Quaternion-based CNN for heart rate prediction from PPG

Area of Science:

• Biomedical Engineering
• Computer Vision
• Signal Processing

Background:

• Remote photoplethysmography (rPPG) extracts physiological data, like heart rate, from facial videos.
• Existing rPPG methods struggle with real-world conditions like varied illumination and motion artifacts.
• Accurate, non-invasive heart rate monitoring is crucial for healthcare applications.

Purpose of the Study:

• To develop a robust rPPG model for accurate heart rate estimation from facial images.
• To overcome limitations of current signal processing and deep learning rPPG algorithms.
• To enhance the reliability of remote vital sign monitoring systems.

Main Methods:

• Proposed a novel quaternion-valued convolutional neural network (QCNN) for rPPG signal estimation.
• Explored various color formats (RGB, YUV, HSL) to enrich input data within the quaternion domain.
• Conducted experiments under diverse illumination and simulated motion artifact conditions.

Main Results:

• The QCNN model demonstrated consistently high accuracy in heart rate estimation across various environments.
• The approach proved robust against challenging conditions, including diverse lighting and motion artifacts.
• The model effectively utilizes spatial-color information for improved rPPG signal extraction.

Conclusions:

• The proposed QCNN model offers a reliable and efficient method for remote, non-contact heart rate monitoring.
• This advancement supports the development of advanced telemedicine and eHealth systems.
• Enables continuous and non-invasive vital sign monitoring crucial for future healthcare.