Deep Boltzmann Regression With Mimic Features for Oscillometric Blood Pressure Estimation

Abstract

Oscillometric blood pressure (BP) devices are among the standard automatic monitors, now readily available for the home, office, and hospital. The systolic blood pressure (SBP) and diastolic blood pressure (DBP) are obtained at fixed ratios of the envelope of the maximum amplitude of the oscillometric wave signal. However, these fixed ratios can cause overestimation or underestimation of the real SBP and DBP in oscillometric BP measurements. In this paper, we propose a new regression technique using a deep Boltzmann regression with mimic features based on the bootstrap technique to learn the complex nonlinear relationships between the mimic features vectors acquired from the oscillometric signals and the target BPs. The performance of the proposed model is compared with those of conventional and auscultatory techniques. Our regression model with mimic features provides lower standard deviation of error, mean error, mean absolute error, and standard error of estimates than the conventional techniques, along with a similar fit for the SBP and DBP.