Google Research Introduces SensorFM: A Wearable Health Foundation Model Pretrained on One Trillion Minutes of Sensor Data

Short summary

Google Research and DeepMind introduced SensorFM, a wearable health foundation model using a ViT-1D masked-autoencoder backbone pretrained on over one trillion minutes of sensor data from five million consented participants. Frozen embeddings with a PCA-50 linear probe outperformed feature-engineered baselines on 34 of 35 health prediction tasks. The team also ran an agentic search over 30,516 prediction heads and conducted clinician evaluations to ground a Personal Health Agent.

• •SensorFM is a ViT-1D masked-autoencoder foundation model trained on 1T minutes of sensor data from 5M participants
• •Frozen embeddings plus PCA-50 linear probe beat feature-engineered baselines on 34 of 35 tasks
• •Agentic search across 30,516 prediction heads and clinician evaluations support a Personal Health Agent concept

Generated with AI, which can make mistakes.