1
A Compressive Sensing Codec Architecture for
ECG Signals with Adaptive Quantization and
Stream Entropy Coding
Shailesh Kumar, Surendra Prasad, Senior Member, IEEE, and Brejesh Lall, Member, IEEE
Abstract—Remote monitoring of electrocardiogram (ECG)
signals plays a critical role in the management of cardiovascular
diseases. Current long term ECG monitoring systems generate
a large amount of data that need to be sent across a wireless
body area network. The wearable devices are resource limited.
Hence a low resource consuming compression scheme is desirable.
Compressive sensing (CS) is quite appealing as a low complexity
method for compression of ECG data. This paper investigates
the distribution of the compressive measurements and proposes
a CS based codec architecture comprising simple quantization
and asymmetric numeral systems (ANS) based entropy coding
steps that can significantly boost the compression ratio without
sacrificing reconstruction performance. The encoder has low
computational complexity beyond the sensing operation and can
be implemented entirely using integer arithmetic. The quantized
Gaussian entropy model for the compressive measurements is
estimated directly from the data and can be easily adapted
to achieve better compression. Our decoder architecture is
flexible in choosing any suitable reconstruction method. We have
tested our encoder with bound optimization based block sparse
Bayesian learning (BSBL-BO) reconstruction algorithm on MIT-
BIH Arrhythmia database. We have validated that our encoder
can achieve additional 20-30% of space savings without any
impact on reconstruction quality on top of the savings in terms
of measurements. We have also clearly described the digitization
process for the measurements which has often been ignored
in the literature on CS based compression of ECG data. As
part of reproducible research we have open sourced our codec
(implemented in Python with JAX based GPU acceleration for
the decoder).
I. I NTRODUCTION
Noncommunicable diseases (NCDs) account for 72% of all
global deaths with cardiovascular diseases (CVDs) accounting
for 44% of all NCD mortality [1]. For patients with CVDs,
wearable device based remote ECG monitoring plays a critical
role in their disease management and care.
In wireless body area networks (WBAN) based telemoni-
toring networks [2], the energy consumption on sensor nodes
is a primary design constraint [3]. The wearable sensor nodes
are often battery operated. It is necessary to reduce energy
consumption as much as possible.
However, long term ECG monitoring can generate a large
amount of uncompressed data. For example, each half hour
2 lead recording in the MIT-BIH Arrhythmia database [4]
requires 1.9MB of storage. As shown in [5], in a real time
telemonitoring sensor node, the wireless transmission of data
consumes most of the energy. The real time compression of
ECG data by a low complexity encoder has received significant
attention in the past decade.
ECG signal compression has been an active area of interest
for several decades. Extensive surveys can be found in [6],
[7]. Compressive sensing (CS) based techniques for ECG data
compression have been reviewed in [8], [9].
Transform domain techniques (e.g., Discrete Cosine Trans-
form [10], Discrete Cosine Transform [11], [12], Discrete
Wavelet Transform [13]–[16]) are popular in ECG compres-
sion and achieve high compression ratios (CR) at clinically
acceptable quality. However, they require computationally
intensive sparsifying transforms on all data samples and are
thus not suitable for WBAN sensor nodes [8].
Compressive sensing [17]–[21] uses a sub-Nyquist sampling
method by acquiring a small number of incoherent measure-
ments which are sufficient to reconstruct the signal if the signal
is sufficiently sparse in some basis. For a sparse signal x ∈ Rn ,
one would make m linear measurements where m ≪ n which
can be mathematically represented by a sensing operation
y = Φx (1)
where Φ ∈ Rm×n is a matrix representation of the sensing
process and y ∈ Rm the set of m measurements collected for
x. A suitable reconstruction algorithm can recovery x from y.
Ideally, the sensing process should be implemented at the
hardware level in the analog to digital conversion (ADC)
process. However, much of the use of CS in ECG follows a
digital CS paradigm [5] where the ECG samples are acquired
first by the ADC circuit on the device and then they are
translated into incoherent measurements via the multiplication
of a digital sensing matrix. These measurements are then
transmitted to remote telemonitoring servers. A suitable recon-
struction algorithm is used on the server to recover the original
ECG signal from the compressive measurements. Reconstruc-
tion algorithms for ECG signals include: greedy algorithms
[22] (simultaneous orthogonal matching pursuit), optimization
based algorithms [23], [5] (SPG-L1), Bayesian learning based
algorithms [24]–[26], deep learning based algorithms [27].
