Google Research has introduced SensorFM, a foundation model that learns health patterns from wearable devices by analyzing over 1 trillion minutes of sensor data collected from 5 million people across 100+ countries. The breakthrough addresses a fundamental problem in digital health: most wearable models are built to predict one health outcome at a time, which becomes impractical when trying to predict dozens of conditions simultaneously. SensorFM flips this approach by learning general patterns from massive amounts of sensor data first, then adapting to specific health tasks with minimal labeled examples.

How Does SensorFM Learn From Wearable Sensors?

SensorFM processes data from five types of sensors commonly found in smartwatches and fitness trackers: heart rate monitors (PPG), accelerometers, electrodermal activity sensors, skin temperature sensors, and altimeters. The model ingests 34 different measurements derived from these sensors, organized into seven categories and analyzed over a 24-hour window. The training data spans from September 2024 through September 2025 and includes information from over 5 million consented participants using more than 20 different Fitbit and Pixel Watch models.

The model uses a specialized neural network architecture called a Vision Transformer adapted for one-dimensional time-series data. During training, the model learns to reconstruct sensor readings that have been intentionally hidden, similar to how a student learns vocabulary by filling in blanks. This self-supervised learning approach means Google didn't need to manually label every data point with a diagnosis or health outcome, which would be prohibitively expensive and time-consuming.

What Health Conditions Can SensorFM Predict?

Google evaluated SensorFM on 35 different health prediction tasks across six categories: cardiovascular health, metabolic conditions, mental health, sleep quality, demographic information, and lifestyle factors. The evaluation used separate data from 13,985 participants across three independent clinical studies approved by institutional review boards. These studies focused on metabolic and cardiac health, sleep patterns, and mental health outcomes.

The largest version of SensorFM, trained on data from all 5 million participants, achieved a mean accuracy score of 0.752 on classification tasks (predicting whether someone has a condition) and a correlation of 0.612 on regression tasks (predicting numerical health metrics). To put this in perspective, the model won 33 of 35 prediction tasks compared to smaller baseline models, demonstrating that scaling up both the model size and training data produces measurably better health predictions.

How Does SensorFM Handle Missing Data From Wearables?

Real-world wearable data is messy. Devices lose connection during charging, people take them off, and battery-saving modes interrupt data collection. Rather than discarding incomplete data or filling gaps with guesses that introduce bias, SensorFM uses a technique called Adaptive and Inherited Masking. The model learns to reconstruct missing sensor signals by treating gaps as a learning opportunity, much like how humans infer what happened during a gap in a conversation based on context.

This approach proved remarkably effective. When tested on reconstructing missing data, SensorFM improved random imputation methods by 74.8% and sensor signal imputation by 83.7%. The model retained 99.7% accuracy on step counts and 99.9% accuracy on deep sleep measurements even when 60 consecutive minutes of data were removed.

Steps to Apply SensorFM for Health Screening and Monitoring

• Screening and Risk Stratification: Healthcare providers can use SensorFM's frozen encoder plus a simple linear prediction head to flag patients who may need confirmatory lab work, without requiring the model to be retrained for each new health system or patient population.
• Repairing Daily Health Summaries: When wearable data contains gaps due to device disconnection or removal, SensorFM can reconstruct missing metrics with high fidelity, ensuring that daily health summaries remain accurate and actionable for users and clinicians.
• Label-Scarce Research Studies: In clinical studies where obtaining diagnostic labels is expensive or time-consuming, researchers can probe SensorFM's learned representations instead of training models from scratch, dramatically reducing the amount of labeled data needed.
• Grounded Health Coaching: AI assistants can generate personalized health insights by combining SensorFM predictions with natural language explanations, avoiding raw numbers or boolean flags that may confuse users.

Can Doctors Trust SensorFM's Predictions?

Google tested whether physicians would find SensorFM predictions useful in real clinical scenarios. Four board-certified doctors, blinded to which predictions came from the model versus actual test results, reviewed health summaries for 31 real patient profiles. The doctors rated summaries on five dimensions including accuracy, usefulness, and confidence. Adding SensorFM predictions significantly improved overall quality compared to summaries with only demographic information and manually engineered metrics. Remarkably, the doctors rated SensorFM predictions as statistically indistinguishable from actual ground-truth test results, with no significant difference detected.

To automate the process of adapting SensorFM to new health tasks, Google's researchers deployed a "classroom" of five large language models (LLMs), ranging from Gemini 2.5 Flash to Gemini 3.1 Pro Preview. These AI agents generated, tested, and refined Python code for 20 iterations, running a total of 30,516 experiments. The AI-generated prediction heads outperformed hand-tuned linear models on 16 of 20 classification tasks and improved correlation scores on 12 of 15 regression tasks. Interestingly, the best solutions remained conservative: most reduced the embedding space to 50 to 100 dimensions, and linear models outnumbered complex non-linear approaches.

What Makes SensorFM Different From Previous Wearable AI Models?

The scale of SensorFM's training data is unprecedented. The model was pretrained on over two billion hours of sensor data, equivalent to more than one trillion minutes. This dwarfs previous wearable health models, which typically trained on much smaller datasets. Google created four variants of SensorFM, each with proportional amounts of training data, ranging from 5,000 participants to 5 million. The largest variant contains 110 million parameters and was trained on data from all 5 million people.

The research shows that scaling matters. The largest SensorFM model reduced reconstruction loss by 31% compared to the smallest variant, and downstream health predictions improved by an average of 28%. However, scaling only works when data volume matches model capacity. When researchers trained the largest model on data from only 5,000 people, it overfitted and performed worse than smaller models, suggesting that foundation models for wearable health require both sufficient model capacity and sufficient training data.

SensorFM represents a shift in how AI approaches digital health. Instead of building separate models for each health condition, foundation models learn general patterns from massive amounts of unlabeled sensor data, then adapt to specific tasks with minimal labeled examples. This approach mirrors recent breakthroughs in large language models and computer vision, bringing the efficiency and flexibility of foundation models to wearable health monitoring.