Prediction of arterial blood pressure waveforms from photoplethysmogram signals via fully convolutional neural networks

https://doi.org/10.1016/j.compbiomed.2021.104877Get rights and content

Highlights

  • Propose an end-to-end ABP-Net to predict ABP waveforms from PPG signals for deriving more physiological parameters.

  • Employ the original, the first and the second derivatives of PPG signals as the inputs of the ABP-Net inputs to predict the ABP waveforms.

  • Conduct the calibration-based subject-independent experiment via meta-learning to reduce the amount of ABP reference data.

Abstract

Cardiovascular disease (CVD) is one of the most serious diseases threatening human health. Arterial blood pressure (ABP) waveforms, containing vivid cardiovascular information, are of great significance for the diagnosis and the prevention of CVD. This paper proposes a deep learning model, named ABP-Net, to transform photoplethysmogram (PPG) signals into ABP waveforms that contain vital physiological information related to cardiovascular systems. In order to guarantee the quality of the predicted ABP waveforms, the structure of the network, the input signals and the loss functions are carefully designed. Specifically, a Wave-U-Net, one kind of fully convolutional neural networks (CNN), is taken as the core architecture of the ABP-Net. Besides the original PPG signals, its first derivative and second derivative signals are all utilized as the inputs of the ABP-Net. Additionally, the maximal absolute loss, accompany with the mean squared error loss is employed to ensure the match of the predicted ABP waveform with the reference one. The performance of the proposed ABP network is tested on the public MIMIC II database both in subject-dependent and subject-independent manners. Both results verify the superior performance of the proposed model over those existing methods accordingly. The mean absolute error (MAE) and the root-mean-square error (RMSE) between the predicted waveforms via the ABP-Net and the reference ones are 3.20 mmHg and 4.38 mmHg during the subject-dependent experiments while those are 5.57 mmHg and 7.15 mmHg during the subject-independent experiments. Benefiting from the predicted high-quality ABP waveforms, more ABP related physiological parameters can be better obtained, which effectively expands the application scope of PPG devices.

Introduction

Even under today's highly prosperous medical conditions, cardiovascular disease (CVD) is still one of the most fatal diseases, especially for the elderly [[1], [2]]. The CVD will significantly affect the arterial blood pressure (ABP) waveform [3,4]. Various ABP waveform related physiological parameters, such as cardiac output (CO) [5], systolic blood pressure (SBP), diastolic blood pressure (DBP), mean arterial pressure (MAP) [6,7] and blood pressure variability (BPV) [8] etc., are good indicators to reflect the cardiovascular status. Besides, the shape of the ABP waveform is also important in CVD diagnosis of atrial fibrillation and arterial stiffness, etc. [4]. Therefore, it is of physiological significance to measure ABP waveforms. However, many traditional ways of ABP waveform measurements are usually invasive and are only used in intensive care unit (ICU).

Existing non-invasive ABP measurements are usually only limited to estimate BP values. For example, cuff-based methods like mercury or electronic sphygmomanometers [9,10], can measure BP values during each measuring period. With the development of sensor techniques, photoplethysmography (PPG) is another widely used way to measure BP values [11,12]. The principle is that PPG makes use of the low-intensity infrared light to capture the changes in blood volume proportional to the changes in the intensity of the absorbed light. Considering that pulse wave velocity (PWV) is the velocity of the BP wave flowing through blood vessels, it is strongly related to blood volume and BP values [13]. The PWV can be calculated by measuring the pulse transmission time when the transmission distance is fixed. In this case, the features, such as pulse transit time (PTT) [14] and pulse arriving time (PAT) [11], can be extracted to estimate BP values [[15], [16], [17], [18]]. However, the measurement of these parameters often need to combine the PPG signals with the other signals (such as the ECG signals, or the other PPG signals from a different body position) [19], which leads to a limitation of the scope of application.

