Tackling inter-subject variability in smartwatch data using factorization models.

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Machine Learning for Health

Background:

• Smartwatches offer continuous health data collection for remote monitoring.
• Inter-subject variability in user data presents a significant challenge for generalizable AI models.
• Accurate classification of health states from wearable sensor data is crucial for effective remote health solutions.

Purpose of the Study:

• To address inter-subject variability in smartwatch data for improved health monitoring.
• To evaluate the effectiveness of different data transformation and normalization strategies.
• To propose and validate novel factorization-based autoencoder models for enhanced classification accuracy.

Main Methods:

• Utilized heart rate and step counter data from smartwatches for binary classification tasks (night/day, inactive/active, sleep, SpO2).
• Explored per-subject and population-based time series transforming and normalization techniques.
• Developed and applied a modified factorized autoencoder, including generalized and triplet factorized autoencoder variants.

Main Results:

• The proposed generalized factorized autoencoder improved night/day classification accuracy from 74.8% to 83.1%.
• The triplet factorized autoencoder achieved a similar night/day classification accuracy of 83.4%.
• Modest gains were observed for inactive/active classification, improving from 84.3% to 86.9% and 86.6% respectively.

Conclusions:

• Factorization models effectively address inter-subject variability in smartwatch data.
• The developed autoencoder models offer more robust and personalized remote health monitoring.
• This research paves the way for more reliable health insights from diverse user populations using wearable technology.