Domain-independent PPG signal quality assessment framework via representation learning

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Photoplethysmography (PPG) is crucial for non-invasive vital sign monitoring but is prone to artifacts that reduce accuracy.
• Existing PPG signal quality assessment (SQA) methods lack generalizability across diverse sensors and users.
• There is a need for robust, adaptable SQA techniques for reliable PPG data interpretation.

Purpose of the Study:

• To develop a novel, real-time compatible, and sensor-location invariant SQA approach for PPG signals.
• To leverage deep representation learning for improved PPG signal quality assessment.
• To create a flexible SQA framework that generalizes across different devices and user populations.

Main Methods:

• Fine-tuned a pre-trained PPG transformer (PPG-PT) using triplet loss for deep representation learning.
• Quantified signal quality by measuring embedding distances to reference embeddings of high-quality PPG segments.
• Validated the framework on diverse datasets including finger, wrist, and neck PPG signals.

Main Results:

• Achieved an F1-score of approximately 90% on expert-annotated finger PPG data (CSL dataset).
• Demonstrated robust generalization with an F1-score of about 95% on wrist PPG data (WCS dataset).
• Showed moderate performance (≈70% F1-score) on neck PPG data (AA dataset) due to dataset complexity.

Conclusions:

• The proposed deep learning SQA framework offers a generalizable solution for PPG signal quality assessment.
• The reference-based scoring method allows adaptation to new devices without retraining.
• This approach has significant potential for improving the reliability of PPG-based vital sign monitoring in various applications.