Cuffless blood pressure estimation based on composite neural network and graphics information

https://doi.org/10.1016/j.bspc.2021.103001Get rights and content

Highlights

  • A new method is provided for the noninvasive blood pressure estimation.

  • The blood pressure prediction is based on the real-time signals.

  • Features are extracted from graph information of inputs signals.

  • Stable blood pressure prediction can be performed in a large number of people.

Abstract

Blood pressure (BP) assessment and dynamic detection are of great significance for timely detection of the morbidity of hypertension, which is a major risk factor for most cardiovascular diseases (CVDs). It has been proved that the dynamic BP can be effectively predicted by using the combined input of photoplethysmogram (PPG) and electrocardiographic (ECG) signals. In this paper, we proposed a hybrid neural network architecture, which contains CNN-Sequential-Adapt layer, ResNet25_BP layer with squeeze and excitation (SE) block and fully connected layers, for BP estimation. The structure based on the convolutional network aims at the current inputs, which can effectively absorb the graph information of the inputted biological signals and make the model more stable and reliable. We evaluated the performance of two datasets including 1216 and 40 subjects, based on the criterions of British Hypertension Society (BHS) and the Association for the Advancement of Medical Instrumentation (AAMI). According to the BHS and AAMI standards, the outputs of the model achieved grade A on BHS and met the AAMI criteria. The mean absolute errors (MAE) of systolic BP and diastolic BP are 3.70 and 2.81 mmHg in the large dataset, and 1.37 and 0.93 mmHg in the small dataset, respectively.

Introduction

Cardiovascular diseases (CVDs) including hypertension and arrhythmia are the main causes of death worldwide, especially for the elderly living in countries that do not have adequate medical treatments [1]. The risk of CVDs has shifted to the young people who have been sub-healthy and inactive in the past few years. The occurrence of CVDs is usually sudden and fatal. Monitoring of blood pressure (BP) is considered as one of the necessary methods to avoid CVDs caused by the variation of BP [2].

The clinical noninvasive methods of BP measurement are mainly used by the blood cuff, including auscultation, oscillometry and volume clamping [3]. These BP measurement methods have been lasted for a long time and have some limitations, such as discontinuity and inconvenience. Obviously, these methods cannot be used to monitor the continuous BP of patients. Meanwhile, the invasive methods that can measure BP continuously were only used in patients who have serious disorder and medical observation in intensive care unit. The invasive methods are the gold standard, but cannot be the mainstream method because its cost is unacceptable for people without serious illness. Due to medical needs, simple continuous BP monitoring methods are urgently needed to be applied in clinical diagnosis.

The cost of continuous BP detection will be greatly reduced if BP can be obtained by correlating it with some biological signals which are easily detectable, such as photoplethysmography (PPG) and electrocardiograph (ECG). Previous works have proved that BP can be effectively predicted by using the combined PPG and ECG. The ECG and PPG which are easy to obtain from the wearable device or instrument have become the signals to estimate BP in recent decades [4], [5].

In the early applications of these signals, the pulse wave velocity (PWV) and the pulse transit time (PTT) were used to build a complex linear or nonlinear relationship with BP [6], [7]. The relationship between BP and pulse wave has been studied for a long time. Mukkamala et al. used a model for the relationship between PTT, age, and sex and BP to determine the PTT-BP calibration curves for each age and sex [8]. Singla et al. adopted wavelet transform on ECG and PPG signals, and estimated BP value through the joint detection of these signals [9]. Pulse arrival time (PAT) that is defined as the time delay from the R-peak of ECG to the next trough of PPG during a specific period was also considered as a parameter to estimate BP [10]. Liu et al. studied the information based on the waveform of PPG and estimated BP with multiple regression model [11]. A method based on the combination of PPG signals morphology and ECG was proposed by Li et al. [12]. Shin et al. simplified the PPG waveform by using an approximate model and then analyzed it as blood flow velocity and acceleration using the derivative of PPG with the pressure index (PI) introduced as a new factor [13].

