Training and Evaluation of LSM-2: Leveraging Wearable Data
In the realm of health and wellness technology, training robust models is crucial for effective insights and improvements. This blog post delves into the training and evaluation methods employed in the LSM-2 model, utilizing an extensive dataset of wearable data.
A Comprehensive Dataset
We utilized a unique dataset comprising 40 million hours of wearable data from over 60,000 participants collected between March and May 2024. This dataset has been anonymized to ensure the privacy of participants, who wore various Fitbit and Google Pixel devices and consented to the use of their data for health-related research. Participants also self-reported their sex, age, and weight for a complete profile.
AIM SSL Technique for Pre-training
To pre-train the LSM-2 model, we employed the AIM Self-Supervised Learning (SSL) technique. This innovative method uses a masked reconstruction training objective to understand the intricacies of naturally missing data while also learning to impute artificially masked data. The AIM framework enables LSM-2 to grasp the underlying structures inherent in wearable sensor data.
Curating Downstream Evaluation Tasks
We developed a range of downstream tasks to evaluate the pre-trained model effectively. These tasks include user-annotated activities across 20 categories such as running, skiing, kayaking, and playing golf, along with self-reported diagnoses for conditions like hypertension and anxiety. To ensure unbiased evaluation, we separated the data into fine-tuning and evaluation sets, ensuring no overlap between the datasets used for pre-training and evaluation.
Evaluating Generative Capabilities
The LSM-2 model’s generative capabilities are assessed through various tasks, including random imputation, temporal interpolation, and forecasting. These evaluations build on insights from our previous work on LSM-1, helping us to refine our understanding of wearable data.
Utility of LSM-2 Embeddings
We further evaluated the LSM-2 embeddings using linear probes across several discriminative tasks, which include:
- Binary classification of hypertension
- Binary classification of anxiety
- 20-class activity recognition
In addition, we assessed LSM-2’s ability to model physiological traits through age and BMI regression tasks.
Conclusion
Through the careful selection and assessment of a comprehensive dataset, along with effective training strategies, LSM-2 emerges as a powerful tool for advancing health and wellness technology. Its robust evaluation methods highlight its potential in accurately interpreting complex wearable data, paving the way for further innovations in the field.
Related Keywords: wearable data, health technology, self-supervised learning, model evaluation, Fitbit, Google Pixel, health research.
