Google's SensorFM: A Trillion-Minute Foundation Model That Rivals Clinical Lab Tests for Wearable Health AI
核心洞察
Google Research (搜索)'s SensorFM, trained on over one trillion minutes of wearable sensor data from five million people, predicts 35 health outcomes including cardiovascular risk, depression (搜索), and anxiety (搜索) from consumer smartwatch signals.
In a blinded physician evaluation, AI-generated health summaries grounded in SensorFM predictions were statistically indistinguishable from summaries built on actual clinical ground-truth measurements.
The model's scaling experiments show no performance saturation across four orders of magnitude in data and model size, suggesting entities with larger consented biometric datasets hold a compounding structural advantage.
Google Research (搜索) has published SensorFM, a wearable health foundation model that predicts cardiovascular risk, depression (搜索), anxiety (搜索), and 32 other conditions from consumer smartwatch signals — and when powering an AI health agent, its predictions are as useful as ground-truth clinical lab data. The finding, published as a preprint on arXiv in May 2026, upends a decade-old assumption that meaningful health prediction from wrist-worn sensors requires dedicated, laboriously labeled datasets for every clinical endpoint.
The model was trained on de-identified data from five million Fitbit and Pixel Watch users who explicitly opted into health research, spanning all 50 U.S. states and more than 100 countries across 20-plus device models. The resulting corpus totals more than two billion hours — over one trillion minutes — of continuous, minute-resolution sensor signals collected between September 2024 and September 2025. Lead author Girish Narayanswamy and 39 co-authors from Google Research (搜索), Google DeepMind (搜索), and academia contributed to the work.
From Single-Task Models to a Generalist Architecture
For roughly a decade, wearable health AI has been built one outcome at a time: one supervised pipeline for atrial fibrillation, another for sleep apnea, a third for depression (搜索). Each demands its own architecture, hyperparameter search, and prospectively collected labeled training data. That approach, the SensorFM paper states, "breaks down at thirty-five endpoints."
SensorFM applies the solution that transformed natural language processing — pretraining a large model on vast unlabeled data to learn a general-purpose representation, then adapting it for specific tasks with minimal labeled examples. The architecture uses a Vision Transformer adapted for one-dimensional time series (ViT-1D), trained with a masked autoencoder objective. Rather than raw sensor output, the model ingests 34 aggregate features per minute from five sensor modalities: photoplethysmography for heart rate and heart-rate variability, accelerometry for motion, electrodermal activity, skin temperature, and altimetry. These features are organized across a 24-hour context window, giving SensorFM a full daily arc of physiological behavior to reason over.
Adaptive and Inherited Masking: Learning from Missing Data
Real wearable data is never complete — sensors power-cycle, batteries die, users remove devices. Conventional masked autoencoder approaches either impute gaps or discard incomplete windows. SensorFM inherits a solution from Google's earlier LSM-2 model: Adaptive and Inherited Masking (AIM), which treats real-world gaps as signal rather than noise. The model learns to reconstruct both naturally missing and artificially masked tokens, producing a representation that is missingness-aware by construction.
The practical payoff is substantial. 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 also represents a methodological contribution with potential beyond health, applicable to any field dealing with naturally fragmented time-series data.
Scaling Without Saturation
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). Performance improved predictably across the entire range with no sign of saturating. SensorFM-B reduced reconstruction loss by 31% over the smallest variant, and those pretraining gains carried downstream: a 9% improvement in average AUC on classification tasks and a 21% improvement in average Pearson correlation on regression tasks.
Critically, scaling data and model capacity together produced near-linear downstream gains. SensorFM-B trained on only 5,000 participants produced 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.
35 Health Tasks from One Frozen Representation
SensorFM was evaluated across 35 discriminative health prediction tasks drawn from three independent, IRB-approved prospective studies totaling 13,985 participants covering metabolic, cardiac, respiratory, sleep, and mental health domains. The evaluation kept the SensorFM encoder frozen and trained only a lightweight linear classifier on top — a deliberately stringent test.
SensorFM's linear probes beat the feature-engineered supervised baseline on 34 of 35 tasks without any task-specific architecture. Gains were largest for mental health conditions — depression (搜索) and anxiety (搜索) — where population-level signals are typically buried under individual variation. The paper suggests the model may be learning individual physiological patterns that capture person-specific depression signatures conventional approaches miss.
AI Agents That Build Better Prediction Heads
For production use, Google deployed an agentic "classroom" of collaborating large language models from the Gemini family. These agents iteratively generated, executed, scored, and refined Python code to build prediction heads on frozen SensorFM embeddings, exploring 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, with solution quality improving monotonically throughout the search and scaling with the capability of the underlying language model.
Predictions Match Ground-Truth Clinical Data
The paper's most consequential experiment asked whether SensorFM's predictions could substitute for real clinical measurements when grounding a personal health agent. Google integrated SensorFM into its Personal Health Agent framework and compared three conditions: summaries grounded in SensorFM inferences, summaries grounded in actual ground-truth lab measurements, and summaries from only basic demographics and daily wearable metrics.
A panel of four board-certified physicians, blinded to condition, rated 1,860 outputs across five dimensions: context, relevance, justifiability, personalization, and potential for harm. SensorFM predictions significantly improved clinician-rated summary quality on every dimension compared to the ungrounded baseline. There was no statistically significant difference between summaries grounded in SensorFM inferences and those grounded in actual clinical ground-truth measurements.
The researchers emphasize that SensorFM is appropriate for screening and risk stratification, not clinical diagnosis. The blinded physician evaluation assessed summary quality — not the accuracy of specific medical claims or the safety of clinical decisions based on those summaries.
Structural Implications for Wearable Health AI
SensorFM's scaling curves carry competitive implications. Because more data keeps helping and the limiting resource is consented, continuous biometric data from large, diverse populations, organizations with larger wearable ecosystems gain a compounding structural advantage. Google's access to Fitbit and Pixel Watch data represents a training resource that academic labs and smaller competitors cannot easily replicate.
The model has not been independently validated outside Google's research environment, and the evaluation populations — while large and multi-national — are Fitbit and Pixel Watch users, not representative of every demographic group or disease burden globally. Separately, European privacy advocacy group noyb filed GDPR complaints against Fitbit in 2023 alleging coercive consent frameworks; those complaints remain under investigation by Ireland's Data Protection Commission.
The full paper, "Towards a General Intelligence and Interface for Wearable Health Data" by Narayanswamy et al., is available on arXiv (2605.22759).