Recently, the machine learning (ML) methods were also applied in the field of cuffless BP estimation based on the clinical data or the relevant physiological signals [14]. Early ML methods typically used logistic regression analysis, support vector machine (SVM) and other methods based on the manually selected feature extraction. Features selected from PPG signals include the width of the 1/2 and 2/3 amplitudes, the foot extracted PAT, the midpoint and peak value, the systolic upstroke time, and the diastolic time [15]. The eigenvalues, such as the Womersley number, were also suggested as feature inputs which contained information of waveform for BP estimation [16]. Kachuee et al. discussed a ML method to extract feature values from PPG and then predict BP with SVM [17]. Kachuee et al. also proposed a framework of AdaBoost which consists of 1000 decision trees for BP estimation with the extraction of two types of features [18]. Ding et al. improved the accuracy of long-term BP monitoring by introducing new indicators, photoplethysmogram intensity ratio (PIR), into the regression model [19]. Lin et al. extracted nineteen eigenvalues based on PPG and PTT on a small sample and estimated BP using the linear regression [20]. Feng et al. proposed a ML strategy based on the regularized linear regression (RLR) to construct BP models with different covariates for the corresponding groups on 28 subjects. The RLR of the individual was used as the initial calibration, and the recursive least squares method was used for re-calibration [21]. Hassani et al. applied a nonlinear mapping to reduce the size of the feature vector by mapping the input parameters to a potential space, used a multi-stage noise reduction technique to effectively smooth the input signals, and then considered SVM to estimate BP values [22]. By evaluating the correlation between various characteristic points and BP, it was found that the diastolic time (DT) which can be computed by the distance from the peak to foot of PPG has a high relevance with BP [23]. Hu et al. used a single-channel PPG signals to estimate systolic BP (SBP) and diastolic BP (DBP) by an integrated ML algorithm of XGBoost [24].

Due to the specificity of the biological signals from the individual bodies, the features extracted from morphology are usually error prone. Furthermore, the morphological features with manual selection may have a good performance, but the ML-based methods need the professional medical knowledge to set up the rule-making formula [25]. Thus, different from the above statistical rule-making ML methods, the deep learning (DL) technology was applied to BP estimation, which abandons the statistical rule-making processes. A hybrid neural network model based on the mean impact value and genetic algorithm was discussed for BP prediction from PPG signal [26]. Tanveer and Hasan used the fully connected layer and LSTM layer neural network to predict BP on a small sample [27]. Slapničar et al. took the PPG and its first and second derivatives as inputs into a spectral-temporal deep neural network with residual connections to mimic the dependence between PPG and BP [28]. A composite network structure using convolutional neural network (CNN) and LSTM was proposed by JEsmaelpoor et al. in which the CNN layer was used to extract eigenvalues, and then sent to LSTM for BP estimation based on temporal variation [29]. Eom et al. proposed a model consisted of CNN, a bidirectional gated recurrent unit, and an attention mechanism [30]. Through the experimental comparison, it was found that the model with attention mechanism shows better performance for BP prediction. Aguirre et al. discussed a cuffless method to estimate the morphology of the arterial BP through a deep learning model based on a seq2seq architecture with attention mechanism [31]. Beyond the BP predication, the deep neural network is also able to diagnose various diseases, such as diabetes and congestive heart failure, by inputting various biological signals, such as blood glucose concentration, diastolic BP, body mass index, heart rate variability [32], [33].

In this work, we focus on the feature value extraction based on graphics information of PPG and ECG, and adopt a carefully designed feature extraction structure, which is more suitable for a better BP prediction with large subject databases. In our model, the input signals are treated as an overall picture, in which the length is the size of the signal segment, and the width is the number of signal types. We consider PPG and ECG as two-dimensional data from the same channel. The CNN is used to compute the hidden features of PPG and ECG, which can replace the artificial feature extraction. In the Section 1, we mainly review the various methods in the field of cuffless BP estimation, including linear correlation, ML and DL models, and propose a new perspective to deal with PPG and EEG. The main content of the Section 2 is the signals preprocessing, the structure of proposed model and the specific experimental details. The comparison of statistical results of the models is described in the Section 3. Discussions and conclusions about the results and the performance of the model are put forward in the 4 Discussion, 5 Conclusion respectively.

