A hybrid neural network for continuous and non-invasive estimation of blood pressure from raw electrocardiogram and photoplethysmogram waveforms

Stephanie Baker,Wei Xiang,Ian Atkinson

doi:10.1016/j.cmpb.2021.106191

Abstract

Background and objectives: Continuous and non-invasive blood pressure monitoring would revolutionize healthcare. Currently, blood pressure (BP) can only be accurately monitored using obtrusive cuff-based devices or invasive intra-arterial monitoring. In this work, we propose a novel hybrid neural network for the accurate estimation of blood pressure (BP) using only non-invasive electrocardiogram (ECG) and photoplethysmogram (PPG) waveforms as inputs.Methods: This work proposes a hybrid neural network combines the feature detection abilities of temporal convolutional layers with the strong performance on sequential data offered by long short-term memory layers. Raw electrocardiogram and photoplethysmogram waveforms are concatenated and used as network inputs. The network was developed using the TensorFlow framework. Our scheme is analysed and compared to the literature in terms of well known standards set by the British Hypertension Society (BHS) and the Association for the Advancement of Medical Instrumentation (AAMI).Results: Our scheme achieves extremely low mean absolute errors (MAEs) of 4.41 mmHg for SBP, 2.91 mmHg for DBP, and 2.77 mmHg for MAP. A strong level of agreement between our scheme and the gold-standard intra-arterial monitoring is shown through Bland Altman and regression plots. Additionally, the standard for BP devices established by AAMI is met by our scheme. We also achieve a grade of ‘A’ based on the criteria outlined by the BHS protocol for BP devices.Conclusions: Our CNN-LSTM network outperforms current state-of-the-art schemes for non-invasive BP measurement from PPG and ECG waveforms. These results provide an effective machine learning approach that could readily be implemented into non-invasive wearable devices for use in continuous clinical and at-home monitoring.

Highlights

Blood pressure (BP) is a key diagnostic tool for a variety of lifethreatening conditions
We propose a hybrid deep neural network (DNN) that incorporates temporal convolutional neural network (CNN) and long short-term memory (LSTM) layers for the estimation of SBP, DBP, and mean arterial pressure (MAP) from raw ECG and PPG waveforms with duration of 5 s
We evaluate the level of agreement between the calculations made by the CNN-LSTM networks and the expected SBP and DBP values as determined from ABP waveforms within the Medical Information Mart for Intensive Care (MIMIC) III database, which were obtained using gold-standard intra-arterial blood pressure measurement

Summary

Introduction

Blood pressure (BP) is a key diagnostic tool for a variety of lifethreatening conditions. Poor organ perfusion can be identified through the measurement of BP-derived parameters, mean arterial pressure (MAP). MAP is useful in determining overall blood flow and the level of nutrient delivery to organs, and is routinely measured when dealing with high-mortality conditions like septic shock [2]. MAP that is too low can lead to shock, syncope, and poor perfusion to organs, while elevated MAP places strain on the cardiovascular system and can eventually to various CVDs including stroke [3]. Blood pressure (BP) can only be accurately monitored using obtrusive cuff-based devices or invasive intra-arterial monitoring. We propose a novel hybrid neural network for the accurate estimation of blood pressure (BP) using only non-invasive electrocardiogram (ECG) and photoplethysmogram (PPG) waveforms as inputs

Objectives

Methods

Results