Continuous Cuffless Blood Pressure Estimation Based on Fractional Order Derivatives via Gramian Angular Field Only Using Photoplethysmograms

Since instantaneous large changes in blood pressure (BP) values would cause the stroke or even death, continuous BP estimation is essential and crucial. Nevertheless, traditional cuffed BP estimation devices are unable to perform continuous BP estimation. Therefore, there has been a growing interest in developing continuous cuffless BP estimation devices. In order to reduce hardware costs, photoplethysmograms (PPGs) are acquired and their integer order derivative signals are computed to extract features related to BP. Then, conventional machine learning models are developed to estimate BP values. However, the nonlinear characteristics of the heart and blood vessels introduce fractional delays to blood flow. Hence, the traditional integer order derivatives of PPGs may not yield high accuracy. To address this issue, this paper proposes a cuffless BP estimation method based on fractional order derivatives (FODs) of PPGs. First, singular spectrum analysis (SSA) is employed to preprocess the PPGs. Then, the fractional order derivatives of the preprocessed PPGs are calculated. Second, a multi-channel Gramian angular field (GAF)-based image encoding method is applied to both the integer order and fractional order derivatives of the PPGs to generate two-dimensional (2D) images. Then, the encoded images from each individual channel are combined to form a multi-channel encoded image. Third, a residual neural network with 18 layers (ResNet-18) and a U-architecture convolutional network (U-Net) are respectively used for BP estimation. To evaluate the effectiveness of our proposed method, computer numerical simulations are conducted using the Queensland dataset. The results show that our proposed method yields the lower errors and higher correlation coefficients compared to existing methods. Furthermore, our proposed method outperforms both the single-channel and three-channel image encoding methods in terms of errors and correlation coefficients.