Google Research has introduced SensorFM, a foundation model designed to learn general representations of physiological and behavioral patterns from wearable sensor data. The model processes data from Fitbit and Pixel Watch devices and outperformed comparison models on 34 out of 35 health and behavioral tasks. According to the researchers, this is the largest and most diverse wearable dataset ever used to train a model of this kind. The model is trained using over a trillion minutes of multimodal sensor data from five million users across more than 100 countries, collected with more than 20 different Fitbit and Pixel Watch models.
SensorFM processes 34 features drawn from five types of sensor data, including optical heart rate monitoring, acceleration, skin conductance, skin temperature, and barometric altitude. The model is trained in a self-supervised way by reconstructing deliberately masked data segments. The technique, called 'Adaptive and Inherited Masking' (AIM), flags both genuinely missing values and values that were artificially hidden during training, so SensorFM learns to handle both types of data gaps. As model size and pretraining data grow together, SensorFM's performance improves for both pretraining reconstruction and downstream prediction tasks.
The researchers tested SensorFM on data from three separate studies with a total of 13,985 participants, none of whom had seen the data during pretraining. They evaluated the model on 35 prediction tasks covering cardiovascular and metabolic health, mental health, sleep, demographics, and lifestyle. Simple task-specific head models built on top of SensorFM's learned representations outperformed supervised baselines with hand-crafted wearable features on 34 of 35 tasks. Scaled pretraining also made SensorFM more label-efficient compared to the supervised baselines.
Source: thedecoder
