Pulse plethysmography based cuffless blood pressure estimation system using time-series neural networks

Author: Sitansu Sekhar

Sekhar, Sitansu, 2022 Pulse plethysmography based cuffless blood pressure estimation system using time-series neural networks, Flinders University, College of Science and Engineering

Terms of Use: This electronic version is (or will be) made publicly available by Flinders University in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. You may use this material for uses permitted under the Copyright Act 1968. If you are the owner of any included third party copyright material and/or you believe that any material has been made available without permission of the copyright owner please contact copyright@flinders.edu.au with the details.


Blood pressure monitoring is vital in health care. Non-invasive monitoring systems such as manual sphygmomanometers and digital oscillometers are used by the medical professionals due to their ease of use. However, their accuracy is limitedExisting non-invasive monitoring systems such as manual sphygmomanometers and digital oscillometers are inaccurate and cannot be used in 24-hour monitoring. To address this, substantial research has been done in cuff-less continuous pressure monitoring systems using electrocardiography (ECG) and pulse-plethysmography (PPG) sensors. The focus has been extracting heuristic features from these signals and using machine learning algorithms such as deep neural networks. However, problems such as low sample size (10 – 100 patients) and sample-wise rather than patient-wise distribution in training and testing datasets have provided a false sense of high accuracy in studies. Further, there has been no justification for how the heuristic signal features (rise time, fall time, and others) benefit estimating pressure measurements. Whole-based features such as the entire signal as input have been explored but not trained and evaluated on big data.

Addressing these gaps, a two-part stacked model to estimate arterial blood pressure waveform from the PPG waveform is devised. The first part uses whole-based features derived using maximally overlap discrete wavelet transform (MODWT). It then uses Bi-Directional LSTM to estimate continuous ABP wavelet coefficients from PPG wavelet coefficients and reconstruct a continuous pulse pressure (ΔP) waveform. In the second part, a Convolutional LSTM (Conv-LSTM) network has been devised, which uses the estimated ABP wavelet coefficients from Bi-LSTM and a few PPG features as inputs to estimate the systolic (SBP) and diastolic (DBP) blood pressure. Training on 8,614 patients and testing on 2,828 patients from Physionet MIMIC database (collected using Philips Monitor), the RMSE of ΔP_Pred to ΔP_Actual was at 5.03 mmHg. The SBP and DBP estimates from the Conv-LSTM network provided an error of ±6.15 mmHg and \+6.46 mmHg, respectively.

Further, to test the scalability of the model, a two-part experiment was conducted. In the first part, ABP and PPG signals were collected from 129 patients from the Intensive Care Unit (ICU) of Flinders Medical Centre (FMC), South Australia, using a similar Philips System. The evaluation of the model showed a ΔP RMSE of 7.499 mmHg, SBP error of 1.98±7.47 mmHg and DBP error of -0.71±7.67 mmHg. In the second part, a custom PPG system was designed from which the PPG signals were collected along with the SBP and DBP readings from the invasive monitor from 78 new patients from FMC ICU. The evaluation showed the SBP error at -1.42±5.86 mmHg and the DBP error at -0.50±7.39 mmHg. The results were within the American Association of Medical Instrumentation (AAMI) error standards of 5±8 mmHg.

The use MODWT-BiLSTM stacked models is a unique approach which was developed and validated in this thesis to estimate the shape of the pulse pressure waveform from the PPG waveform. Further to estimate the absolute blood pressure readings, a convolutional-LSTM architecture was developed to take the predicted pulse pressure waveform from MODWT-BiLSTM and estimate the SBP and DBP values. This convoluted architecture helped in estimating a continuous ABP waveform from a PPG waveform. Trained on an extensive dataset and further evaluating the scalability in a different dataset, this thesis is able to present a non-invasive cuff-less blood pressure estimation algorithm which can help in improving accuracy of non-invasive monitors. With unique whole-based features and architecture, further training and validating on by far the most extensive available dataset, and further evaluating the scalability, it is concluded that non-invasive cuff-less blood pressure estimation could be made accurately using deep neural networks.

Keywords: continuous blood pressure, cuff-less systems, machine learning, deep learning, time series neural networks, pulse plethysmography

Subject: Engineering thesis

Thesis type: Doctor of Philosophy
Completed: 2022
School: College of Science and Engineering
Supervisor: Prof Karen Reynolds