Multi-Task Deep Learning Approach for Contactless Simultaneous Heart Rate and Oxygen Saturation Estimation.

Area of Science:

• Biomedical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Pulse oximeters are standard for monitoring heart rate (HR) and peripheral oxygen saturation (SpO2).
• Non-contact, camera-based vital signs monitoring is gaining traction for improved hygiene and comfort.

Purpose of the Study:

• To introduce and validate a novel multi-task neural network for simultaneous, non-contact estimation of HR and SpO2 from multispectral video.
• To leverage interrelationships between HR and SpO2 using a specialized network architecture.

Main Methods:

• Preprocessing multispectral video across spatial, temporal, and frequency domains.
• Utilizing a multi-task neural network with shared feature extractors, task-specific branches, channel attention, and task weight balancing.
• Validation using a 3D multispectral camera system on 23 healthy participants in a breath-holding study.

Main Results:

• Mean absolute errors of 2.97 bpm for HR and 2.34% for SpO2 were achieved in a leave-one-participant-out validation.
• Pearson correlation coefficients were 0.54 for HR and 0.68 for SpO2 when compared to commercial pulse oximeters.
• Demonstrated good precision and agreement with existing pulse oximetry methods.

Conclusions:

• The proposed approach shows potential as a hygienic and comfortable alternative for simultaneous HR and SpO2 measurement.
• This camera-based vital signs monitoring method offers a promising non-contact solution.
• Further development could enhance non-contact health monitoring capabilities.