Google Research Unveils SensorFM Wearable Health Foundation Model

Google Research has introduced SensorFM, a foundation model designed for wearable health technology that leverages a massive dataset of one trillion minutes of sensor information. By pretraining on data from five million consented participants across more than 100 countries, the model aims to move beyond the limitations of building individual health models for specific outcomes, instead providing a versatile architecture capable of transferring to 35 distinct health prediction tasks.

Scaling Model Capacity and Data

SensorFM functions as a Large Sensor foundation Model, utilizing a ViT-1D encoder backbone to process time-series data. The model ingests 34 aggregate features drawn from five primary sensors: PPG, accelerometer, EDA, skin temperature, and altimeter. These features are organized into seven categories over a 24-hour context window. To ensure optimal performance, the research team scaled model capacity alongside data volume across four orders of magnitude, ranging from the XXS variant to the B variant.
The results demonstrate that scale significantly improves performance. When comparing the largest model, SensorFM-B, to the smallest variant, the research team observed a 31% reduction in reconstruction validation loss. Across the 35 evaluated tasks—which span cardiovascular, metabolic, mental health, sleep, demographic, and lifestyle categories—the B variant outperformed smaller models in 33 instances. The research indicates that the trend in performance gains has not yet saturated, suggesting further potential for scaling.

Handling Missing Data with AIM

A significant challenge in wearable health data is the frequent occurrence of missing signals due to device charging, off-wrist periods, or power-saving modes. Traditional methods often rely on biased imputation or simply discard data gaps. SensorFM addresses this through Adaptive and Inherited Masking (AIM). By utilizing a two-stage token masking approach, the model computes loss only on artificially masked patches that possess ground truth.
This architecture allows the decoder to learn to reconstruct ablated observations, effectively providing imputation and forecasting capabilities as a native feature. Compared to conventional baselines, SensorFM improves random imputation by 74.8% and signal imputation by 83.7%. This robustness ensures that the model remains effective even when real-world sensor streams are fragmented.

Practical Application and Agentic Integration

To adapt the model for specific predictions, researchers use a frozen encoder and aggregate embeddings per person, which are then reduced to 50 principal components. A linear head is subsequently trained using person-independent cross-validation. This approach proved successful, with the linear probe outperforming supervised feature-engineered baselines on 34 of 35 tasks.
Beyond standard benchmarking, the team explored automating the tuning process using a classroom of LLM student agents. These agents generated and refined Python heads for the model, successfully beating the linear probe on 16 of 20 classification tasks. Furthermore, when tested as a tool for generating health summaries, SensorFM-enhanced predictions were rated by board-certified physicians as statistically indistinguishable from ground truth, highlighting the model's potential for clinical screening and risk stratification.