Few-Shot User-Adaptable Radar-Based Breath Signal Sensing

Area of Science:

• Biomedical Engineering
• Signal Processing
• Human-Computer Interaction

Background:

• Vital signs monitoring is crucial for health assessment but often requires wearables or invasive methods.
• Contactless and privacy-preserving methods are needed for continuous health monitoring, especially in shared spaces like offices.

Purpose of the Study:

• To develop a radar-based, user-adaptable system for non-contact respiratory signal prediction.
• To enable privacy-friendly vital signs estimation in office environments with minimal training data.

Main Methods:

• Utilized a 60 GHz frequency-modulated continuous wave radar to collect data from 24 subjects.
• Employed episodic optimization with a convolutional variational autoencoder for signal prediction and generalization.
• Incorporated autocorrelation analysis to assess and mitigate motion-induced data corruption.

Main Results:

• Achieved rapid model adaptation, requiring less than 1-2 seconds for 1-5 training examples.
• Demonstrated effective respiratory signal prediction with a contact-free, privacy-friendly approach.
• Showcased the system's ability to adjust predictions based on motion corruption levels.

Conclusions:

• The proposed radar system offers a novel, rapidly adaptable, non-contact solution for respiratory monitoring in office settings.
• This technology minimizes the need for extensive user training and respects user privacy.
• The system effectively handles motion artifacts, enhancing the reliability of vital signs estimation.