Recently, a large number of studies have shown that it is feasible to estimate BP values using the single-channel PPG signals [[20], [21], [22], [23]]. Since both the ABP and the PPG signals have the same source of heart excitation, it is reasonable to expect a similarity between them in both time and frequency domains [24]. Gloria et al. [24] pointed out that BP monitoring via the PPG signals is potentially a feasible replacement of the invasive ABP waveform, provided that both time and frequency domain features are carefully extracted from both the raw single-channel PPG signals and its corresponding derivative signals [25]. In addition to estimate BP values from hand-crafted features of the PPG signals [21,22,24,25], the BP values can also be directly predicted from the PPG signals in an end-to-end way [19,26,27] using deep learning technology. Although the feasibility of predicting BP values from a single PPG sensor has been verified, it is still important to measure other ABP-related parameters such as CO and BPV for the diagnosis of CVDs. Therefore, it is worth developing non-invasive measurements of ABP waveforms from the PPG signals. In this way, more cardiovascular related physiological parameters will be measured.

In this paper, we propose an end-to-end model, named ABP-Net, to predict the ABP waveforms from the PPG signals. The ABP-Net is built with the Wave-U-Net [28], which is efficient to translate homologous waves through feature downsampling and concatenation at different time scales. To better predict the ABP waveform, both the inputs and the loss functions of ABP-Net are well designed. The original PPG signal, its first derivative and second derivative signals, termed as velocity plethysmography (VPG) and acceleration plethysmography (APG), respectively, are all employed as the inputs. The mean squared error (MSE) loss is taken to constrain the overall deviation between the predicted ABP waveform and the reference ABP waveform, while the consistency of local characteristics is restricted by the maximal absolute loss function as [29]. The two loss functions are combined together to ensure the match of the predicted ABP waveform with the reference ABP one. Previous studies [15,19,30], have deeply discussed the impact of subject-dependent and subject-independent data division methods to the performance of parameter estimation. Due to individual physiological differences, subject-dependent manner can achieve improved experimental results while subject-independent manner has more practical significance. Therefore, we test the proposed ABP-Net model on the public database MIMIC-II [31] both under subject-dependent and subject-independent conditions and the results prove that the quality of the predicted ABP waveforms is satisfactory and the quality of extracted physiological parameters outperforms those ones from most existing methods accordingly.

The main contributions of this work are summarized as follows:

  • We propose a new end-to-end method named ABP-Net to predict ABP waveforms from PPG signals. More physiological parameters representing cardiovascular status can be estimated in a non-invasive way from the predicted high-quality ABP waveforms, which effectively broadens the potential application scope of PPG devices.

  • The inputs of the proposed ABP-Net have been carefully chosen and the loss functions have also been customized to restrict the consistency of both global shape and local characteristic. The effectiveness of the inputs and loss functions have been verified to improve the quality of the predicted ABP waveforms.

  • Both subject-dependent and calibration-based subject-independent experiments have been conducted on the public database. Experimental results have demonstrated the effectiveness of our proposed ABP-Net.

The rest of the paper is structured as follows: Section II introduces related work. Section III provides the details of the proposed methodology. Section IV presents the results and the analysis. Finally, Section V concludes the paper.

Section snippets

Related work

The existing BP estimation methods based on PPG signals can be divided into two types. One type of methods are first to extract appropriate features from the PPG signals, and then to map these features to BP values through regression algorithms or neural networks [[32], [33], [34]]. The other type of methods map the input PPG signals to BP values in an end-to-end way via deep-learning-based techniques. Before sending to the neural networks, the PPG signals may be denoised and normalized [19,26,

ABP-net model

The core architecture of the proposed ABP-Net is the Wave-U-Net, one kind of fully CNNs, and is originally designed for the blind source separation (BSS) of speech signals [28]. The generation of the ABP waveforms from the PPG signals can be treated as the separation of the target component (the ABP waveform) from the observed PPG signals containing cardiovascular information. In order to ensure the predicting performance of the ABP waveform, we carefully design the inputs and the loss

Experiments

The flow chart of the proposed ABP-Net is illustrated in Fig. 2. It includes three stages, called pre-processing, training and testing. All the deep learning models are implemented in Pytorch 1.8.1. Training and evaluation are done on Nvidia GA102 [GeForce RTX 3090] and Intel core i9-10900x @3.70 GHz. The average time to obtain each segment of ABP waveform is about 0.5 s.

