Remote Breathing Rate Tracking in Stationary Position Using the Motion and Acoustic Sensors of Earables

Title of the Paper

Remote Sensing of Respiratory Rate in Stationary Conditions Using Motion and Acoustic Sensors in Ear-Worn Devices

Paper Information

• Field of Study: Human Health Monitoring and Mobile Computing
• Keywords: respiration, ear-worn devices, respiratory rate, remote monitoring, wearable devices, motion sensors, acoustic sensors, respiratory rate estimation, signal processing, machine learning

Research Background and Problem Statement

• Problems and Challenges:

  • Respiratory rate is a vital sign, but current detection methods require expensive equipment (e.g., chest straps) or manual counting, which are inconvenient.
  • The potential of ear-worn devices for monitoring respiratory rate has not been fully explored, but background noise and minor non-respiratory head movements reduce accuracy.
  • Respiratory monitoring solutions based on a single sensor lose significant data in noisy environments, resulting in low retention rates.

• Significance:

  • Respiratory rate is a critical indicator for assessing respiratory health and is particularly important for patients with respiratory diseases (e.g., COPD and asthma).
  • Daily respiratory monitoring can help in the early detection of severe disease exacerbations and is closely linked to psychological stress, emotions, and cognitive load.

• Motivation and Related Work:

  • Although some existing studies have demonstrated the potential of smartphones, smartwatches, and ear-worn devices for respiratory monitoring, they often overlook the issue of data retention under noise filtering.
  • Multimodal solutions, as a potential way to reduce noise impact, have not been thoroughly studied, especially in the context of ear-worn devices.

Solution

• Methodology and Innovations:

  • A multimodal approach combining motion sensors and acoustic sensors for respiratory rate monitoring in ear-worn devices is proposed.
  • Signal processing algorithms are used to extract respiratory signals from motion sensor data, while lightweight random forest machine learning models process acoustic data from microphones.
  • The multimodal integration ensures that acoustic data supplements motion sensor data in scenarios with significant head movement where motion sensors alone cannot provide accurate estimates.

• Key Techniques and Steps:

  1. Data Collection: Training data is collected using Samsung Galaxy Buds Pro combined with an FDA-approved chest strap.
  2. Motion Sensor Processing: Signal processing techniques such as zero-crossing rate, fast Fourier transform (FFT), and peak detection are applied.
  3. Acoustic Data Processing: Machine learning classifiers (random forest and multilayer perceptron) detect respiratory "transition points" in sound, optimized with noise filtering algorithms.
  4. Multimodal Integration: When respiratory signals are weak, motion and acoustic sensors complement each other to provide higher retention rates and accuracy.

Research Outcomes

• Specific Results:

  • Demonstrated superior performance in real-world tests with 30 participants (2635 breathing sessions) in both lab and home environments.
  • The multimodal approach achieved an average absolute error (MAE) of <2 BPM and a retention rate of 75.1%.

• Advantages Over Existing Solutions:

  • Improved data retention rates (49.4% and 53.1% for motion and acoustic sensors alone, respectively, compared to 75.1% for the multimodal approach).
  • Comparable or superior accuracy to existing methods, with better adaptability across different postures (standing, sitting, lying down).

• Experimental and Evaluation Results:

  • Performed exceptionally well in noise-free lab environments (MAE < 2 BPM); maintained high retention rates in home environments with background noise through noise filtering algorithms.
  • Achieved stable estimates across a broader respiratory rate range (10–30 BPM) compared to single-sensor solutions.

• Limitations and Future Directions:

  • The current system requires high user compliance, and performance may be affected in real-world scenarios with noise and significant head movements.
  • Performance for abnormal respiratory ranges (e.g., rates above 30 BPM or below 10 BPM) has not been evaluated; future work could adjust parameters to support a broader pathophysiological range.
  • Explore optimized algorithms better suited for long-term passive respiratory rate monitoring (e.g., incorporating duty cycle mechanisms).