A PPG-Based Calibration-Free Cuffless Blood Pressure Estimation Method Using Cardiovascular Dynamics

Figures

• 

  Figure 1. Flowchart of the proposed method in this paper for cuffless blood pressure estimation, which shows the process of BP estimation for design, evaluation and testing. During the design phase, the ANN model is based on leave-one-out (LOO) method, where it is evaluated on data from a single patient at a time and repeated over all patients. The same structure is used during the final testing phase where the ANN is trained on the complete dataset, and then, tested on a new set of data from 25 patients that have never been seen by the structure. In this process the choice of whether to use extracted features from PPG morphology depends on whether the analysis is based on morphology or interbeat interval (IBI) features (as explained in the text).

• 

  Figure 2. Detection of peaks, troughs and corresponding IBIs in the PPG signal. (a) Detected peaks are marked with vertical red lines and sample IBIs, used for SBP estimation, are marked with horizontal red lines; (b) Detected troughs are marked with vertical blue lines and sample IBIs, used for DBP estimation, are marked with horizontal red lines.

• 

  Figure 3. Extracted parameters from the PPG signal to generate the 21 features used in this study.

• 

  Figure 4. Bland–Altman plots presenting systolic and diastolic blood pressure estimation using three different methods (IBI features from PPG, PPG morphology features and calibrated mathematical model). The left column shows the results for systolic BP and the right column for diastolic BP. Each of the plots is named corresponding to the method that is used for estimation. In all the plots, the black horizontal line is the bias and the two dashed red lines are the 95% limits of agreement.

• 

  Figure 5. Bland–Altman plots presenting systolic and diastolic blood pressure estimation using fusion on three different methods (IBI features from PPG, PPG morphology features and calibrated mathematical model). The left column (a–c) shows the results for systolic BP, and the right column (d–f) for diastolic BP. Each of the plots is named accordingly. In all the plots the black horizontal line is the bias and the two dashed red lines are the 95% limits of agreement.

• 

  Figure 6. Bland–Altman plots presenting systolic and diastolic blood pressure estimation using three different sets of features from PPG waveforms (IBI and morphology, IBI and calibrated mathematical model and IBI, morphology and calibrated mathematical model). The left column (a–c) shows the results for systolic BP and the right column (d–f) for diastolic BP. Each of the plots is named accordingly. In all the plots, the black horizontal line is the bias and the two dashed red lines are the 95% limits of agreement.

• 

  Figure 7. Scatter diagrams for blood pressure estimation using three different methods. The top row has the results for systolic BP and bottom row shows the results from diastolic BP. For both rows, the first plot from the left shows correlation between BP estimation using PPG dynamics and PPG morphology, the second plot displays BP estimation correlation between PPG dynamics and the calibration model and the third plot illustrates blood pressure estimation correlation using morphology and the calibration model. All the results are based on collected PPG waveforms from 30 patients. All estimates in this figure are in mmHg.

• 

  Figure 8. Bland–Altman plots presenting systolic blood pressure estimation using different methods: (a) Estimation using PPG dynamics features and ANN model; (b) Estimation using PPG morphology features and ANN model; (c) Estimation using PPG morphology features and feedforward deep NN model; (d) Estimation using PPG morphology features and LSTM model; (e) Estimation using PPG morphology features and GRU model; (f) Estimation combining the estimated BP result from (b) and PPG dynamics and use them in ANN model; (g) Estimation combining the estimated BP result from (c) and PPG dynamics and use them in ANN model; (h) Estimation combining the estimated BP result from (d) and PPG dynamics and use them in ANN model; (i) Estimation combining the estimated BP result from (e) and PPG dynamics and use them in ANN model. In all the plots, the black horizontal line is the bias and the two dashed red lines are the 95% limits of agreement.

• 

  Figure 9. Bland–Altman plots presenting diastolic blood pressure estimation using different methods: (a) Estimation using PPG dynamics features and ANN model; (b) Estimation using PPG morphology features and ANN model; (c) Estimation using PPG morphology features and feedforward deep NN model; (d) Estimation using PPG morphology features and LSTM model; (e) Estimation using PPG morphology features and GRU model; (f) Estimation combining the estimated BP result from (b) and PPG dynamics and use them in ANN model; (g) Estimation combining the estimated BP result from (c) and PPG dynamics and use them in ANN model; (h) Estimation combining the estimated BP result from (d) and PPG dynamics and use them in ANN model; (i) Estimation combining the estimated BP result from (e) and PPG dynamics and use them in ANN model. In all the plots, the black horizontal line is the bias and the two dashed red lines are the 95% limits of agreement.

• 

  Figure 10. Structure of the proposed neural network model. This network consists of three subnetworks. The one on the left has seven inputs that are PPG morphology features, three hidden layers that use sigmoid activation functions and an output layer that uses a linear activation function. The outputs of this subnetwork are estimated SBP and DBP that are then used in the subsequent subnetworks to estimate the final BP values. Each of these two subnetworks (on the right) has eight inputs (seven features related to cardiovascular dynamics plus the estimated BP value from morphology features from the prior subnetwork). Furthermore, each of these two networks has a single hidden layer with a sigmoid activation function and an output layer with a linear activation function. The resulting SBP and DBP from this stage are the final BP estimation values.

• 

  Figure 11. Bland–Altman plots presenting systolic and diastolic blood pressure estimations using the proposed model on the test data from 25 never-seen-before patients. In both plots, the black horizontal line is the bias and the two dashed red lines are the 95% limits of agreement.