Cuff-less and Calibration-free Blood Pressure Estimation Using Convolutional Autoencoder with Unsupervised Feature Extraction

Monitoring blood pressure (BP) in people's daily life in an unobtrusive way is of great significance to prevent cardiovascular disease and its complications. However, most of the current cuff-less BP estimation methods still suffer from two drawbacks including calibration and tedious feature selection. In this study, we first attempt to validate the feasibility of convolutional autoencoder (CAE) to estimate continuous BP without calibration and hand-crafted feature extraction. 62 subjects were recruited in this experiment. We first trained the CAE on all the data to extract the unsupervised features. Then, we trained a regressor to estimate BP values using the features learning from the CAE. 10-fold cross-validation tests were used to examine the performance of our models. Finally, it has been demonstrated that the accuracy of the predicted BP satisfied the Grade B standard of British Hypertension Society. Due to its calibration-free and unsupervised feature learning ability from the collected signal, the proposed method has high prospects for application in wearable BP monitoring device.