Studying from a trillion minutes of sensor information
To construct the pre-training corpus, we sampled de-identified information from 5 million individuals who had consented to using their information for well being and wellness analysis, captured between September 2024 and September 2025. The dataset spans greater than 100 international locations, all 50 U.S. states, and over 20 Fitbit and Pixel Watch machine fashions. From every individual we drew a number of weeks of information, yielding over two billion hours — greater than a trillion minutes — of minute-resolution indicators.
SensorFM ingests 34 one-minute mixture options derived from 5 sensor modalities: photoplethysmography (PPG), accelerometry, electrodermal exercise (EDA), pores and skin temperature, and altimetry. Collectively these seize coronary heart fee and heart-rate variability, blood-oxygen saturation, sleep phases, movement and steps, pores and skin conductance, and temperature over a full 24-hour window.
Slightly than counting on labels, SensorFM learns via self-supervised reconstruction, constructing on the LSM-2 method and its Adaptive and Inherited Masking (AIM) framework. This can be a essential design alternative, as a result of lacking and fragmented information (e.g., stretches of time the place information isn’t out there) is the norm with wearable units, attributable to a wide range of components akin to sensors’ power-cycle, units coming off the wrist, energy saving modes of operation, and sensors switching on and off. Typical self-supervised strategies assume full, uninterrupted inputs and so are pressured to both impute the gaps (which might introduce bias) or discard incomplete home windows (which throws away priceless information). AIM takes neither path: it treats real-world missingness as a pure artifact and learns immediately from incomplete recordings, combining the tokens inherited from real gaps with these artificially masked for the reconstruction goal and treating the 2 as equal. The result’s a illustration that’s missingness-aware by development. SensorFM doesn’t simply tolerate fragmented information, it makes use of it productively, because the generative outcomes beneath present.
