Boosting Lying Posture Classification with Transfer Learning.

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Health Monitoring

Background:

• Wrist-based trackers offer unobtrusive health monitoring but struggle with accurate lying posture tracking due to sensor noise.
• Existing research on wrist-based lying posture tracking is limited, especially concerning noisy data from sleep movements.

Purpose of the Study:

• To develop an accurate and efficient lying posture tracking model using noisy data from wrist-based sensors.
• To improve lying posture tracking performance by leveraging deep transfer learning.

Main Methods:

• A deep transfer learning approach was developed, utilizing LSTM sequence regression to map noisy sensor data to clean data.
• Knowledge transfer from a dataset with both clean and noisy data was employed to train the model.
• Clean synthesized data was reconstructed for settings lacking noisy sensor data.

Main Results:

• The proposed method significantly improved lying posture tracking accuracy.
• An increase in F1-Score of 24.9% for left-wrist and 18.1% for right-wrist sensors was achieved compared to baseline methods.
• The deep transfer learning model effectively handled unpredictable noise from wrist movements during sleep.

Conclusions:

• Deep transfer learning offers a viable solution for enhancing lying posture tracking accuracy with noisy wrist-worn sensor data.
• The developed method demonstrates potential for reliable, long-term human health monitoring using wearable technology.
• This approach addresses a critical gap in wearable-based health monitoring by improving posture tracking robustness.