In order to keep the sensing matrix multiplication simple
and efficient, sparse binary sensing matrices are a popular
choice [5], [24].
A. Problem Formulation
We consider the problem of efficient transmission of com-
pressive measurements of ECG signals over the wireless body
area networks under the digital compressive sensing paradigm.
Let x be an ECG signal and y be the corresponding stream

of compressive measurements. Our goal is to transform y into
a bitstream s with as few bits as possible without losing the
signal reconstruction quality.
We consider whether a digital quantization of the compres-
sive measurements affects the reconstruction quality. Further,
we study the empirical distribution of compressive measure-
ments to explore efficient ways of entropy coding of the
quantized measurements.
A primary constraint in our design is that the encoder
should avoid any floating point arithmetic so that it can be
implemented efficiently in low power devices.
B. Related Work
The literature on the use of CS for ECG compression is
mostly focused on the design of the specific sensing matrix,
sparsifying dictionary, or reconstruction algorithm for the high
quality reconstruction of the ECG signal from the compressive
measurements. To the best of our knowledge, (digital) quan-
tization and entropy coding of the compressive measurements
of ECG data hasn’t received much attention in the past.
Mamaghanian et al. [5] use a Huffman codebook which is
deployed inside the sensor device. They don’t use any quan-
tization of the measurements. However, they don’t provide
much detail on how the codebook was designed or how should
it be adapted for variations in ECG signals. They clearly
define the compression ratio in terms of a ratio between the
uncompressed bits bitsu and the compressed bits bitsc . They
u −bitsc
define it as bitsbits × 100. This is often known in literature
u so that all the operations can be implemented by integer arith-
as percentage space savings.
Simulation based studies often send the real valued com-
pressive measurements to the decoder modules and don’t
consider the issue of number of bits required to encode each
measurement. Zhang et al. [24], Zhang et al. [27] define
n−m
n × 100 as compression ratio. Mangia et al. [28] define
n
m as the compression ratio. Polania et al. [29] define m
as compression ratio. Picariello et al. [30] use a non-random
sensing matrix. They infer a circulant binary sensing matrix
directly from the ECG signal being compressed. The sensing
matrix is adapted as and when there is significant change in
the signal statistics. They represent both the ECG samples
and compressive measurements with same number of bits
per sample/measurement. However, in their case, there is the
additional overhead of sending the sensing matrix updates.
n a short overview on entropy coding.
Their compression ratio is slightly lower than m where they
n
call m as the under-sampling ratio.
Huffman codes have frequently been used in ECG data
compression for non-CS methods. Luo et al. in [31] proposed
a dynamic compression scheme for ECG signals which con-
sisted of a digital integrate and fire sampler followed by an
entropy coder. They used Huffman codes for entropy coding of
the timestamps associated with the impulse train. Chouakri et
al. in [32] compute the DWT of ECG signal by Db6 wavelet,
select coefficients using higher order statistics thresholding,
then perform linear prediction coding and Huffman coding of
the selected coefficients.
Asymmetric Numeral Systems (ANS) [33] based entropy
coding schemes have seen much success in recent years for
2
lossless data compression. zstd by Facebook Inc. [34] and
LZFSE by Apple Inc. are popular lossless data compression
formats based on ANS. To the best of our knowledge, ANS
type stream codes have not been considered in the past for the
entropy coding of compressive measurements.
C. Contributions
In this paper, we present an adaptive quantization and
entropy coding scheme for the digital compressive measure-
ments in the encoders running on the WBAN sensor nodes.
The adaptive quantization scheme has been designed in a
manner so that the quantization noise is kept within a specified
threshold. Once the measurements are quantized and clipped,
they are ready to be entropy coded. A quantized Gaussian
distribution of the measurements is estimated directly from
the data and the parameters for this distribution are used for
the entropy coding of the measurements. Asymmetric numeral
systems (ANS) based entropy coding is then used for efficient
entropy coding of the quantized and clipped measurements.
The ECG signal is split into frames for encoding purposes.
We have designed a bitstream format with a stream header to
describe the encoding parameters and a frame header for each
encoded frame containing frame wise quantization parameters.
We have designed a decoder which accepts this encoded
bitstream and reconstructs the ECG signal frame by frame
in real time.
