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

Photoplethysmography PPG is a non-invasive and effective method for continuously measuring blood pressure BP without the use of a cuff. However, its accuracy can be affected by noise, motion artifacts, and inter-individual variability. We implement a deep learning framework that utilizes the PPG signal to accurately reconstruct arterial blood pressure ABP waves, enabling reliable estimation of systolic and diastolic blood pressure SBP/DBP values from the reconstructed ABP waves. To capture temporal and morphological cardiovascular dynamics, the TCN-BiLSTM model combines temporal convolutional networks and bidirectional long short-term memory networks within a framework of derivative-enhanced PPG representations. The Mean Absolute Error and Root Mean Square Error MAE/RMSE values of the proposed method were 3.61/5.24 mmHg for SBP and 1.79/2.37 mmHg for DBP when using 2226 patients in VitalDB database. The completed model was tested on an independent the MIMIC-IV database cohort as an additional external assessment, without retraining. These findings are consistent with those of previous studies, suggesting that using PPG to reconstruct the ABP waveform provides an accurate and physiologically valid method for estimating arterial pressure. This study is feasible for continuous blood pressure monitoring without the use of a cuff.