Continuous cuffless blood pressure measurement using feed-forward neural network

Abstract

High blood pressure (BP) or hypertension is an extremely common and dangerous condition affecting more than 18–27 % of the world population. It causes many cardiovascular diseases that kill 7.6 million people around the world per year. The most accurate way to detect hypertension is ambulatory monitoring of blood pressure lasting up to 24 h and even more. Traditional non-invasive methods for measuring BP are oscillometric and auscultatory, they use an occlusal cuff as an external pressure source. Unfortunately, cuffed BP measurement creates some inconvenience for the patient and can be inaccurate due to incorrect cuff placement. In connection with the problems caused by cuff methods, it has become necessary to develop cuffless methods for measuring blood pressure, which are based on the relationship of blood pressure with various manifestations of cardiac activity and hemodynamics, which can be recorded and measured non-invasively, without the use of a compression cuff and with simple technical means. Over the past decade, there have been many publications devoted to estimating blood pressure based on pulse wave velocity (PWV) or pulse wave transit time (PTT). However, this approach has few disadvantages. First, the measurement of BP using only PTT parameter is valid only for a given patient. Second, the linear model of the relationship between BP and PTT is valid only in a small range of BP variations. To solve this problem neural networks or linear regression models were used. The main problem with this approach is the accuracy of blood pressure measurement. This study builds one feed-forward neural network (FFNN) for determining systolic and diastolic blood pressure based on features extracted from electrocardiography (ECG) and photoplethysmography (PPG) signals without a cuff and calibration procedure. The novelty of this work is the discovery of five new key points of the PPG signal, as well as the calculation of nine new features of the ECG and PPG signals, which improve the accuracy of blood pressure measurement. The object of the study was the ECG and PPG signals recorded from the patient's hand. The target of the study was to obtain systolic and diastolic blood pressure based on an FFNN, the input arguments of which are the parameters of the ECG and PPG signals. Algorithms for estimating signal parameters based on the determination of characteristic points in the PPG signal, the position of R-peaks in the ECG signal, and parameters calculated from the relationship of time parameters and amplitude ratios of these signals are described in detail. The Pearson correlation coefficients for these parameters and BP are determined, which helps to choose the set of signal parameters valuable for BP estimation. The results obtained show that the mean absolute error ± standard deviation for systolic and diastolic BP is equal to 1.72±3.008 mmHg and 1.101±1.9 mmHg, respectively; the correlation coefficients for the estimated and true BP are equal to 0.94. Conclusions. The model corresponds to the AAMI standard and the “A” grade in the BHS standard, which proves the high accuracy of BP assessment by the proposed approach. Comparison to other known methods was performed, which confirmed the advantages of the proposed approach.

Keywords

• blood pressure
• electrocardiography
• photoplethysmography
• neural network
• feedforward neural network