Elsevier

Measurement

Volume 183, October 2021, 109803
Measurement

ECG compressed sensing method with high compression ratio and dynamic model reconstruction

https://doi.org/10.1016/j.measurement.2021.109803Get rights and content

Highlights

  • Compress sensing of electrocardiographic signal.

  • Frames with offset and amplitude normalization and requantization.

  • High compress ratio.

  • Reconstruction based on dynamic ECG model.

  • Optimization by differential evolution.

Abstract

This paper introduces an alternative method for compressed sensing and reconstruction of ECG that is patient agnostic and offers a high compression ratio. The high compression ratio is achieved by high decimation of the measurement signal and its post requantization, further decreasing the number of bits needed for information transfer. The sensing method also incorporates a QRS detector to detect exact R wave positions for signal segmentation before compression. ECG signal is also normalized in amplitude and offset, which maintains the bit resolution during requantization. The reconstruction employs a simple dynamic ECG model, parameters of which are calculated from the measurement signal by the Differential Evolution algorithm. The proposed method was evaluated using the MIT-BIH arrhythmia database and compared with two wavelet dictionary reconstruction methods. The proposed method keeps the structure of heartbeats preserved including the exact positions of R waves, and it reduces the noise interfering with ECG signals.

Introduction

The importance of remote health monitoring systems is arising, and the issues related to power efficiency are becoming more and more significant. Each health monitoring system uses measured bio-signals in the process of analysis. Sensor nodes employed in modern remote health monitoring systems and Wireless Body Area Networks (BAN) [1], [2] are required to meet certain specifications, including limited size, adequate costs and performance. One of the most significant challenges is the reduction of power consumption. It is proven that wireless data transmission is the most power-demanding activity of a sensor node impacting the battery life [3], [4]. One of the possible solutions is to lower the amount of transmitted data by employing a suitable compression technique [5]. Such technique should not demand a significant amount of additional power to perform the compression, which means it should be computationally simple and effective. The standard or well-known compression techniques based on Nyquist sampling and redundancy suppression are not sufficient. These methods rely on signal transformation into some spectral domain, i.e. Discrete Fourier Transform, Discrete Cosine Transform, Walsh Transform, and Discrete Wavelet Transform (DWT) [6], [7], [8]. The transform-based methods showed sufficient compression efficiency with accurate reconstruction. However, these systems are rather hardware demanding [9], [10], [11]. The transform-based methods combined with other approaches were investigated in [9]. These approaches incorporate standard deviation to split ECG signals into two frames, plain and complex. The different compression techniques are then used for these frames. Another method [12] is based on the combination of wavelet and Principal Component Analysis (PCA).

The compressed sensing (CS) appears to be more effective while the information is preserved even when the Nyquist rule is not fulfilled. Furthermore, the CS is not hardware demanding and the compression is achieved directly during the signal sensing. CS is a compression technique which exploits signal sparsity to represent the signal using fewer samples than needed according to the Nyquist-Shannon sampling theorem. It has recently emerged as a low-power alternative to the traditional lossy compression techniques [13], [14]. Various methods for the application of CS directly to analog signals were proposed [15], [16], [17], [18], [19], [20], [21], [22], knowns as the Analog to Information Converters (AIC). The common AIC incorporating a significant amount of analog circuitry in addition to one or more conventional ADCs do not seem to have significant power-saving benefits [23], [24]. Moreover, the analog circuitry may introduce distortion into the acquired signal causing the reconstruction more difficult [22]. The ECG is usually acquired using a conventional Nyquist rate ADC. Consequently, CS is applied digitally, providing a computationally simple compression [25]. There are several frame-based approaches as well. The approach proposed in [26] is based on the generation of a sensing matrix, which is adapted to the acquired signal frame. In [27], the QRS detection is performed directly from compressed data. This is done thanks to a trained dictionary. All the approaches rely on the power of the receiver node, where numerically complex tasks are solved to reconstruct the original signal. Many CS and reconstruction methods for ECG signals with excellent reconstruction quality were proposed [28], [29]. The methods use either a trained reconstruction dictionary, where a large database of training records or even a patient-specific database is required [30]. Nowadays, the deep learning approaches have become very popular as well [31], [32], [33]. Other methods are based on a suitable wavelet or a spline basis for the signal reconstruction (patient-agnostic methods) [26], [34], [35], [36]. The drawback of these methods is that the reconstruction quality gets significantly worse for high compression ratios. Moreover, these algorithms are evaluated using denoised ECG and with QRS annotation present in the MIT-BIH database.

