- 334 views
- 647 downloads
Motion Compensated Continuous Blood Pressure Measurements Using Recurrent Neural Networks
-
- Author / Creator
- Ghosh, Shrimanti
-
High blood pressure (BP) is the leading cause of death and disability in the world, affecting nearly 1.5 billion adults. It leads to many complications, including stroke, heart failure, kidney disease, and coronary disease. In current clinical practice, BP is measured either invasively by an intra-arterial catheter that is invasive and causes pain or non-invasively by cuff using the oscillometric method, which does not allow continuous BP measurement. Therefore, development of an accurate, continuous, and non-invasive BP measurement method is needed for hypertension diagnosis and management. Prediction models based on pulse transit time (PTT) have been used for continuous and non-invasive BP estimation. PTT has been reported to be highly correlated with BP, which makes PTT a good candidate to be used for continuous BP monitoring, including for continuous ambulatory monitoring. PTT can be defined as the time lag between the R-peak of the ECG signal and the peak of the blood oxygenation signal (PPG) signal, when measured within the same cardiac cycle. In this study, windowed cross-correlation between the adjacent peak points of ECG and PPG signals in the same cardiac cycle is used to compute PTT automatically. We have performed sparsification of PPG signal by computing a moving window maximum on that signal. This pre-processing converts the PPG signal into a very sparse signal that ultimately increased the accuracy and the efficiency of PTT calculation. We then use a linear regression model to calculate BP continuously from PTT measurement where unknown constants of that model are subject dependent and must be calibrated. To improve the classical PTT-based prediction during motion, along with the ECG and the PPG signals, for the first time to our knowledge, we use signals from accelerometers and gyroscopes to predict BP. Accelerometers and gyroscopes have been widely accepted as useful and practical sensors for wearable devices to measure and assess physical activity. To make a prediction model, we use recurrent neural networks (RNNs), which can effectively learn the multi-timescale dependencies from a sequential time series of BP values. For this study, deep long short-term memory network (LSTM) and gated recurrent units (GRU) were used. BP was monitored in five different scenarios (seating, standing, recumbent, walking and cycling) from 50 healthy volunteers. We propose that the RNN models can predict continuous BP sequences from physiological signals like ECG, PPG and other parameters like PTT in posture and activity. Walking and cycling introduce baseline noise into both the ECG and PPG signals, making it more difficult to accurately estimate the BP values. To predict BP more accurately during activity, here we incorporated the accelerometric and gyroscopic values in the model. After including these data, we are able to achieve significant boost in the accuracy for all positions. The mean ±standard deviation is 0.08 ± 4.5 for SBP and 1.7 ± 3.4 for DBP in seated position and 2.3 ± 5.7 for SBP and 2.5 ± 4.2 for DBP while walking which is permissible according to the accepted threshold for accuracy using GRUs. Also, the root mean squared error between the reference standard and estimated SBP and DBP are 4.22 and 3.73 respectively for motionless position and are 6.10 and 4.80 for walking in case of GRUs. It can be stated that the difference between the estimated BP from RNN and the reference standard was less than the accepted threshold in all scenarios. The deep learning based method applied in this study appears sufficiently accurate in measuring BP not only in motionless conditions but also for walking and cycling, where motion artefacts are present. This novel approach has a significant potential contribution in continuous BP measurement in different postures and activity for hypertension management.
-
- Graduation date
- Spring 2018
-
- Type of Item
- Thesis
-
- Degree
- Master of Science
-
- License
- Permission is hereby granted to the University of Alberta Libraries to reproduce single copies of this thesis and to lend or sell such copies for private, scholarly or scientific research purposes only. Where the thesis is converted to, or otherwise made available in digital form, the University of Alberta will advise potential users of the thesis of these terms. The author reserves all other publication and other rights in association with the copyright in the thesis and, except as herein before provided, neither the thesis nor any substantial portion thereof may be printed or otherwise reproduced in any material form whatsoever without the author's prior written permission.