SensorFM: Towards a general intelligence and interface for wearable health data (research.google)

Google Research has introduced SensorFM, a groundbreaking foundation model designed to harness data from wearable health devices by utilizing over one trillion minutes of sensor data collected from five million individuals. This innovative model learns a generalized representation of human physiology, enabling it to effectively transfer knowledge across 35 health prediction tasks, including cardiovascular, metabolic, sleep, and mental health assessments. The significance of SensorFM lies in its ability to adapt without relying on extensive labeling, which is often time-consuming and costly, thereby transforming the landscape of wearable health analytics. Technical advancements in SensorFM include its self-supervised learning approach, particularly through the LSM-2 method and Adaptive and Inherited Masking (AIM) framework, which allows the model to handle incomplete data—an inherent characteristic of wearable devices—without introducing bias. Performance results from scaling experiments demonstrate that increasing both data volume and model size leads to substantial improvements in reconstruction and predictive accuracy across various health tasks. Additionally, SensorFM's integration into a Personal Health Agent has shown that it can generate health summaries as reliably as real measurements, reinforcing its potential as a versatile tool in personalized health management. This shift from single-outcome models to a generalized approach heralds a new era in wearable health research, emphasizing efficiency and adaptability.