Application of Neural Networks in the Problem of Indirect Estimation of Blood Pressure According to the Heart Rate Variability Features

The article presented the results of testing the hypothesis that it was possible to predict blood pressure using the significant features of heart rate variability signals. The hypothesis was based on previous results related to the task of diagnosing arterial hypertension using heart rate variability data. One of the PhysioNet's databases was used as dataset, which contains simultaneously recorded biomedical signals of electrocardiogram and blood pressure. In this work, fully connected neural networks were considered as a method for regression model evaluation. Comparison of different variations of neural network architectures, including varying number of hidden layers and number of neurons in the hidden layer, was carried out and their results were compared. A comparison was made with the previously obtained results based on the application of evolutionary programming for significant feature selection. Various possibilities for improving the results obtained are discussed, including the application of other signals, convolutional neural networks, the inclusion of additional parameters such as gender, age, anthropomorphic characteristics of the subjects.