PulseGAN: Learning to Generate Realistic Pulse Waveforms in Remote Photoplethysmography

Area of Science:

• Biomedical Engineering
• Signal Processing
• Computer Vision

Background:

• Remote photoplethysmography (rPPG) non-invasively measures cardiac signals from facial videos, crucial for health monitoring and emotion recognition.
• Existing rPPG methods often yield inaccurate pulse signals, limiting their use to average heart rate (HR) estimation.
• High-quality rPPG signals are needed to derive precise physiological metrics.

Purpose of the Study:

• To introduce PulseGAN, a novel generative adversarial network framework for generating realistic rPPG pulse signals.
• To improve the accuracy of rPPG-derived features, including HR, interbeat interval (IBI), and heart rate variability (HRV).
• To enhance the quality of rPPG waveforms by denoising chrominance (CHROM) signals.

Main Methods:

• Utilized a generative adversarial network (GAN) architecture, PulseGAN, for rPPG signal generation.
• Employed denoising of chrominance (CHROM) signals from facial videos as input.
• Incorporated error losses in both time and spectrum domains, alongside adversarial loss, to ensure accurate pulse waveform generation.

Main Results:

• PulseGAN significantly improved rPPG waveform quality compared to input CHROM signals.
• Demonstrated substantial reductions in mean absolute error for AVNN (average of normal-to-normal intervals) by 40.45-41.63% and SDNN (standard deviation of NN intervals) by 37.53-58.41% across three public databases.
• Validated the framework's effectiveness in enhancing the accuracy of HR, IBI, and HRV features.

Conclusions:

• PulseGAN effectively generates high-fidelity rPPG pulse signals, overcoming limitations of existing methods.
• The framework enhances the accuracy of critical cardiac metrics derived from rPPG.
• PulseGAN offers a versatile solution that can be integrated with other rPPG techniques to improve reliability and expand applications.