1.6 million day-long windows examined for that project contained completely uninterrupted data.

That observation reframes missing data from an edge case into a normal operating condition. A useful wearable model must function when users behave like people, not study instruments.

SensorFM also learns to fill selected missing metrics. In practical terms, it can estimate portions of a sensor stream from surrounding signals and relationships among modalities.

This capability has several research uses. It can support retrospective studies with incomplete recordings, normalize inputs across device models, or recover daily metrics needed by downstream analyses.

However, reconstructed values are still model estimates. They should not be presented as measurements that a device actually captured.

That distinction becomes especially important in health applications. An estimated sleep stage or physiological pattern can support analysis, but it should not silently enter a medical record as observed fact.

Google’s core argument is narrower and more credible. Missingness-aware pretraining preserves more usable information than pipelines that automatically fill or discard every incomplete window.

The model also takes a full 24-hour view. This window lets it relate daytime movement, nighttime sleep, resting patterns, and other daily signals within one representation.

That approach differs from models trained around short events, such as individual heartbeats or brief activity segments. Each granularity supports different questions.

SensorFM’s daily window favors broader physiological and behavioral patterns. It is less obviously suited to detecting an acute event requiring second-level or millisecond-level resolution.

That limitation does not weaken the foundation-model concept. It defines the layer where SensorFM operates: longitudinal interpretation rather than every form of real-time sensing.

SensorFM Versus Apple’s Wearable Model Strategy

Google and Apple are converging on reusable wearable models, but they emphasize different representations and platform advantages.

Apple researchers described a Wearable Behavior Model in 2025. Their work trained foundation models on more than 2.5 billion hours of data from 162,000 participants.

The Apple wearable study evaluated 57 health-related tasks. It emphasized behavioral summaries such as movement, sleep, exercise, and mobility patterns.

SensorFM uses more participants and centers minute-level multimodal signals. Apple’s published behavioral work covers more reported hours because it derives long-term behavior histories from its study cohort.

The figures are not a simple model leaderboard. The companies use different data definitions, cohorts, tasks, architectures, and evaluation protocols.

Apple’s model explores whether behavior can predict health states better than low-level sensor signals alone. Google’s SensorFM focuses on learning a broad physiological representation from multiple sensor modalities.

These routes are complementary as much as competitive. Behavioral summaries capture meaningful activity over longer timescales, while lower-level signals preserve physiological variation that summaries can hide.

Apple also previously developed foundation models for PPG and electrocardiogram signals. That research treated individual biosignal types as large pretraining domains.

Google is now arguing for a unified layer spanning five modalities and many device configurations. Its advantage comes from the Fitbit and Pixel Watch data available across its ecosystem.

Apple’s strategic advantage is similarly clear. Apple Watch, HealthKit, and long-running research studies provide an integrated route from sensor collection to consumer software.

Neither company has established that a general wearable model can safely support routine clinical decisions. Their current results describe research evaluations, not regulated diagnostic products.

The rivalry still matters because each company can connect model research with devices, operating systems, and consumer health interfaces. A successful foundation layer can reduce the cost of adding future health features.

It can also influence the developer interface. Instead of exposing only raw metrics, a platform could eventually provide embeddings, predictions, or agent tools grounded in a user’s longitudinal patterns.

That possibility creates a second competitive boundary. Wearable platforms may compete over who controls the physiological representation between raw sensor data and the final user-facing advice.

Open research efforts offer another route. OpenMHC, introduced in 2026, describes more than 60 million wearable hours and an open framework based on the MyHeart Counts study.

Open models can improve reproducibility and give independent teams a common benchmark. They still face scale, consent, hardware diversity, and clinical-label constraints.

SensorFM’s paper and blog post do not announce public model weights, a general developer API, or a shipping Fitbit feature. That keeps Google’s immediate advantage inside research rather than the broader developer market.

The important comparison is therefore not a consumer feature checklist. It is which organization can turn a general representation into validated, accessible, and governable health tools.

What SensorFM’s Results Do Not Establish

SensorFM’s scale makes it notable, but scale alone cannot resolve representativeness, privacy, clinical validity, or deployment risk.

The pretraining corpus includes consenting participants across more than 100 countries. Geographic breadth does not automatically establish demographic or clinical representativeness.

Wearable owners differ from the general population in income, technology access, age distribution, health engagement, and device usage. Those differences can shape the patterns a model learns.

Google says demographic features provided a smaller incremental benefit as model size increased. The researchers interpret this as evidence that larger models implicitly capture relevant physiological traits.

That finding requires careful handling. Inferring demographic or health-related traits from sensor patterns can improve predictions, but it also introduces privacy and fairness questions.

A de-identified sensor stream can still contain distinctive behavioral and physiological patterns. Researchers have warned that de-identification may not eliminate every re-identification risk in wearable datasets.

Consent also needs a precise scope. Participants agreed to research use, but future commercial or clinical applications would require their own governance, product review, and legal analysis.

SensorFM’s reported downstream studies included 13,985 people. That is a substantial research evaluation, yet it remains far smaller than the five-million-person pretraining corpus.

The evaluations also use selected tasks and study cohorts. Independent replication will need to test different populations, device brands, missing-data patterns, and care settings.

Mental-health predictions deserve particular caution. Google says scale delivered notable benefits for depression and anxiety tasks, where sensor signals can be faint and highly individual.