Conclusion

In this paper, we have proposed a deep learning model, named ABP-Net, to predict the ABP waveforms from the PPG signals. This model has been carefully designed with the PPG, VPG and APG signals as inputs, while the MSE and the maximal absolute loss functions are taken to guarantee the consistency of the predicted and reference ABP waveforms. Many ABP related parameters, rather than only BP values, could be well predicted with the obtained high-quality ABP waveforms, for the diagnosis of the

Conflict of interest and authorship conformation form

Please check the following as appropriate:

All authors have participated in (a) conception and design, or analysis and interpretation of the data; (b) drafting the article or revising it critically for important intellectual content; and (c) approval of the final version.

This manuscript has not been submitted to, nor is under review at, another journal or other publishing venue.

The authors have no affiliation with any organization with a direct or indirect financial interest in the subject

Acknowledgement

This work was supported in part by the National Natural Science Foundation of China (Grants 62171176, 61922075, 62176081 and 41901350), in part by the National Defense Basic Scientific Research Program of China (Grant JCKY2019548B001), in part by the Anhui Key Project of Research and Development Plan (Grant 202104d07020005), and in part by the Fundamental Research Funds for the Central Universities (Grants JZ2021HGPA0061, PA2021KCPY0051, JZ2020HGPA0111 and JZ2021HGTB0078).

References (53)

  • R. Miotto et al.

    Deep learning for healthcare: review, opportunities and challenges

    Briefings Bioinf.

    (2018)
  • D. Ravì et al.

    Deep learning for health informatics

    IEEE Journal of Biomedical and Health Informatics

    (2016)
  • G. Mancia et al.

    Blood pressure and heart rate variabilities in normotensive and hypertensive human beings

    Circ. Res.

    (1983)
  • W.M. Kirkendall et al.

    Recommendations for human blood pressure determination by sphygmomanometers

    Circulation

    (1967)
  • J. Jilek

    Electronic sphygmomanometers: the problems and some suggestions

    Biomed. Instrum. Technol.

    (2003)
  • M. Sharma et al.

    Cuff-less and continuous blood pressure monitoring: a methodological review

    Technologies

    (2017)
  • J.C. Bramwell et al.

    The velocity of pulse wave in man

    Proc. R. Soc. Lond. - Ser. B Contain. Pap. a Biol. Character

    (1922)
  • L. Geddes et al.

    Pulse transit time as an indicator of arterial blood pressure

    Psychophysiology

    (1981)
  • M. Kachuee et al.

    Cuffless blood pressure estimation algorithms for continuous health-care monitoring

    IEEE (Inst. Electr. Electron. Eng.) Trans. Biomed. Eng.

    (2017)
  • J. Dey et al.

    Instabp: cuff-less blood pressure monitoring on smartphone using single PPG sensor

  • P. Su et al.

    Long-term blood pressure prediction with deep recurrent neural networks

  • S. Baek et al.

    End-to-end blood pressure prediction via fully convolutional networks

    IEEE Access

    (2019)
  • A.S. Zadi et al.

    Arterial blood pressure feature estimation using photoplethysmography

    Comput. Biol. Med.

    (2018)
  • M. Radha et al.

    Estimating blood pressure trends and the nocturnal dip from photoplethysmography

    Physiol. Meas.

    (2019)
  • Y.-C. Hsu et al.

    Generalized deep neural network model for cuffless blood pressure estimation with photoplethysmogram signal only

    Sensors

    (2020)
  • G. Martínez et al.

    Can photoplethysmography replace arterial blood pressure in the assessment of blood pressure?

    J. Clin. Med.

    (2018)
  • Cited by (40)

    View all citing articles on Scopus
    View full text