Multi-task learning for estimation of remote PPG and respiration signals with complex valued convolutional neural...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Remote and continuous monitoring of physiological signals is vital for early disease detection.
• Traditional contact sensors face challenges like virus transmission risk and patient discomfort.
• Deep learning models for enhanced diagnostics often require large architectures, unsuitable for embedded systems.

Purpose of the Study:

• To develop a multitask learning model for simultaneous remote photoplethysmogram (PPG) and respiratory rate estimation from facial videos.
• To create a compact model architecture suitable for embedded edge devices.

Main Methods:

• Utilized RGB facial video streams and constructed a complex-numbered dataset.
• Developed a complex-valued multitask learning neural network model.
• Trained and evaluated the model on a public dataset of face video streams.

Main Results:

• The complex-valued multitask learning model achieved simultaneous estimation of remote PPG and respiratory rate.
• The proposed model demonstrated higher performance compared to conventional real-valued networks.
• The model exhibited a smaller network structure, beneficial for embedded devices.

Conclusions:

• The developed model shows significant potential for accurate and efficient remote physiological disorder monitoring.
• Complex-valued neural networks offer advantages in reducing model size for edge computing applications.
• This approach facilitates non-invasive, continuous health monitoring via facial video analysis.