Training and evaluation
During the period from March to May 2024, we utilized a dataset containing 40 million hours of wearable data sampled from over 60,000 participants. Subjects wore a variety of Fitbit and Google Pixel smartwatches and trackers, and agreed that the data would be used to develop new health and wellness products and services. Subjects were asked to self-report their sex, age and weight.
To pre-train the LSM-2, we employ the AIM SSL technology introduced in the previous section. AIM will implement the objectives of masked reconstruction training, understand data that is missing in nature, and learn to support artificially masked data. This unified framework allows LSM-2 to learn the underlying structures (including missing) specific to wearable sensor data.
A series of downstream tasks are curated to evaluate pre-trained models using metadata collected along with sensor signals for research and development purposes. These include user-annotated activities from 20 different categories (running, skiing, kayaking, playing golf, etc.), as well as self-report diagnosis of hypertension and anxiety. These data were split into fine-tuning and evaluation sets, either in tuning or evaluation sets, rather than both. Data from individuals used during the pre-deletion stage were also not included in the fine-tuning or evaluation stage.
The generation ability of LSM-2 is assessed through the tasks of random assignment, temporal interpolation, temporal extrapolation (prediction), and sensor assignment as described in the LSM-1 task.
The utility of LSM-2 embedding is assessed via linear probes on many identification tasks. Specifically, we evaluate the applicability of LSM-2 embeddings to binary hypertension classification, binary anxiety classification, and 20 classes of activity recognition tasks. Assess the ability of LSM-2 to model physiology via age and BMI regression tasks.
