Robust Federated Video-based Remote Physiological Measurement for Heterogeneous Multi-source Data.

Area of Science:

• Biomedical Engineering
• Computer Science
• Signal Processing

Background:

• Remote photoplethysmography (rPPG) enables non-contact physiological monitoring using facial videos.
• Current rPPG methods face challenges with data transmission costs and privacy concerns.
• Federated learning (FL) offers a solution but struggles with cross-domain heterogeneity in multi-source rPPG data.

Purpose of the Study:

• To address the challenges of privacy and cross-domain heterogeneity in federated training for remote photoplethysmography.
• To propose a novel federated learning framework, FedGRC, for effective multi-source rPPG data analysis.
• To characterize the heterogeneity of multi-source rPPG data from both input and output domain perspectives.

Main Methods:

• Characterized multi-source rPPG data heterogeneity from input and output domains.
• Introduced pseudo-labeling techniques within a federated learning framework.
• Developed FedGRC with automatic gradient regularization calibration and pseudo-label alignment using handcrafted rPPG methods.

Main Results:

• FedGRC effectively mitigates input domain discrepancies through gradient regularization calibration.
• Pseudo-labeling aligns output domains and reduces label inconsistencies across datasets.
• Validated across six public datasets, FedGRC demonstrated significant advantages over existing approaches.

Conclusions:

• The FedGRC framework successfully addresses privacy protection and heterogeneity challenges in multi-source rPPG data.
• This approach enhances the feasibility of federated learning for non-contact physiological measurements.
• FedGRC offers a robust solution for privacy-preserving and heterogeneous rPPG data analysis.