Google Unveils SensorFM, a Health Foundation Model Built on 1 Trillion Minutes of Data

This research, involving Google Research, DeepMind, and others, proposes SensorFM, which learns human physiological characteristics from vast amounts of unlabeled wearable sensor data.

Key Features and Achievements:

• Large-Scale Dataset: Trained using over 1 trillion minutes (2 billion hours) of multimodal sensor signals collected from 5 million participants.
• Self-Supervised Learning: Unlike existing approaches that target specific diseases, it uses a MAE (Masked Autoencoder)-family architecture that masks and reconstructs portions of the data, learning general-purpose representations without labels.
• High Generalizability: Shows excellent transfer learning performance across various health indicators including cardiovascular, metabolic, sleep, and mental health.
• Automated Adaptation: Introduces a 'Classroom' approach in which an LLM agent automatically explores prediction head code, reducing the manual engineering burden of adapting the model to new tasks.

This model can be used as a core tool for Personal Health Agents, and clinical evaluations have demonstrated performance superior to existing approaches.