SiamQuality: a ConvNet-based foundation model for photoplethysmography signals

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Signal Processing

Background:

• Low-quality physiological data frequently hinders effective health monitoring.
• Developing robust models to manage data variability is crucial for reliable healthcare technologies.

Purpose of the Study:

• To create a foundation model capable of handling low-quality physiological data.
• To enhance the accuracy and reliability of health monitoring systems.

Main Methods:

• Introduced SiamQuality, a self-supervised learning approach utilizing convolutional neural networks (CNNs).
• Trained the model on a large dataset of photoplethysmography (PPG) signals from intensive care patients.
• Employed a Siamese network architecture to learn similar representations for high and low-quality PPG signals.

Main Results:

• Fine-tuned SiamQuality on six cardiovascular monitoring tasks, achieving significant performance improvements.
• Exceeded state-of-the-art performance in respiratory rate estimation (by 75%) and atrial fibrillation detection (by 5%).
• Demonstrated consistent effectiveness across all evaluated tasks, particularly for heart monitoring applications.

Conclusions:

• CNNs provide a robust backbone for foundation models in physiological data analysis.
• SiamQuality effectively handles real-world, variable-quality data, paving the way for advanced healthcare monitoring.
• The model's success highlights potential for more reliable and efficient wearable health technologies.