Blood Pressure Measurement: From Cuff-Based to Contactless Monitoring

Figures

• 

  Figure 14. A pipeline of a deep learning model for rPPG singals proposed in [147,229,230]. The pixels of the ROIs of face images are preprocessed and then extracted to form rPPG signals. These rPPG signals are then preprocessed, and those features of preprocessed rPPG signals are extracted to feed the respective models for training to predict SBP and DBP signals.

• 

  Figure 1. A timeline of PPG-related publications every year from 1990 to 2022. For 2022, the number of publications was recorded up to 26 July 2022.

• 

  Figure 2. This figure shows waveforms from various measurements and the relation between PAT and PTT as in Equation (1). ECG = electrocardiography, SCG = seismocardiography, and PPG = photoplethysmography. The letters in the ECG waveforms represent the parts of the waveforms.

• 

  Figure 3. Biological signals from various parts of the body. This figure was modified and enhanced from [52]. The abbreviations represent BP measurement signals at various parts on the human body. BCG = ballistocardiography, PPG = photoplethysmography, IPG = impedance photoplethysmography, SBS = strain-based sensor, GCG = gyrocardiography, ECG = electrocardiography, SCG = seismocardiography, ICG = impedance cardiography, ABP = arterial blood pressure (by invasive measurement on the arms), and rPPG = remote photoplethysmography (by cameras).

• 

  Figure 4. A working principle of rPPG. This figure was modified from [53]. The specular component gives the information on the skin’s surface, which does not have any physiological signals. The diffused counterpart gives the subtle change in blood flow, which provides physiological information after meticulous signal processing.

• 

  Figure 5. A timeline of rPPG-related publications every year from 1990 to 2022. For 2022, the number of publications was recorded up to 26 July 2022.

• 

  Figure 6. A flowchart showing the research situation in BP measurement. There are 5, 11 and 30 pieces of work using non-ML, ML and DL methods respectively to predict BP from PPG signals, while there are 4, 8 and 7 pieces of work using non-ML, ML and DL methods respectively to predict BP from rPPG signals. ROI preprocessing means preprocessing of data received from some regions of interest on a human face recorded on video. The red dotted line means that there are no research papers talking about the techniques from a video for BP measurement.

• 

  Figure 7. A flowchart showing the organization of this review paper.

• 

  Figure 8. A pipeline of traditional machine learning (TML) methods for PPG signals. Features inside training data are extracted based on various author-preferred criteria (e.g., highest correlation with the ground truth BP) to train their studied TML models and give the predictions for SBP and DBP.

• 

  Figure 9. The flowchart of general PPG waveform-based methods. (upper flowchart) [123,125,127,128,136]. Features inside training data of raw PPG signals are manually extracted to train their studied models and give the predictions for SBP and DBP. (lower flowchart) [133,135,137,145]. The whole waveform profiles of raw PPG signals are put into models, which automatically pick up useful features and give predictions for SBP and DBP.

• 

  Figure 10. The structure of the algorithm in [131]. PPG signals are preprocessed by continuous wavelet transform (CWT) to produce two dimensional scalograms for identifying the low-frequency or fast-changing frequency components. These scalograms are converted into RGB images for the convolutional neural network pretrained by GoogLeNet. This model automatically extracts features in the scalograms to predict the BP categories: normotension (NT), prehypertension (PHT), and hypertension (HT).

• 

  Figure 11. The structure of the algorithm in [141]. PPG signals are preprocessed by the wavelet transform with the removal of the very low and very high frequency components, the soft Rigrsure thresholding [198,199], and the normalization. The filtered PPG signals are processed through the approximation network (one-dimensional deep supervised U-net model), which estimates the ABP based on the inputs, and then refined by the refinement network. The maximum and minimum of the refined ABP signals are taken as SBP and DBP, respectively.