Section snippets

Materials and method

In this section, we describe the preprocessing phase including data slicing and resampling and introduce a more robust model structure to extract features effectively. The data extracted from the Multiparameter Intelligent Monitoring in Intensive Care (MIMIC-II) database [34] include PPG, ECG and ambulatory blood pressure (ABP).

Result

Since the gold standard of BP value is determined by the invasive method, the evaluation standard of BP calculated by PPG and ECG is to minimize the error between the estimation performance and the invasive BP. For evaluating the reliability of the proposed model, we extract 1216 and 40 records containing many of waveform information from the database, respectively. The advantage of the model proposed in this paper is that it can extract macro features. Thus, the analysis of the statistical

Discussion

In this paper, we proposed a composite neural network consisting of CNN-Sequential-Adapt layer, Renet25_BP layer with SE block, and fully connected layers. Inspired by ML in the field of cuffless BP estimation, this paper innovatively proposed a completely graph-based neural network structure for BP prediction. The first layer of the compound neural network transforms the initial input graph information into the input shape that can be accepted by Renet25_BP layer. The accuracy improvement

Conclusion

The deep learning model based on graph information proposed for BP prediction in this paper is more effective in the performance of the database with more subjects. The previous researches on the extraction of PPG and ECG features with ML methods indicate that the biological signals such as PPG and ECG have more graph information to be mined. The accuracy of BP prediction can be significantly improved by adding CNN modules and SE block in deep learning model to promote the feature extraction

CRediT authorship contribution statement

Ye Qiu: Conceptualization, Methodology, Software, Formal analysis, Writing - original draft. Dongdong Liu: Data curation, Witing - review & editing. Guoyu Yang: Witing - review & editing. Dezhen Qi: Writing - review & editing. Yuer Lu: Validation. Qingzu He: Investigation. Xiangyu Qian: Resources. Xiang Li: Visualization. Jianwei Shuai: Supervision, Project administration.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

We acknowledge supports from the National Natural Science Foundation of China (Grant Nos. 11874310, 12090052, and 11704318), the China Postdoctoral Science Foundation (Grant No. 2016M602071), the 111 Project (Grant No. B16029) and the Fujian Province Foundation (Grant No. 2020Y4001).

References (41)

  • S.S. Virani et al.

    Heart Disease and Stroke Statistics-2021 Update: A Report From the American Heart Association

    Circulation

    (2021)
  • R. Mukkamala et al.

    Toward ubiquitous blood pressure monitoring via pulse transit time: theory and practice

    IEEE Trans. Biomed. Eng.

    (2015)
  • P.M. Nabeel et al.

    Single-source PPG-based local pulse wave velocity measurement: a potential cuffless blood pressure estimation technique

    Physiol. Meas.

    (2017)
  • T.H. Huynh et al.

    Noninvasive cuffless blood pressure estimation using pulse transit time and impedance plethysmography

    IEEE Trans. Biomed. Eng.

    (2019)
  • R. Mukkamala et al.

    Toward ubiquitous blood pressure monitoring via pulse transit time: predictions on maximum calibration period and acceptable error limits

    IEEE Trans. Biomed. Eng.

    (2018)
  • M. Singla, P. Sistla, S. Azeemuddin, Cuff-less Blood Pressure Measurement Using Supplementary ECG and PPG Features...
  • Y.-Z. Yoon et al.

    Cuff-less blood pressure estimation using pulse waveform analysis and pulse arrival time

    IEEE J. Biomed. Health. Inf.

    (2018)
  • S.H. Liu, S.H. Lai, J.J. Wang, et al., The Cuffless Blood Pressure Measurement with Multi-dimension Regression Model...
  • Y. Li et al.

    Characters available in photoplethysmogram for blood pressure estimation: beyond the pulse transit time

    Australas. Phys. Eng. Sci. Med.

    (2014)
  • H. Shin et al.

    Feasibility study for the non-invasive blood pressure estimation based on ppg morphology: normotensive subject study

    Biomed. Eng. Online

    (2017)
  • Cited by (17)

    View all citing articles on Scopus
    View full text