Coaching and analysis
We leverage a dataset with 40 million hours of wearable information sampled from over 60,000 members throughout the interval from March to Could 2024. The dataset was totally anonymized or de-identified to make sure that participant data was eliminated and privateness was maintained. Topics wore quite a lot of Fitbit and Google Pixel smartwatches and trackers and consented for his or her information for use for analysis and growth of latest well being and wellness services. The topics had been requested to self-report intercourse, age, and weight.
To pre-train LSM-2, we make use of the AIM SSL approach launched within the earlier part. AIM implements a masked reconstruction coaching goal, and learns to grasp information that’s naturally lacking, and impute information that’s artificially masked. This unified framework permits LSM-2 to be taught the underlying construction (together with missingness) inherent in wearable sensor information.
We curate a set of downstream duties to judge the pre-trained mannequin, utilizing meta-data that was collected alongside the sensor alerts for the needs of analysis and growth. These embrace consumer annotated actions from a set of 20 completely different classes (corresponding to working, snowboarding, kayaking and enjoying golf) and self-reported diagnoses of hypertension and anxiousness. These information had been break up into fine-tuning and analysis units the place information from every particular person was solely in both the tuning or the analysis set and never each. Information from people used within the pretraining stage was additionally not included within the fine-tuning or analysis phases.
The generative capabilities of LSM-2 are evaluated by means of the duties of random imputation, temporal interpolation, temporal extrapolation (forecasting), and sensor imputation, described in our LSM-1 work.
The utility of the LSM-2 embeddings are evaluated through linear probe on numerous discriminative duties. Particularly we gauge the applicability of the LSM-2 embeddings to the duties of binary hypertension classification, binary anxiousness classification, and 20-class exercise recognition. We consider LSM-2’s potential to mannequin physiology through age and BMI regression duties.