A Multifunctional Network with Uncertainty Estimation and Attention-Based Knowledge Distillation to Address Practical...

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Respiration rate is a key health indicator, but traditional measurement methods are limited during physical activity.
• Wearable devices estimate respiration rate using physiological signals like ECG and photoplethysmography, with deep learning showing promise for improved accuracy.
• Existing deep learning models face challenges in interpretability, uncertainty estimation, and compactness for wearable deployment.

Purpose of the Study:

• To develop a multifunctional deep learning framework for robust respiration rate estimation.
• To address challenges of interpretability, uncertainty estimation, and model compactness in wearable respiration monitoring.
• To enhance the accuracy and practical applicability of respiration rate estimation during ambulatory movements.

Main Methods:

• Proposed a framework integrating an attention mechanism, uncertainty estimation (Monte Carlo dropouts), and knowledge distillation.
• Evaluated the framework on two datasets featuring ambulatory movements.
• Utilized attention mechanisms for improved instantaneous respiration rate estimation and Monte Carlo dropouts for uncertainty quantification.

Main Results:

• The attention mechanism improved instantaneous respiration rate estimation accuracy.
• Uncertainty estimation via Monte Carlo dropouts allowed rejection of high-uncertainty data, reducing mean absolute error by 7.99%.
• Attention-aware knowledge distillation reduced model parameters by 49.5% and inference time by 38.09% without compromising accuracy.

Conclusions:

• The proposed framework effectively enhances respiration rate estimation accuracy and robustness during movement.
• The integration of attention, uncertainty estimation, and knowledge distillation addresses key challenges for wearable deployment.
• This work represents a significant step towards deploying advanced respiration monitoring in edge wearable devices.