Few-shot pulse wave contour classification based on multi-scale feature extraction.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Pulse wave contour (PWC) annotation is expensive and time-consuming, hindering large-scale dataset creation for deep learning.
• Few-shot learning approaches are needed to overcome data limitations in PWC analysis.

Purpose of the Study:

• To develop a novel, small-parameter unit structure and a multi-scale feature-extraction model for few-shot PWC analysis.
• To improve the efficiency and accuracy of cardiovascular disease classification using limited PWC data.

Main Methods:

• A small-parameter unit structure utilizing state variables and a forgetting gate to capture long-term PWC dependencies.
• A multi-scale feature-extraction model integrating Convolutional Neural Networks (CNNs) for spatial and rhythm features, and Recursive Neural Networks (RNNs) for temporal dependencies.
• An inference layer for final classification based on extracted multi-scale features.

Main Results:

• The proposed multi-scale feature-extraction model achieved 80% classification accuracy on a photoplethysmography dataset.
• The model reached 96% classification accuracy on a continuous non-invasive blood pressure dataset.
• The approach demonstrates effectiveness in few-shot learning scenarios for PWC analysis.

Conclusions:

• The developed small-parameter unit structure and multi-scale feature-extraction model offer an efficient solution for PWC analysis with limited data.
• This method significantly enhances the classification of cardiovascular diseases, showing high accuracy on diverse datasets.
• The proposed model addresses the data scarcity challenge in deep learning for PWC-based medical diagnostics.