Blood Pressure Estimation From Korotkoff Sound Signals Using an End-to-End Deep-Learning-Based Algorithm

Abstract

While measurement of blood pressure (BP) is now widely carried out by automated non-invasive BP (NIBP) monitoring devices, as they do not require skilled clinicians and do not carry risk of complications, their accuracy is in doubt. A novel end-to-end deep learning-based algorithm was developed in this study that estimates NIBP directly from sequences of Korotkoff sounds (KSs) rather than oscillometric waveforms. First, sequences of segments of KSs were formed using different signal segmentation techniques, i.e., segmentation using sliding window with or without overlap and segmentation using the cardiac period estimation. Each segment within each sequence was then labeled as (i) after-systolic and before-diastolic (AB), or (ii) before-systolic or after-diastolic (BA) such that a binary sequence-to-sequence classification problem was achieved. To deal with the resultant sequence-to-sequence classification problem, an algorithm was developed by combining one-dimensional (1D) convolutional neural networks (CNNs) and recurrent neural networks (RNNs). The segments associated with systolic and diastolic blood pressure (SBP and DBP) are then identified as the segments at which the output target sequence switches from class BA to class AB and later from class AB to class BA. Lastly, the values of SBP and DBP are obtained by mapping the center point of the switching segments to the deflation curve. To evaluate the performance of the proposed NIBP estimation method, we used a database of 350 NIBP samples collected from 155 participants (87 male, age: 23-97 years, arm circumference: 10-35 cm, SBP: 81-104 mmHg, and DBP: 37-104 mmHg), and the achieved estimation errors for SBP and DBP, relative to the reference values, using a 5-fold cross validation approach, were 1.6±3.9 mmHg (mean absolute error ± standard deviation of error) and 2.5±4.0 mmHg, respectively. We finally conclude that the proposed end-to-end deep learning-based NIBP estimation algorithm from sequences of KSs is a novel technique that requires modest preprocessing steps and can measure BP accurately.