Development of a multi-wear-site, deep learning-based physical activity intensity classification algorithm using raw...

Area of Science:

• Biomedical engineering
• Wearable technology
• Machine learning for health

Background:

• Accelerometers are common for measuring physical activity but rely on site-specific algorithms.
• Non-compliance with wear instructions can lead to inaccurate activity intensity estimations.
• Existing algorithms are limited by their dependence on precise device placement.

Purpose of the Study:

• To develop deep neural network models for classifying accelerometer wear-site and activity intensity.
• To evaluate the performance of these models against ground truth and existing count-based algorithms.

Main Methods:

• Trained Long Short-Term Memory deep neural networks on raw acceleration data from 54 participants.
• Collected activity data with accelerometers on hip, wrist, and chest.
• Used portable COSMED K5 to measure metabolic equivalents for ground truth intensity.

Main Results:

• Models achieved over 90% accuracy in classifying wear-sites and activity intensities.
• Performance surpassed traditional count-based algorithms, which had lower accuracy.
• Including participant age, height, and weight improved model accuracy to over 95%.

Conclusions:

• Deep learning models offer accurate, non-wear-site-specific activity intensity classification.
• These models outperform existing count-based algorithms.
• Allowing participant autonomy in accelerometer placement can improve compliance and data accuracy.