GMM-HMM-Based Blood Pressure Estimation Using Time-Domain Features

Abstract

This article presents a novel method of estimating systolic blood pressure (SBP) and diastolic BP (DBP) from time-domain features extracted from auscultatory and oscillometric waveforms and using Gaussian Mixture Models and Hidden Markov Model (GMM-HMM). The nine time-domain features selected include the cuff pressure (CP), the cardiac period (T), the energy of the Korotkoff pulses (KE), the oscillometric waveform envelope (OWE), the lag between the trough of the oscillometric waveforms (OWs) and the peak of the Korotkoff energy (Lag), the time between the trough and the peak of the OW (OWD), the slopes of the KE and OWE (SKE, SOWE), and the maximum upslope of individual OWs (MSOW). Adopting a fivefold cross-validation scheme and using a database of 350 noninvasive BP (NIBP) recordings gave an average mean error (± standard deviation of error) of -0.3 ± 4.2 mmHg for SBP and 2.9 ± 8.1 mmHg for DBP relative to reference values derived from a visual method of determining SBP and DBP. The significantly larger spread of DBP estimates relative to SBP, suggests that the criteria for determining DBP are poorly defined and would benefit from further experimental studies involving simultaneous invasive and noninvasive methods of measuring arterial pressure. We conclude that the proposed GMM-HMM BP estimation method outperforms previously reported methods in the literature and is a very promising method in improving the accuracy of automated noninvasive measurement of BP.