The structure of the ECG signal consists of QRS components. It is possible to model these components using different function sets. Thus, we proposed a novel approach to ECG CS. The proposed method is based on the model presented in [37], which describes the 3D electrical activity of the heart by a couple of differential equations capable capturing morphological features of typical ECG signals. We decided to perform a study to implement the mentioned model in Compress Sensing. The model is capable capturing the morphological feature of the heart. Therefore, we assumed that the CS based on this model should offer a much higher compression ratio than methods based on a different reconstruction function basis.

This paper aims to propose an alternative and completely patient-agnostic compression and reconstruction method, which can be used in the case when high compression ratios are needed and a suitable training dataset is not available. The method does not require prior patient-specific information and relies only on the proposed dynamical model to create a synthetic reconstructed signal. The signal acquisition is performed framewise, where the frame splitting positions are found using a QRS detector. Parameters obtained by the QRS detector serve as additional information used by the decompressing procedure to guarantee the exact position of the R wave in the reconstructed signal. The frames are cycle-normalized before compression in order to simplify the reconstruction and increase the achievable compression ratio. Such an approach also guarantees that exact R wave positions are maintained in the reconstructed signal. A preliminary version of the proposed method was shortly introduced by the authors in [38]. The current paper presents a further improved and assessed method. In particular, the reconstruction is more deeply described, and the method is compared with two state-of-the-art methods performing CS-based ECG signal acquisition.

The organization of the paper is as follows: Section 2 briefly introduces the theoretical background of CS. The detailed description of the proposed modified CS acquisition method with the QRS detector is in Section 3. Section 4 describes the reconstruction method utilizing the dynamic ECG model. The performance of the proposed method is assessed using the MIT-BIH arrhythmia database. The results are compared with a wavelet dictionary reconstruction in Section 5. Finally, conclusions and future works are presented in Section 6.

Section snippets

Compressed sensing background

Compress sensing is based on assumption that signal may be represented by the reduced linear combination of suitable basis functions. If only a small number of basis functions is needed, the signal is called a sparse signal. Consider input signal vector x of length N that can be written as a linear combination of basis functions of the arbitrary domain, which is multiplied by expansion coefficients θ (1).x=Ψθwhere Ψ is the N × K size basis matrix, containing K columns of basis functions ψk k =

The proposed ECG compression algorithm

The proposed ECG compression algorithm is outlined in Fig. 1. The input signal is acquired using a conventional ADC at the sampling frequency of fs. An analog or digital QRS detector returns the exact positions of R waves nRi. Here a suitable QRS detection algorithm or a low power analog detection circuit in IoT such as [48], [49] can be used.

There is a wide range of digital QRS detectors including topological mapping for measurement of QRS complex energy, adaptive matched filtering based upon

Reconstruction method

The reconstruction method slightly differs from the abovementioned basic theory of CS and commonly used methods. Rather than using a dictionary, it relies on the dynamic ECG signal model, which is suitable for the generation of synthetic ECG. The Differential Evolution (DE) optimization algorithm [55] was utilized to find the optimal ECG model parameters for signal reconstruction (Fig. 2).

Experimental results and discussion

The proposed method was simulated in a LabVIEW programming environment, employing the MIT-BIH arrhythmia database [58] as a set of test signals. The database consists of 48 ECG records sampled at 360 Hz with 11 bit resolution.

