SensorFM is a large sensor foundation model that learns directly from unlabeled wearable-device data at population scale. The project aims to produce a single, reusable representation of sensed human physiology that transfers across cardiovascular, metabolic, sleep and mental health, as well as lifestyle and demographic factors.
Data corpus and time coverage
For pre-training the team used de-identified, consented data from five million participants, with data collected between September 2024 and September 2025. The dataset covers more than 100 countries, all 50 U.S. states, and over 20 Fitbit and Pixel Watch device models. From each person several weeks of data were sampled, producing over two billion hours — more than one trillion minutes — of minute-resolution sensor signals.
SensorFM ingests 34 one-minute aggregated features derived from five sensor modalities: photoplethysmography (PPG), accelerometry, electrodermal activity (EDA), skin temperature and altimetry. Together these capture heart rate and heart-rate variability, blood-oxygen saturation, sleep stages, motion and steps, skin conductance and temperature across a full 24-hour window.
Self-supervised learning and the AIM approach
Rather than relying on labeled outcomes, SensorFM is trained with a self-supervised reconstruction objective, building on the LSM-2 approach and its Adaptive and Inherited Masking (AIM) framework. This design choice addresses a core practical issue for wearable data: missing and fragmented recordings are typical (due to power cycles, devices being removed, power-saving modes, sensors toggling on/off). Conventional self-supervised methods assume complete inputs and must either impute gaps (introducing bias) or discard incomplete windows (losing data). AIM treats real-world missingness as a natural artifact and learns directly from incomplete recordings by combining tokens inherited from genuine gaps with tokens artificially masked for the reconstruction objective and treating them equivalently. The resulting representation is missingness-aware by construction and uses fragmented data productively.
Scaling data and model together
The team ran systematic scaling experiments across four orders of magnitude in both pre-training data volume (roughly 2 million to 2 billion sensor-hours) and model size (100K to 100M parameters). Pre-training loss decreased predictably as data and capacity increased, and importantly, those gains translated to downstream health tasks.
The largest variant, SensorFM-B, trained on the full five-million-person corpus, reduced reconstruction loss by 31% relative to the smallest variant and improved downstream performance by an average 9% AUC on classification tasks and 21% Pearson coefficient on regression tasks. The largest gains appeared when scaling data and capacity together: proportional increases produced near-linear improvements in both generative pre-training and discriminative downstream performance, and the curve showed no sign of saturating. Across model variants, SensorFM-B outperformed others on 33 of 35 tasks.
One representation across many health domains
To evaluate generality, SensorFM was tested on 35 discriminative health tasks drawn from three independent, IRB-approved prospective studies totaling 13,985 participants. Tasks covered six categories: cardiovascular health, metabolic risk, mental health, sleep, demographics and lifestyle.
To probe embedding quality directly, the SensorFM encoder was frozen and only a lightweight linear head was trained; results were compared to supervised baselines trained on engineered features. Key findings:
- Broad generalization: linear probes on SensorFM embeddings outperformed the feature-engineered supervised baseline on 34 of 35 tasks without any task-specific architecture.
- Physiology learned implicitly: adding demographic features (age, sex, etc.) offered a modest boost, but that boost diminished as model scale increased, suggesting larger models implicitly captured physiologically relevant traits during pre-training.
- Strength on hard-to-measure conditions: scaled pre-training was especially valuable for conditions such as depression and anxiety that vary across individuals and leave faint sensor traces; SensorFM appears to learn individual differences that help extract cross-person patterns.
- Label efficiency: with only a small fraction of labeled examples, SensorFM quickly surpassed both demographic-only and feature-engineered baselines, an important property when high-quality labels are scarce.
An "agentic classroom" to build prediction heads
Turning a general embedding into a strong predictor for a new endpoint traditionally requires manual feature engineering, architecture selection and hyperparameter tuning. To automate this, the team built an agentic “classroom”: a set of collaborating and competing LLM agents that iteratively generate, test and refine executable code to build prediction heads on SensorFM embeddings. The system explored more than 30,000 candidate solutions.
Agent-designed heads beat a simple linear probe on 16 of 20 classification tasks and 12 of 15 regression tasks. Two patterns emerged: solution quality improved monotonically over the search, and it scaled with the capability of the underlying LLM — more capable models (for example, newer Gemini versions) produced better solutions, while collaboration between agents helped less-capable models close the gap.
Grounding a Personal Health Agent
Finally, the team evaluated whether SensorFM can be used end-to-end to ground an AI health coach in a person’s physiology. They integrated SensorFM into a Personal Health Agent and compared three conditions for generating health summaries from 31 real participant profiles:
- Demographics + daily wearable metrics + SensorFM predictions
- Demographics + daily wearable metrics + ground-truth measurements
- Demographics + daily wearable metrics only (baseline)
A blinded panel of clinicians rated the summaries across five rubric dimensions — context, relevance, justifiability, personalization and potential for harm — producing 1,860 ratings over more than 40 hours of expert evaluation. Adding SensorFM predictions significantly improved ratings over the baseline on every rubric dimension. There was no statistically significant difference between grounding the agent in SensorFM predictions versus actual ground-truth measurements, indicating the model’s inferences served the agent about as well as real labels would.
Conclusion
SensorFM points toward a shift in wearable health research from many bespoke, single-outcome models toward a single, generalist representation of human physiology that can be adapted flexibly, efficiently and at scale. By learning from over a trillion minutes of unlabeled sensor data, SensorFM generalizes across cardiovascular, metabolic, sleep and mental health; supports label-efficient adaptation and robust daily-metric estimation; can be specialized automatically via an agentic classroom; and can ground a Personal Health Agent in a person’s own signals.
Acknowledgements
The authors thank co-authors and collaborators from Google Research, Google DeepMind and academic partners for their contributions to this work.
