2 F. Picariello et al.

/ Measurement 167 (2021) 108259

on the extraction of some features of the ECG signal such as the P dom matrices are used as it has been proven their incoherence
wave (the earliest wave of the ECG cycle, corresponding to the atria with Fourier and wavelet dictionaries. In [16], the authors evalu-
depolarization), the T wave (corresponding to the repolarization of ated the Toeplitz, Circulant and Triangular structured sensing
the ventricles) or the QRS complex (corresponding to the depolar- matrices, because they do not require the random number genera-
ization of the right and left ventricles). A method falling in this cat- tion and therefore are easier to implement and less complex. In
egory was presented in [8]. This method is composed by a [17,18], a CS–based sensing scheme using a Deterministic Binary
preprocessing phase, where the signal is divided in beats and each Block Diagonal matrix (DBBD) as sensing matrix is utilized. In par-
beat is further segmented to find the P section, the QRS section and ticular, in [17] a hardware implementation of the sensing scheme
the T section. Each section is filtered with a different filter and the is also proposed and compared with the random demodulator
baseline is removed. Then, an encoder searches the best match of Analog-to-Information Converter architecture. In [18], a digital CS
the preprocessed beat with the entries of a codebook and performs scheme using the same sensing matrix is proposed and compared
a Long-Term Prediction (LTP). Finally, the residue between the out- with Gaussian random matrices. About reconstruction algorithms,
put of the codebook and the predicted waveform from the LTP is Basis Pursuit (BP) and Orthogonal Matching Pursuit (OMP) are
furtherly coded. In transform domain methods, the ECG signal is mainly used in the literature. In [16], also Compressed Sampling
projected to a transform domain by means of a linear orthogonal Matching Pursuit (CoSaMP) and Normalized Iterative Hard Thresh-
transformation. Then the coefficients in the transformed domain olding (NIHT) have been evaluated for ECG reconstruction. Alterna-
are properly encoded. Some widely used transforms include Dis- tive approaches have been proposed in [19] and in [20]. In the
crete Fourier Transform (DFT), Discrete Cosine Transform (DCT), former, a parametric model of the ECG signal is used and the signal
Walsh Transform and Discrete Wavelet Transform (DWT) [5,9]. In recovery is achieved by means of the Differential Evolution algo-
[10], the ECG is compressed by first evaluating the DWT based rithm. In the latter, a time-normalized agnostic dictionary created
on the Db6 wavelet function. Then, the obtained coefficients are by the PCA of training signals is instead used as signal base.
selected by the application of a higher order statistics thresholding. This paper aims to propose a new digital CS–based ECG com-
The compressed signal is obtained by linear prediction coding and pression method where a deterministic sensing matrix, adapted
Huffman coding. The transform domain methods have gained a to the acquired signal, is utilized. Differently from random sensing
significant attention due to their capability of accurately represent matrices, the sensing matrix does not require the random number
the ECG signal also at moderate compression ratios. However, most generation. Moreover, being adapted to the signal, the sensing
of them are unsuitable for a real–time implementation in acquisi- matrix contains more information on the signal features and there-
tion nodes with constrained resources, due to their computational fore it can guarantee a better reconstruction quality, than other
complexity and large buffer memory requirements [11]. Hybrid deterministic matrices.
approaches, combining methods from different categories have The present research aims to realize the ECG signal compression
been proposed in [11,12]. The method in [11] first detects the to be integrated in the smart wearable device of the Ambient-
beats, it divides each beat into plain and complex blocks, based intelligent Tele-monitoring and Telemetry for Incepting & Catering
on their standard deviation. Then, specific compressions are over hUman Sustainability (ATTICUS) project [21]. The project
applied to the two block types. The method in [12] faces the prob- main objective is to develop a telemedicine system where physio-
lem of unacceptable distortion of abnormal beats for high com- logical signals are acquired by means of a smart T-shirt and possi-
pression ratios. It uses a Support–Vector Machine (SVM) binary ble anomalies in the parameters are discovered.
classifier to identify abnormal beats. Then a wavelet–based com- A preliminary version of the proposed method was presented in
pression is used for abnormal beats, while normal beats are com- [22]. In the work presented here, the aforementioned method was
pressed in groups by means of a hybrid encoder, employing a further improved and assessed. In particular, the compression
combination of wavelet and Principal Component Analysis (PCA). algorithm was modified such that the sensing matrix is not evalu-
Alternatively, Compressive Sampling (CS) has been proposed for ated in each frame, but only whether a significant change in the
ECG signal compression. The advantage of CS is its capability of signal distribution is found. Moreover, the proposed method has
achieving performance comparable with transform-domain meth- been compared against two state of the art methods performing
ods, while moving the computational load from the acquisition CS–based ECG signal acquisition.
node to the receiving node. Since the receiving node is usually The rest of the paper is organized as follows. In Section 2, some
located on the Internet cloud, it has much greater computational background knowledge about CS is given. In Section 3, the pro-
resources available. Therefore, CS methods allows to realize ECG posed CS–based ECG compression scheme is presented. The exper-
compression also on sensing nodes with constrained resources. imental evaluation is presented in Section 4. Finally, Section 5
For this reason, CS represents a widely proposed solution for data presents the conclusions and future work.
compression in IoT systems [13], where the acquisition and com-
pression can be implemented in resource-constrained nodes, while
the reconstruction is actually carried out (where needed) in the 2. Fundamentals of compressive sampling
cloud.
The literature regarding the application of CS for ECG compres- CS is based on the assumption that the ECG signal is sparse in a
sion is mostly concentrated on the quality assessment and transform domain, i.e. its representation in such domain is a vector
improvement of the reconstructed signal by investigating the with few elements. In particular, let us consider a vector x of N
impact on the quality of a specific sensing matrix, dictionary matrix, samples acquired in a certain time window, at Nyquist rate,
or reconstruction algorithm. (see Section 2 for further details about expressed as:
CS) [14]. In [15], the authors compared the performance of the ECG
x ¼ Wh; ð1Þ
reconstruction with the application of different dictionary matri-
ces, based on the Discrete Wavelet Transform, belonging to several where W is a N  N matrix describing the domain transformation,
wavelet families, such as Coiflet, Daubechies, Symlet, biorthogonal often called dictionary matrix, and h is the vector of the coefficients
and reverse biorthogonal. Another comparison of several dictionar- in the transformed domain. The vector h is assumed to be K-sparse,
ies was reported in [14], where the dictionary matrix was based on i.e. it contains at most K elements. Under the sparsity assumption,
the Daubechies db4, Gabor, and mexican hat and spline wavelets. the compression is actually performed by multiplying the acquired
Regarding the sensing matrix, often Gaussian and Bernoulli ran- ECG samples by a M  N sensing matrix U, with M < N: