Google's SensorFM trains on 1 trillion minutes of wearable data
Google Research has released details on SensorFM, a foundation model pre-trained on sensor data from five million people that aims to generalize across dozens of health prediction tasks.
What is SensorFM and what can it actually do?
SensorFM is a foundation model for wearable health data, pre-trained on more than one trillion minutes of sensor signals collected from five million consented participants. Google Research says the model learns a single, reusable representation of human physiology that transfers to 35 health prediction tasks, including cardiovascular, metabolic, sleep, and mental health outcomes, without being purpose-built for any one of them.
The data problem it is designed to solve
Most wearable health models are trained in a supervised, task-specific way, which requires confirmed diagnoses, lab results, or validated questionnaires as labels. Those labels are expensive to collect, slow to acquire, and nearly impossible to gather after the fact. SensorFM bypasses that bottleneck by pre-training entirely on unlabeled sensor data, then adapting to specific tasks with far fewer labeled examples.
The researchers describe this as a co-scaling approach: growing both model size and dataset size together, similar to how large language models improved by scaling text data alongside parameter counts.
What the training data looks like
Google assembled the pre-training corpus from de-identified data captured between September 2024 and September 2025. The dataset spans more than 100 countries, all 50 U.S. states, and over 20 Fitbit and Pixel Watch device models. Each participant contributed several weeks of data, yielding more than two billion hours of minute-resolution signals in total.
SensorFM ingests 34 one-minute aggregate features derived from five sensor modalities:
| Modality | Example signals captured |
|---|---|
| Photoplethysmography (PPG) | Heart rate, heart-rate variability, blood-oxygen |
| Accelerometry | Movement, activity levels |
| Electrodermal activity (EDA) | Skin conductance |
| Skin temperature | Thermal variation |
| Altimetry | Elevation changes |
Google says this is, to its knowledge, the largest and most diverse wearable dataset used to train any model to date.
Three things the model is designed to support
Beyond standard health prediction, Google describes two additional use cases that distinguish SensorFM from earlier wearable models.
First, the model supports data infilling, meaning it can reconstruct missing or corrupted sensor windows. Wearable data is notoriously patchy, with gaps from charging, poor sensor contact, or device removal, so this capability has practical value for downstream analysis.
Second, SensorFM is positioned as a grounding tool for a Personal Health Agent. The framing suggests Google envisions the model as a backend component that a conversational health assistant could query to contextualize a user's physiological state, though the source documents do not elaborate on a deployed product.
Why individual variation makes this hard
One of the core challenges the researchers cite is that baseline physiology differs substantially from person to person. A heart-rate pattern that indicates elevated risk in one individual may fall within normal range for another. A model trained to generalize across populations has to learn these individual baselines implicitly, which is part of why the team argues that population-scale pre-training, rather than small cohort studies, is necessary.
The longitudinal nature of the data, spanning weeks per participant rather than isolated snapshots, gives the model exposure to how signals evolve over time, including across sleep cycles, activity patterns, and day-to-day variation.
Label efficiency after pre-training
Google highlights label-efficient adaptation as a key property of SensorFM. Because the model has already learned a rich physiological representation during pre-training, fine-tuning it for a specific outcome, such as detecting a sleep disorder or estimating a metabolic marker, should require substantially fewer labeled examples than training a task-specific model from scratch. The source documents do not provide specific benchmark figures on how many labeled samples are needed.
The work is described in a paper titled "Towards a General Intelligence and Interface for Wearable Health Data," authored by Xin Liu and Daniel McDuff of Google Research.