HGCTNet: Handcrafted Feature-Guided CNN and Transformer Network for Wearable Cuffless Blood Pressure Measurement.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning for Healthcare

Background:

• Wearable biosignals like ECG and PPG contain redundant information and complex temporal dependencies.
• Extracting accurate features for cuffless blood pressure (BP) estimation from these signals remains a significant challenge.
• Existing methods struggle to efficiently process both local and global temporal information in biosignals.

Purpose of the Study:

• To develop an advanced deep learning model, HGCTNet, for accurate cuffless BP estimation using wearable device data.
• To leverage a hybrid CNN-Transformer architecture to capture comprehensive features from biosignals.
• To integrate handcrafted features and demographic data to further enhance BP estimation performance.

Main Methods:

• Proposed HGCTNet, a hybrid CNN-Transformer network integrating convolutional operations and self-attention mechanisms.
• Implemented a handcrafted feature-guided attention module to reduce redundancy in learned features using handcrafted signal features.
• Employed a feature fusion module combining learned features, handcrafted features, and demographic data for final BP prediction.

Main Results:

• HGCTNet achieved low estimation errors: 0.9 ± 6.5 mmHg (DBP) and 0.7 ± 8.3 mmHg (SBP) on the CAS-BP dataset.
• On the Aurora-BP dataset, errors were -0.4 ± 7.0 mmHg (DBP) and -0.4 ± 8.6 mmHg (SBP).
• Significantly reduced SBP Mean Absolute Error by 10.68% (CAS-BP) and 9.84% (Aurora-BP) compared to state-of-the-art methods.

Conclusions:

• HGCTNet effectively extracts discriminative features from wearable biosignals for cuffless BP estimation.
• The hybrid architecture and feature fusion strategy enhance accuracy and reduce redundancy in BP monitoring.
• Demonstrated significant improvements over existing methods, highlighting HGCTNet's potential for advanced wearable BP measurement.