Blood Pressure Morphology Assessment from Photoplethysmogram and Demographic Information Using Deep Learning with Attention Mechanism

Figures

• 

  Figure 1. Block diagram of the proposed methodology.

• 

  Figure 2. (a) Summarized processing stage. (b) In blue, two 15-s segments analyzed with 5 min gap between each other, in black. (c) An example of ABPM¯ computed. Red, green, and blue lines correspond to the classes [onset—systolic peak], [systolic peak—dicrotic notch] and [dicrotic notch–end], respectively, while gray lines represent ABP pulses and dashed magenta lines represent the limits to consider the average pulse.

• 

  Figure 3. Selected dataset distributions: (a) Number of subjects with their corresponding quantity of segments (gray bars) and segments cumulative percentage (red line). (b) Age and gender. (c) Systolic blood pressure (SBP) and diastolic blood pressure (DBP).

• 

  Figure 4. (a) recurrent NN (RNN) Encoder-Decoder architecture. (b) Structure of gated recurrent units (GRU) unit.

• 

  Figure 5. Fixed-size target signal composed by a completed ABPM¯ and its repetition. Red, green, blue, and black lines correspond to classes [onset - systolic peak], [systolic peak—dicrotic notch], [dicrotic notch—end] and [Ended], respectively. Vertical magenta dotted line represents the end of the error’s mask.

• 

  Figure 6. Model architecture.

• 

  Figure 7. Relationship between photoplethysmogram signals (PPG) and PPG input signals and ABPM^ via attention weight’s heatmap. Additionally, ABPM¯v is shown in comparison with ABPM^.

• 

  Figure 8. Comparison between different ABPM¯v and ABPM^ examples.

• 

  Figure 9. Regression plots (a,d), Bland–Altman plots (b,e), and histograms of errors (c,f), corresponding to Mixno+DI scenario and DBP and SBP values.