GHC-net: A gramian angular field based hybrid CNN for cuffless blood pressure classification using PPG signals.

GHC-Net achieved normotensive, prehypertensive, and hypertensive triple classification accuracy of 76.39% with an F1 score of 76.68% on the public PPG-BP dataset.

• Classification was performed on three categories: normotensive (NT), prehypertensive (PHT), and hypertensive (HT).
• The public dataset used was PPG-BP.
• Overall accuracy was 76.39% and F1 score was 76.68% on this dataset.

GHC-Net achieved a triple-classification accuracy of 92.66% with an F1 score of 93.33% on a private dataset.

• The private dataset results substantially outperformed the public PPG-BP dataset results.
• Classification accuracy was 92.66% and F1 score was 93.33% for NT, PHT, and HT categories.
• The private dataset was distinct from the publicly available PPG-BP dataset.

Gramian Angular Field (GAF) encoding was used to transform one-dimensional PPG signals into two-dimensional matrices for spatial-temporal feature extraction.

• GAF coding converts original 1D PPG signals into 2D matrix representations.
• This transformation enables the model to capture multi-scale spatial-temporal correlations.
• The encoding strategy was designed to transcend the limitations of traditional temporal feature extraction.
• The 2D representations allow use of advanced computer vision architectures.

The GHC-Net architecture incorporates a Hybrid Dilation Convolution (HDC) module to increase the receptive field.

• HDC was specifically designed to expand the receptive field of the convolutional network.
• The hybrid dilation convolution is a key architectural component of GHC-Net.
• This design choice was intended to improve multi-scale feature extraction from the 2D GAF-encoded inputs.

The GHC-Net architecture includes a Remaining Effective Channel Attention (RECA) module to capture cross-channel dependencies and feature interactions.

• RECA captures cross-channel dependencies within the network.
• The module enables interactions between features across channels.
• RECA is integrated alongside the HDC module as a core component of GHC-Net.