Cuffless blood pressure estimation based on composite neural network and graphics information
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)
- et al.
Blood pressure estimation from appropriate and inappropriate PPG signals using A whole-based method
Biomed. Signal Process. Control
(2019) - et al.
A novel dynamical approach in continuous cuffless blood pressure estimation based on ECG and PPG signals
Artif. Intell. Med.
(2019) - et al.
A review of machine learning in hypertension detection and blood pressure estimation based on clinical and physiological data
Biomed. Signal Process. Control
(2021) - et al.
Investigation on the effect of Womersley number, ECG and PPG features for cuff less blood pressure estimation using machine learning
Biomed. Signal Process. Control
(2020) - et al.
A review of machine learning techniques in photoplethysmography for the non-invasive cuff-less measurement of blood pressure
Biomed. Signal Process. Control
(2020) - et al.
Cuffless blood pressure estimation from electrocardiogram and photoplethysmogram using waveform based ANN-LSTM network
Biomed. Signal Process. Control
(2019) - et al.
A multistage deep neural network model for blood pressure estimation using photoplethysmogram signals
Comput Biol Med
(2020) - et al.
Performance of small-world feedforward neural networks for the diagnosis of diabetes
Appl. Math. Comput.
(2017) - et al.
Multi-stage classification of congestive heart failure based on short-term heart rate variability
Chaos, Solitons Fractals
(2019) - et al.
Hypertension: overly important but under-controlled,
Eur. J. Prevent. Cardiol.
(2017)
Heart Disease and Stroke Statistics-2021 Update: A Report From the American Heart Association
Circulation
Toward ubiquitous blood pressure monitoring via pulse transit time: theory and practice
IEEE Trans. Biomed. Eng.
Single-source PPG-based local pulse wave velocity measurement: a potential cuffless blood pressure estimation technique
Physiol. Meas.
Noninvasive cuffless blood pressure estimation using pulse transit time and impedance plethysmography
IEEE Trans. Biomed. Eng.
Toward ubiquitous blood pressure monitoring via pulse transit time: predictions on maximum calibration period and acceptable error limits
IEEE Trans. Biomed. Eng.
Cuff-less blood pressure estimation using pulse waveform analysis and pulse arrival time
IEEE J. Biomed. Health. Inf.
Characters available in photoplethysmogram for blood pressure estimation: beyond the pulse transit time
Australas. Phys. Eng. Sci. Med.
Feasibility study for the non-invasive blood pressure estimation based on ppg morphology: normotensive subject study
Biomed. Eng. Online
Cited by (17)
Prediction of arterial blood pressure waveforms based on Multi-Task learning
2024, Biomedical Signal Processing and ControlAutomatic identification of hypertension and assessment of its secondary effects using artificial intelligence: A systematic review (2013–2023)
2024, Computers in Biology and MedicineFully degradable, highly elastomeric and adhesive silk fibroin electronic skin for microdynamic pressure monitoring
2023, Chemical Engineering JournalBP-Net: Cuff-less and non-invasive blood pressure estimation via a generic deep convolutional architecture
2022, Biomedical Signal Processing and ControlCalibration-free blood pressure estimation based on a convolutional neural network
2024, Psychophysiology