Analysis of CNN-based remote-PPG to understand limitations and sensitivities

Area of Science:

• Biomedical Engineering
• Computer Vision
• Signal Processing

Background:

• Deep learning, particularly convolutional neural networks (CNNs), excels in vision tasks and is emerging in camera-based vital signs monitoring.
• Current CNN applications for photoplethysmography (PPG) extraction prioritize performance over methodological understanding.

Purpose of the Study:

• To enhance understanding of CNN-based PPG extraction methodology.
• To investigate factors influencing CNN performance in vital signs monitoring.
• To explore potential improvements for motion-robustness and signal accuracy.

Main Methods:

• Experimental analysis of CNNs for PPG signal extraction from visual data.
• Investigation into the role of blood absorption variations in CNN signal detection.
• Evaluation of reference-signal parameters and multi-site measurement strategies.
• Assessment of the utility of prior physiological knowledge in CNN models.

Main Results:

• CNNs leverage blood absorption variations for physiological signal extraction.
• Reference-signal characteristics (phase, spectral content) significantly impact PPG extraction.
• Multiple convolutional kernels enhance flexibility but may not match multi-site knowledge-based methods for motion robustness.
• Incorporating PPG-related prior knowledge can benefit CNN-based extraction.

Conclusions:

• CNN-based PPG extraction relies on understanding blood absorption dynamics.
• Optimizing reference signals and integrating prior knowledge are crucial for robust vital signs monitoring.
• Hybrid CNN approaches combining deep learning with physiological principles warrant further research.