Google SensorFM Wearable AI: General Model for Health Data

TL;DR: Google’s SensorFM was trained on more than one trillion minutes of Fitbit and Pixel Watch data and beats specialist wearable models on 34 of 35 health prediction tasks. In a blinded physician study, its health summaries were statistically indistinguishable from those based on actual lab tests, signaling a shift toward unified AI health agents.

Key facts

• SensorFM was pre-trained on more than one trillion minutes (over two billion hours) of wearable sensor data from five million consented participants [1][4].
• The training data span Fitbit and Pixel Watch devices across more than 100 countries, all 50 U.S. states and over 20 device models collected between September 2024 and September 2025 [4].
• The model ingests 34 one-minute aggregate features from five sensors – photoplethysmography, accelerometer, electro-dermal activity, skin temperature and altimeter – to learn a reusable representation of physiology without clinical labels [1].
• Evaluation on 35 downstream health tasks across cardiovascular, metabolic, sleep, mental-health, lifestyle and demographic domains showed SensorFM outperformed supervised baselines on 34 tasks, delivering mean gains of 0.09–0.21 AUC/r² and a 28% reduction in mean-squared error on generative tasks [5][8].
• SensorFM can impute missing sensor data 74.8% more accurately than prior methods and retains >99.7% step-count accuracy and >99.9% deep-sleep accuracy even when an hour of data is missing [5][8].
• Scaling experiments revealed near-linear performance improvements as model size grew from 100 K to 100 M parameters, with no sign of saturation across four orders of magnitude in data and model size [4].
• A large-language-model (LLM)-driven agent automatically designed and tested over 30 000 predictive heads to adapt the representation to specific health endpoints [5].
• In a blinded physician evaluation involving 1,860 ratings, health summaries based on SensorFM predictions were statistically indistinguishable from summaries built on actual clinical lab measurements and consistently outperformed a no-data baseline [4][5][8].

A New Kind of Wearable Intelligence

Google Research unveiled SensorFM, a foundation model built to understand the continuous streams of data produced by modern wearables. A foundation model is a large neural network that learns general patterns from massive, unlabeled datasets, then can be fine-tuned for many downstream tasks. SensorFM applies this concept to health-focused sensor streams, aiming to replace the patchwork of single-purpose algorithms that currently power fitness trackers and health apps.

Massive, Global Training Corpus

The model was pre-trained on more than one trillion minutes – over two billion hours – of minute-resolution data gathered from five million consented participants who used Fitbit or Pixel Watch devices. The data span more than 100 countries, all 50 U.S. states and over 20 device models, and were collected between September 2024 and September 2025 [4].

From Raw Signals to a Unified Representation

SensorFM ingests 34 one-minute aggregate features drawn from five sensor modalities: photoplethysmography (PPG), accelerometer, electro-dermal activity (EDA), skin temperature and altimeter. By training on this unlabeled corpus, the model learns a general, reusable representation of physiological and behavioral patterns – a kind of “health language” that can be repurposed for many prediction tasks without hand-engineered features [1].

Benchmarking Across 35 Health Tasks

The research team evaluated SensorFM on 35 downstream tasks covering cardiovascular risk, metabolic markers, sleep quality, mental-health indicators, lifestyle habits and demographic attributes. In 34 of the 35 tasks the model outperformed traditional supervised baselines that rely on handcrafted wearable features. Average gains ranged from 0.09 to 0.21 in AUC or r², and generative tasks saw a 28% reduction in mean-squared error [5][8].

Robustness to Gaps and Noise

Wearable data are often incomplete due to device removal or signal loss. SensorFM can impute missing sensor data 74.8% more accurately than prior methods. Even when an entire hour of data is removed, the model retains over 99.7% step-count accuracy and over 99.9% deep-sleep accuracy, showing strong resilience to real-world gaps.

Scaling Benefits

The authors ran scaling experiments by varying model size from 100 K to 100 M parameters. Performance improved near-linearly across this four-order-of-magnitude range, with no sign of saturation, confirming that larger models and more data continue to yield better health insights [4].

Automated Head Design via an LLM Agent

To adapt the general representation to specific health endpoints, the team employed an LLM-driven agent (a large language model that can generate code and experiment configurations). This agent autonomously designed and tested over 30 000 predictive heads, dramatically reducing the engineering effort required to launch new health applications [5].

Clinical-Grade Summaries in a Blinded Study

A key validation came from a blinded physician study where 1,860 health summaries generated from SensorFM predictions were compared against summaries built on actual clinical laboratory measurements. Physicians rated the AI-generated summaries as statistically indistinguishable from the lab-based ones, and both outperformed a baseline that provided no sensor data [4][5][8]. This result suggests that continuous wearable streams, when interpreted by a model like SensorFM, can deliver insights on par with traditional lab tests.

Implications for Thailand’s Health-Tech Landscape

Thailand has a growing ecosystem of wearable manufacturers, tele-health platforms and government-backed health initiatives. SensorFM’s unified approach could lower the barrier for local startups to build personal health agents that address multiple conditions – from diabetes monitoring to sleep disorder detection – without developing separate models for each metric. Hospitals could integrate SensorFM-derived insights into electronic health records, enabling earlier interventions and more personalized care. Moreover, the model’s ability to handle missing data aligns well with the variable adherence patterns observed in rural Thai populations.

Looking Ahead

SensorFM marks a shift from isolated, single-purpose wearable algorithms toward a general-purpose health intelligence layer. While the model’s performance is impressive, broader adoption will require careful handling of privacy, consent and regulatory compliance, especially in jurisdictions with strict data-protection laws. If these challenges are addressed, SensorFM could become the backbone for the next generation of AI-driven personal health assistants worldwide.

All figures and claims are drawn from Google’s pre-print and accompanying coverage in reputable tech and business outlets.

Sources

4. Google’s SensorFM: A Trillion-Minute Foundation Model That Rivals Clinical Lab Tests for Wearable Health AI (trial.medpath.com) — 2026-07-11
5. SensorFM: Towards a general intelligence and interface for wearable health data (research.google) — 2026-06-30
6. Google’s SensorFM Reveals Where AI Takes Wearable Health (www.forbes.com) — 2026-07-10
7. Wearable AI Now Rivals Lab Tests: Google’s SensorFM Trained on 1 Trillion Minutes (www.techtimes.com) — 2026-07-10