Arterial blood pressure waveform reconstruction estimation from PPG using TCN–BiLSTM

Figures

• 

  Fig. 1. Overview of the proposed waveform-based cuffless blood pressure estimation framework. PPG signals from the VitalDB dataset undergo preprocessing, signal quality assessment, and fixed-window segmentation. A hybrid TCN–BiLSTM model reconstructs the ABP waveform, from which SBP and DBP are derived.

• 

  Fig. 2. Example of raw and filtered PPG signals, and raw and filtered ABP signals, after physiological quality control and signal filtering.

• 

  Fig. 3. Example of a segment showing the PPG signal and its first-order and second-order derivatives (PPG, ΔPPG, Δ2PPG).

• 

  Fig. 4. Example of ABP waveform analysis after Normalization and Segmentation, showing detected systolic peaks, diastolic valleys, and the median SBP and DBP values within a valid window

• 

  Fig. 5. Overview of the proposed Temporal Convolutional Network TCN module.

• 

  Fig. 6. Illustration of one-dimensional dilated causal convolution with progressively increasing dilation factors.

• 

  Fig. 7. The overall framework of the proposed TCN–BiLSTM architecture for the waveform-to-waveform reconstruction of arterial blood pressure ABP from photoplethysmography PPG is provided. The input signal is of three channels for an eight-second window of 1000 samples, and the output is a reconstructed ABP signal of the same window.

• 

  Fig. 8. Representative example of true and predicted arterial blood pressure.

• 

  Fig. 9. Scatter plots and Bland–Altman analyses of predicted versus reference systolic blood pressure (SBP) and diastolic blood pressure (DBP) on the test set.

• 

  Fig. 10. Histogram of SBP prediction errors (Predicted − Reference), centered around zero, indicating low systematic bias and balanced estimation performance.

• 

  Fig. 11. Histogram of DBP prediction errors (Predicted − Reference), showing a narrow distribution and high robustness of DBP estimation.

• 

  Fig. 12. Error distribution of blood pressure estimation results for different models.