A meta-learning algorithm for respiratory flow prediction from FBG-based wearables in unrestrained conditions

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Artificial Intelligence in Healthcare

Background:

• Continuous breathing monitoring is crucial for health assessment and early detection of health issues.
• Existing wearable solutions often lack comprehensive respiratory data.
• Smart wearables for breathing monitoring are gaining research and commercial interest.

Purpose of the Study:

• To develop and evaluate a meta-learning algorithm for inferring respiratory flow using wearable sensors.
• To compare the proposed algorithm against conventional machine learning methods.
• To assess the algorithm's accuracy and efficiency under various conditions and device configurations.

Main Methods:

• Implementation of a meta-learning algorithm based on Long Short-Term Memory (LSTM) neural networks.
• Integration of Fiber Bragg Grating (FBG) sensors and inertial units in a wearable system.
• Testing and comparison with conventional machine learning approaches under static and dynamic conditions.

Main Results:

• The meta-learning algorithm demonstrated superior accuracy in predicting respiratory flow, especially for new subjects.
• LSTM's memory capability effectively captured breathing pattern nuances.
• High accuracy was achieved with subject-specific models after short calibration, even with reduced sensors (RMSE down to 9.97 L/min).
• Strong correlation (r=0.9) observed with comprehensive sensor configurations; significant results (r=0.84-0.88) also with thoracic FBG sensors alone.

Conclusions:

• The proposed meta-learning approach offers a reliable method for continuous respiratory flow monitoring.
• Short calibration enables accurate, personalized respiratory prediction with minimal sensors.
• This technology supports the development of advanced AI-driven health applications for daily life.