Google trains SensorFM on a trillion minutes of wearable data

Google Research introduced SensorFM on July 9th, a wearable health foundation model trained on over one trillion minutes of sensor data from five million consented Fitbit and Pixel Watch users. The model uses self-supervised learning to predict health outcomes across 35 tasks, outperforming supervised baselines on 34 of them, marking a significant scale-up from Google's previous wearable AI work.

Google Research @GoogleResearch https://x.com/GoogleResearch/status/2075283854093607016 introduced SensorFM on July 9th, a wearable health foundation model trained on more than one trillion minutes of sensor data from five million consented participants. The release came through a Google Research blog post https://research.google/blog/sensorfm-towards-a-general-intelligence-and-interface-for-wearable-health-data/ by Xin Liu, a senior research scientist, and Daniel McDuff, a staff research scientist. The underlying paper https://arxiv.org/abs/2605.22759 , submitted to arXiv on May 21st, lists 40 authors across Google Research, Google DeepMind and academic collaborators. SensorFM is Google's clearest attempt yet to turn its wearable hardware footprint into a general AI substrate for health prediction. The model was pre-trained on de-identified data captured between September 2024 and September 2025 from Fitbit and Pixel Watch devices, spanning more than 100 countries, all 50 U.S. states and more than 20 device models, according to Google Research. The model ingests 34 one-minute aggregate features from five sensor modalities: photoplethysmography, accelerometry, electrodermal activity, skin temperature and altimetry. Those inputs cover heart rate and heart-rate variability, blood oxygen, sleep stages, motion, steps, skin conductance and temperature over a 24-hour window. Google Research's bet is that wearable health models can move away from narrowly trained classifiers that require expensive labels for each condition. The paper frames the core problem plainly: confirmed diagnoses, lab values and validated questionnaires are slow and costly to collect, while retrospective labeling of historical wearable streams is usually impractical. SensorFM uses self-supervised reconstruction instead, learning from missing and fragmented signals without requiring a label on each minute of data. That design choice matters because consumer wearable data is messy by default. People remove devices. Batteries die. Sensors power-cycle. Some modalities appear in one device generation and disappear in another. Google Research says SensorFM builds on its LSM-2 work and an Adaptive and Inherited Masking framework, treating real gaps in data and artificially masked training tokens as part of the same reconstruction problem. The claimed scale is a step change from Google's previous work in the category. In November 2024, Google Research published scaling work on wearable foundation models https://research.google/blog/scaling-wearable-foundation-models/ using over 40 million hours of de-identified multimodal sensor data from 165,000 users. In July 2025, Google Research followed with SensorLM https://research.google/blog/sensorlm-learning-the-language-of-wearable-sensors/ , a sensor-language model trained on 59.7 million hours of data from 103,643 people. SensorFM pushes the dataset to more than two billion hours, according to Google Research, and shifts the target from activity and language alignment to broad health prediction. The results Google Research reported are strong, with the usual caveat that they come from the authors' own evaluation. The team evaluated SensorFM across 35 discriminative health tasks drawn from three independent, Institutional Review Board approved prospective studies totaling 13,985 participants. The tasks covered cardiovascular health, metabolic risk, mental health, sleep, demographics and lifestyle. Using a frozen SensorFM encoder and lightweight linear heads, Google Research says SensorFM embeddings beat supervised baselines trained on engineered features on 34 of 35 tasks. The largest variant, SensorFM-B, was trained on the full five-million-person corpus and reduced reconstruction loss by 31% versus the smallest variant. Google Research also reported average downstream gains of 9% on classification tasks measured by AUC and 21% on regression tasks measured by Pearson coefficient. The paper then adds a second layer: LLM agents that write the small predictive models sitting on top of SensorFM embeddings. Google Research describes an agentic "classroom" in which multiple agents generate, test and refine executable code for downstream prediction heads. Across more than 30,000 candidate solutions, those agent-designed heads beat a simple linear probe on 16 of 20 classification tasks and 12 of 15 regression tasks, according to the blog post. That is the operational piece. A large wearable representation is useful only if a team can adapt it to a new endpoint without rebuilding a custom pipeline. Google Research is arguing that a general embedding plus automated head search can collapse the manual work traditionally needed for each health prediction task. The most product-shaped part of the release is the Personal Health Agent evaluation. Google Research says it integrated SensorFM predictions into an AI health agent and compared three inputs: demographics plus daily wearable metrics plus SensorFM predictions; demographics plus daily wearable metrics plus ground-truth measurements; and demographics plus daily wearable metrics alone. Clinicians, blinded to the input condition, rated summaries across context, relevance, justifiability, personalization and potential harm, producing 1,860 ratings over more than 40 hours of expert evaluation. Google Research says SensorFM-grounded responses improved over the baseline across every rubric dimension, and that there was no statistically significant difference between responses grounded in SensorFM predictions and responses grounded in actual labels. That result is the most commercially loaded claim in the release: it suggests a wearable model could give a health assistant enough inferred context to behave closer to one that had access to harder-to-collect clinical measurements. The limits are just as important. Google Research announced a research model, paper and blog post, with no public release of the training data. The data advantage comes from consented Fitbit and Pixel Watch users, a population shaped by device ownership, geography, income, platform preference and willingness to share health data for research. Google Research says the data was de-identified, but the work still depends on a proprietary corpus that outside labs cannot reproduce at comparable scale. That is where the founder-market lesson sits, even in a story led by a research lab rather than a startup. SensorFM shows why wearable health AI may be difficult for smaller companies to copy. Model architecture matters, but the durable asset is longitudinal, consented, multimodal data at population scale. Google has that through Fitbit and Pixel Watch. Most health AI startups have pilots, integrations and labeled study cohorts measured in thousands. Google is training on five million people before any startup gets to the starting line. The release also sharpens the regulatory and product boundary Google will have to manage. SensorFM is framed as a research system for prediction, infilling and grounding health summaries, rather than a diagnostic product. The closer those predictions get to mental health, metabolic risk and cardiovascular signals, the more the model moves from wellness analytics toward clinical decision support. Google's paper gives the company a technical claim. Turning that into a product would require a different kind of proof.