Using machine learning models for cuffless blood pressure estimation with ballistocardiogram and impedance plethysmogram

Figures

• 

  Figure 1. In the flowchart of the proposed method, the blue block is the data labels, the green block is the signal quality classification by a stacking 1D CNN + GRU model, red block is the extraction of calibration-based and calibration-free PTT parameters and the orange block is the BP estimation by RF and XGBoost models.

• 

  Figure 2. The synchronous ECG (blue), PPG (red), DPPG (pink), BCG (black), IPG (green), and DIPG (purple) signals. The PTT1BCG + IPG and PTT2BCG + IPG are the interval between the J wave of BCG and the foot point of IPG, and the peak point of DIPG. PAT1ECG + PPG and PAT2ECG + PPG are the interval between the R wave of ECG and the foot point of PPG, and the peak point of DPPG.

• 

  Figure 3. The trends of PAT2ECG + PPG (green), PTT2BCG + IPG (black), SBP (red), and DBP (orange) for one measurement of subject 1.

• 

  Figure 4. The BCG (black), IPG (green), and DIPG (purple). The label of signal quality for each pulse is the red line. When the pulse belongs to the good quality, the label is 1 (higher horizontal line). Otherwise, the label is 0 (lower horizontal line).

• 

  Figure 5. BCG, IPG, and dIPG signals are segmented with 1024 points as a sample, and the overlap is 512 points.

• 

  Figure 6. The structure of the proposed stacking 1D CNN + GRU model for the signal quality classification. Three input channels, BCG, IPG, and DIPG come through a time-distributed layer comprising a stacking 1D CNN + GRU. Then, weights between GRU layer and full connection layer are dropout 50%. The activation function in output layer is sigmoid function.

• 

  Figure 7. The significant scores of mPTT2, PTT2SYS, PTT2DIA, mPTT1, PTT1SYS, PTT1DIA, mHR, and ratio with the permutation feature importance.

• 

  Figure 8. The results of the stacking 1D CNN + GRU model in the training (blue) and validation (orange) phases, (a) the accuracy curves, and (b) the loss curves.

• 

  Figure 9. The fusion matrix in the testing phase. The numbers of TP, TN, FN, FP are 1174, 5534, 7, and 71, respectively.

• 

  Figure 10. The Pearson correlation coefficients of SBP (blue circle) and DBP (green circle) estimations (a) RF model are both 0.94. (b) XGBoost model are 0.95 and 0.94, respectively.

• 

  Figure 11. Bland–Altman plots for SBP and DBP estimations using the RF model are depicted. (a) For SBP estimation, a bias of 0.13 mmHg is observed, with the upper and lower bounds of agreement at 10.35 and −10.10 mmHg, respectively. (b) For DBP estimation, a bias of −0.19 mmHg is noted, with the upper and lower bounds of agreement at 7.29 and −7.67 mmHg, respectively.

• 

  Figure 12. The distributions of absolute difference between reference and BP estimation by RF model. (a) SBP estimated with 83.2%, 93.8%, and 97.4% differences no more than 5 mmHg, 10 mmHg, and 15 mmHg, respectively. (b) DBP estimated with 89.4%, 96.0%, and 98.5% differences less than 5 mmHg, 10 mmHg, and 15 mmHg, respectively.