Sparsity-constrained one-dimensional convolutional autoencoder for reliable photoplethysmography template quality...

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Photoplethysmography (PPG) signal quality is critical for accurate physiological monitoring.
• Real-world PPG data often suffers from motion artifacts and pathological variations.
• Existing methods may struggle with diverse signal corruptions, especially in arrhythmic cases.

Purpose of the Study:

• To develop an automated, lightweight method for assessing PPG signal quality.
• To improve the reliability of PPG-based physiological analysis in challenging conditions.
• To create a model suitable for resource-constrained edge devices.

Main Methods:

• A lightweight 1D convolutional autoencoder was employed for signal quality classification.
• Template-based segmentation and spline interpolation were used for baseline correction.
• Classification was based on template-level reconstruction error for automatic quality assessment.

Main Results:

• The model achieved high performance: 96% accuracy, 99% precision, 96% recall, and 98% F1-score.
• It effectively preserved valid PPG signals while rejecting corrupted segments, outperforming a skewness-based index, especially with arrhythmias.
• The compact model (1.3 MB) is suitable for edge device deployment.

Conclusions:

• The proposed method offers an efficient and clinically relevant solution for PPG signal quality assessment.
• Its robustness in handling artifacts and arrhythmias makes it valuable for real-world applications.
• The lightweight design facilitates integration into wearable and edge computing systems.