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

Cuffless blood pressure (BP) classification using photoplethysmography (PPG) has been a hot topic for the prevention of cardiovascular system diseases. In this paper, we propose a novel encoding strategy to transcend the limitations of traditional temporal feature extraction. By leveraging Gramian Angular Field (GAF) coding, the original one-dimensional PPG signals are transformed into two-dimensional matrices, enabling the model to capture multi-scale spatial-temporal correlations through advanced computer vision architectures. The triple classification of BP is achieved on the deep neural network GHC-Net based on GAF. The network is designed with a suitable hybrid dilation convolution (HDC) to increase the receptive field. Meanwhile, the Remaining Effective Channel Attention (RECA) module is able to capture cross-channel dependencies and realize interactions between features. In the public dataset PPG-BP, the classification accuracy for normotensive (NT), prehypertensive (PHT), and hypertensive (HT) was 76.39% and the F1 score was 76.68%. In the private dataset, the classification accuracy of the three classifications was 92.66% and the F1 score was 93.33%. The findings reveal that we provide an efficient and accurate automated BP classification method with potential home care applications in the early detection and monitoring of hypertensive disorders.