cbPPGGAN: A Generic Enhancement Framework for Unpaired Pulse Waveforms in Camera-Based Photoplethysmography

Area of Science:

• Biomedical Engineering
• Signal Processing
• Computer Vision

Background:

• Camera-based photoplethysmography (cbP PG) non-invasively measures blood volume changes using facial videos.
• Traditional methods struggle with motion artifacts and illumination variations, impacting heart rate (HR) and heart rate variability (HRV) analysis.
• Deep learning models degrade under illumination changes, despite reconstructing high-quality pulse waveforms.

Purpose of the Study:

• To develop a robust cbP PG framework (cbPPGGAN) combining strengths of traditional and deep learning approaches.
• To enhance waveform fidelity and generalization capabilities for accurate HR and HRV analysis.
• To address limitations of paired training data using a novel cycle consistency loss.

Main Methods:

• Proposed cbPPGGAN framework for cbP PG signal enhancement.
• Integration of both unpaired and paired data sources during training.
• Implementation of a cycle consistency loss for time-frequency consistency.
• Utilized traditional approach waveforms as input for reconstruction.

Main Results:

• cbPPGGAN significantly enhances traditional cbP PG waveform quality across different illuminations and datasets.
• Achieved Mean Absolute Error (MAE) of 1.34 bpm (BH-rPPG) and 1.65 bpm (UBFC-rPPG).
• Demonstrated Average Beat-to-Beat (AVBB) errors of 27.46 ms (BH-rPPG) and 45.28 ms (UBFC-rPPG).
• Outperformed state-of-the-art methods in HR estimation and HRV analysis.

Conclusions:

• The cbPPGGAN framework effectively improves cbP PG signal quality and accuracy.
• The proposed method enables reliable HR estimation and HRV analysis in unconstrained environments.
• cbPPGGAN offers a promising advancement for non-contact physiological monitoring.