To assess the performance let us first define the average compression ratio (CR). CR defined by (14) represents a reduction in the number of bits needed to describe the original I frames of heartbeats:CR=1Ii=0I-1Ni.B0BH+Mi.BC,where B0 is the bit resolution of the original

Conclusion

It was shown that the reconstruction of ECG signal acquired using a CS lossy data reduction can be done with no prior knowledge, patient-specific information, or a large record database just using a suitable time-domain signal model. The proposed method keeps the structure of heartbeats preserved, including the exact positions of R waves and works also for noisy and interfered signals. An advantage of the proposed method is the relatively good reconstruction quality in comparison with

CRediT authorship contribution statement

Ján Šaliga: Supervision, Conceptualization, Funding acquisition, Software, Validation, Writing - review & editing. Imrich Andráš: Methodology, Conceptualization, Investigation, Writing - original draft, Writing - review & editing. Pavol Dolinský: Conceptualization, Investigation, Methodology, Software, Validation, Data curation, Writing - original draft. Linus Michaeli: Supervision, Conceptualization, Methodology, Validation, Writing - review & editing. Ondrej Kováč: Validation, Writing -

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.

Acknowledgment

The work is a part of the project supported by the Science Grant Agency of the Slovak Republic (No. 1/0722/18).

References (58)

  • A. Maalej et al.

    On the wavelet-based compressibility of continuous-time sampled ECG signal for e-health applications

    Measurement

    (2020)
  • S.S. Mehta et al.

    K-means algorithm for the detection and delineation of QRS-complexes in Electrocardiogram

    IRBM

    (2010)
  • J.P. do Vale Madeiro et al.

    Evaluation of mathematical models for QRS feature extraction and QRS morphology classification in ECG signals

    Measurement

    (2020)
  • X. Xu et al.

    A Survey on Energy Harvesting and Integrated Data Sharing in Wireless Body Area Networks

    Int. J. Distrib. Sens. Netw.

    (2015)
  • S.-Y. Lee et al.

    Low-Power Wireless ECG Acquisition and Classification System for Body Sensor Networks

    IEEE J. Biomed. Health Inform.

    (2015)
  • S.C. Mukhopadhyay

    Wearable Sensors for Human Activity Monitoring: A Review

    IEEE Sensors J.

    (2015)
  • M. Ben-Romdhane et al.

    Event-Driven ECG Sensor in Healthcare Devices for Data Transfer Optimization

    Arab. J. Sci. Eng.

    (2020)
  • N. Adochiei et al.

    ECG waves and features extraction using Wavelet Multi-Resolution Analysis

    2011 E-Health Bioengineering Conference (EHB)

    (2011)
  • S. Chouakri, O. Djaafri, A. Taleb-Ahmed, Wavelet transform and Huffman coding based electrocardiogram compression...
  • X. Chen et al.

    A Sub-Nyquist Rate Sampling Receiver Exploiting Compressive Sensing

    IEEE Trans Circuits Syst.

    (2011)
  • P. Bera et al.

    Preserving Abnormal Beat Morphology in Long-Term ECG Recording: An Efficient Hybrid Compression Approach

    IEEE Trans. Instrum. Meas.

    (2020)
  • D.L. Donoho

    Compressed sensing

    IEEE Trans. Inform. Theory

    (2006)
  • E. Balestrieri et al.

    A Wi-Fi IoT prototype for ECG monitoring exploiting a novel Compressed Sensing method

    Acta IMEKO

    (2020)
  • M. Rani et al.

    A Systematic Review of Compressive Sensing: Concepts, Implementations and Applications

    IEEE Access

    (2018)
  • Z. Slavik et al.

    Compressive Sensing Hardware for Analog to Information Conversion

  • R. Agarwal, M. Trakimas, S. Sonkusale, Adaptive asynchronous analog to digital conversion for compressed biomedical...
  • A. Singh Alvarado et al.

    The integrate-and-fire sampler: A special type of asynchronous Σ - Δ modulator

    2011 IEEE International Symposium of Circuits and Systems (ISCAS), Rio de Janeiro, Brazil: IEEE

    (2011)
  • I. Andráš et al.

    Sparse Signal Acquisition via Compressed Sensing and Principal Component Analysis

    Measure. Sci. Rev.

    (2018)
  • L. Michaeli et al.

    Optimization Paradigm in the Signal Recovery after Compressive Sensing

    Measure Sci Rev

    (2019)
  • Cited by (0)

    View full text