Assessment of Blood Pressure by Photoplethysmogram Signal Based on the Combined Configuration of an Artificial Neural Network

The paper presents a technique for creating an individualized model for predicting human blood pressure from a photoplethysmogram (PPG) signal. The signals used had a high level of noise, since their registration was performed using a wearable device in adverse conditions. Therefore, much attention was paid to cleaning the signal from interference. To obtain blood pressure estimates, a combination of two machine learning algorithms with different approaches to data analysis was used, in particular a one-dimensional convolutional neural network and a fully connected direct propagation network. Fragments of the photoplethysmogram signal were fed to the input of the convolutional network, and a set of features calculated according to the cycles of the PPG were fed to the input of the direct propagation network. It is shown that the combination of two alternative configurations of neural networks allows you to get more accurate estimates of blood pressure than each of the networks separately.