A stronger research classifier does not create a diagnosis. Mood, sleep, medication, work schedules, disability, and social conditions can produce overlapping wearable patterns.

False positives can cause distress or unnecessary care. False negatives can create reassurance when someone needs professional assessment.

The same caution applies to cardiovascular and metabolic predictions. A model output can identify a statistical association without explaining its cause or supporting a treatment decision.

A 2026 healthcare model review highlights recurring risks around privacy, demographic bias, interpretability, and repeated clinical validation. Those concerns apply directly to wearable foundation models.

Device measurement error adds another layer. Optical heart sensors, skin temperature readings, and movement estimates vary with fit, motion, environment, skin characteristics, and hardware generation.

SensorFM covers more than 20 Fitbit and Pixel Watch models, which helps test device diversity within Google’s ecosystem. It does not establish equivalent behavior on Apple, Garmin, Samsung, Oura, or clinical devices.

The model’s agent experiments require similar restraint. Google integrated SensorFM into a research Personal Health Agent and evaluated summaries from 31 participant profiles.

Clinicians produced 1,860 blinded ratings across five dimensions, including context, relevance, justifiability, personalization, and potential harm. Google says SensorFM grounding improved every dimension over a baseline using daily metrics alone.

The team also reported no statistically significant difference between summaries grounded with SensorFM predictions and those using ground-truth measurements.

That result is interesting, but the sample includes only 31 profiles. A lack of statistical difference does not establish equivalence across larger populations, medical conditions, or longer interactions.

Google itself has previously described its health agent prototype as a research framework. The company said it was not a description of an available product or feature.

Workflow barriers also remain outside the model. Physicians need validated outputs, clear liability, usable interfaces, and integration with electronic health records.

A 2026 survey of 2,222 physicians found continued interest in wearable data alongside major integration barriers. Clinical validation, workflow support, reimbursement, and liability remained unresolved concerns.

SensorFM addresses the interpretation layer, not those institutional constraints. Better embeddings cannot independently create reimbursement policies, medical accountability, or interoperable records.

Three Signals Will Show Whether SensorFM Matters

SensorFM becomes consequential only when independent validation, broader access, and responsible product deployment follow the research results.

The first signal is independent evaluation. Researchers outside Google need access to enough technical detail, code, weights, or standardized outputs to reproduce the reported gains.

The strongest test would compare SensorFM with supervised baselines and other foundation models across unseen institutions, populations, and wearable brands. Results should include subgroup performance and calibration, not only average rankings.

Calibration measures whether a predicted probability matches observed outcomes. A health model can rank people accurately while still producing risk estimates that are unsuitable for decisions.

Independent testing would strengthen Google’s generalization claim if the representation transfers across hardware and cohorts. Weak subgroup results would narrow that claim to specific research populations.

The second signal is developer access. Google has described the paper and methodology, but it has not announced a general SensorFM API or downloadable model weights.

Access terms will reveal whether SensorFM becomes shared infrastructure or remains an internal research asset. Documentation should clarify supported devices, intended uses, data requirements, and prohibited applications.

A useful interface would also distinguish measured data from reconstructed values and inferred health attributes. Developers should not need to reverse-engineer which outputs are observations and which are estimates.

Access will shape competition with Apple and open initiatives. Closed platform models can integrate deeply with devices, while open models allow broader auditing and adaptation.

The third signal is a reviewed product deployment. A Fitbit or Pixel health feature would show how Google translates the representation into language, safeguards, consent controls, and escalation paths.

The key question will not be whether an AI coach sounds personalized. It will be whether users can inspect the evidence, understand uncertainty, correct errors, and reach professional care when necessary.

Personal health tools also need durable context. Users may combine wearable trends with laboratory results, clinician instructions, medication changes, and their own observations.

General personal knowledge systems illustrate the value of connecting information across time. However, ordinary knowledge tools should not replace regulated medical records or professional advice.

Google’s agentic “classroom” offers another signal worth watching within these deployments. Multiple language-model agents generated and refined prediction heads using SensorFM embeddings.

The system explored more than 30,000 candidate solutions. Its generated heads outperformed a simple linear probe on 16 of 20 classification tasks and 12 of 15 regression tasks.

That workflow can reduce manual experimentation, but it also expands the validation burden. Automatically generated code requires reproducibility, leakage checks, security review, and expert oversight.

More capable Gemini versions produced better candidate solutions, according to Google. Collaboration among agents also helped less capable models close part of the performance gap.

This combination links two foundation-model layers. SensorFM represents the body’s sensor history, while language-model agents search for code that adapts those representations to particular questions.

The result is a more automated research stack. It is not an automated medical authority.

SensorFM deserves attention because it attacks a real technical bottleneck with unusual scale. Wearable programs collect far more unlabeled data than clinically verified outcomes.

Google’s answer is to learn one missingness-aware physiological representation, then reuse it across many tasks. Its reported results make that strategy harder for device platforms and digital-health researchers to ignore.

The next proof must come outside the original benchmark. Watch for independent replication, practical access, and a reviewed product with transparent limits.

Until then, Google Research launches SensorFM as a credible challenge to one-task wearable modeling, not as a finished intelligence for personal health. Readers should judge the next announcement by evidence, access, and accountability, not scale alone.