Estimating Cuff-less Continuous Blood Pressure from Fingertip Photoplethysmogram Signals with Deep Neural Network Model

Abstract

AbstractObjectiveBlood pressure (BP) is an important physiological index reflecting cardiovascular function. Continuous blood pressure monitoring helps to reduce the prevalence and mortality of cardiovascular diseases. In this study, we aim to estimate systolic blood pressure (SBP) and diastolic blood pressure (DBP) values continuously based on fingertip photoplethysmogram (PPG) waveforms using deep neural network models.MethodsTwo models were proposed and both models consisted of three stages. The only difference between them was the method of extracting features from PPG signals in the first stage. Model 1 adopted Bidirectional Long Short-Term Memory (BiLSTM), while the other used convolutional neural network. Then, the residual connection was applied to multiple stacked LSTM layers in the second stage, following by the third stage with two fully connected layers.ResultsOur proposed models outperformed other methods based on similar dataset or framework, while in our proposed models, the model 2 was superior to model 1. It satisfied the standard of Association for the Advancement of the Medical Instrumentation (AAMI) and obtained grade A for SBP and DBP estimation according to the British Hypertension Society (BHS) standard. The mean error (ME) and standard deviation (STD) for SBP and DBP estimations were 0.21 ± 6.40 mmHg and 0.19 ±4.71 mmHg, respectively.ConclusionOur proposed models could extract important features of fingertip PPG waveforms automatically and realize cuff-less continuous BP monitoring, which can be helpful in the identification and early treatment of abnormal blood pressure, thus may reduce the occurrence of cardiovascular malignant events.