Our encoding scheme doesn’t require a fixed codebook for
entropy coding. The encoder has been designed in a manner
metic. We have tested our encoder on MIT-BIH Arrhythmia
database and are able to demonstrate additional space savings
beyond the savings by reduced number of measurements. Our
implementation is available as part of reproducible research
on GitHub [35].
n
D. Paper Organization
The rest of the paper is organized as follows. Section II
describes the ECG database used for the evaluation of the
codec and the performance metrics. Section III describes the
proposed codec architecture. The analysis of the experimental
results is covered in section IV. Section V summarizes our
contributions and identifies future work. Appendix A provides
a short overview on compressive sensing. Appendix C provides
II. ECG DATABASE AND P ERFORMANCE M ETRICS
We use the MIT-BIH Arrhythmia Database [4] from Phy-
sioNet [36]. The database contains 48 half hour excerpts of
two channel ambulatory ECG recordings from 47 subjects.
The recordings were digitized at 360 samples per second for
both channels with 11 bit resolution over a 10mV range.
The samples can be read in both digital (integer) form or
as physical values (floating point) via the software provided
by PhysioNet. We use the integer values in our experiments
since our encoder is designed for integer arithmetic. We use
the MLII signal (first channel) from each recording in our
experiments.
 3

A. Performance Metrics go up to[− nd nd

m × 1024, m × 1023]. For a simple case where
In our encoder, ECG signal is split into windows of n n = 2m and d = 4, we will require 14 bits to represent each
samples each which can be multiplied with a sensing matrix. measurement value. To achieve m n as the compression ratio,
Each window of n samples generates m measurements by the we will have to quantize the measurements in 11 bits. If we do
sensing equation y = Φx. Assume that we are encoding s so, we shall need to provide some way to communicate the
ECG samples where s = nw and w is the number of signal quantization parameters to the decoder as well as study the
windows being encoded. Let the ECG signal be sampled by impact of quantization noise. This issue seems to be ignored
the ADC device at a resolution of r bits per sample 1 . Then in [24].
the number of uncompressed bits is given by bitsu = rs. Another way of looking at the compressibility is how many
1) Compression Ratio: Let the total number of compressed bits per sample (bps) are needed on average in the compressed
bits corresponding to the s ECG samples be bitsc 2 . Then the bitstream. We define bps as:
compression ratio (CR) is defined as
bitsc
bitsu bps ≜ . (6)
CR ≜ . (2) s
bitsc
Since the entropy coder is coding the measurements rather
Percentage space saving (PSS) is defined as
than the samples directly, hence it is also useful to see how
bitsu − bitsc many bits are needed to code each measurement. We denote
PSS ≜ × 100. (3)
bitsu this as bits per measurement (bpm):
Note that often in literature, PSS is defined as compression
bitsc
ratio (e.g., [5]). Several papers ignore the bitstream formation bpm ≜ . (7)
aspect and report m n−m mw
n × 100 (e.g., [37]) or n × 100 (e.g.,
[27]) as the compression ratio which measures the reduction in 2) Reconstruction Quality: The normalized root mean
number of measurements compared to the number of samples square error is defined as
in each window. We shall call this metric as percentage
measurement saving (PMS): ∥x − x̃∥2
N RMSE(x, x̃) ≜ (8)
n−m ∥x∥2
PMS ≜ × 100. (4)
n
where x is the original ECG signal and x̃ is the reconstructed
The ratio m/n will be called as the measurement ratio:
signal. A popular metric to measure the quality of reconstruc-
m tion of ECG signals is percentage root mean square difference
MR = . (5)
n (PRD):
The measurement ratio m n is not a good indicator of
compression ratio. If the sensing matrix Φ is Gaussian, then PRD(x, x̃) ≜ N RMSE(x, x̃) × 100 (9)
the measurement values are real valued. In literature using
Gaussian sensing matrices (e.g., [37]), it is unclear how The signal to noise ratio (SNR) is related to PRD as
many bits are being used to represent each floating point
measurement value for transmission. Under standard 32-bit SNR ≜ −20 log10 (0.01PRD). (10)
IEEE floating point format, each value would require 32-bits.
Then for MIT-BIH data the compression ratio in bits would As one desires higher compression ratios and lower PRD, one
11×n
be 32×m . The only way the ratio m n would make sense if can define a combined quality score (QS) as
the measurements are also quantized at 11 bits resolution.
However the impact of such quantization is not considered CR
QS = × 100. (11)
in the simulations. PRD
Now consider the case of a sparse binary sensing matrix.
Since it consists of only zeros and ones, hence for integer Zigel et al. [38] established a link between the diagnostic
inputs, it generates integer outputs. Thus, we can say that distortion and the easy to measure PRD metric. Table I shows
output of a sparse binary sensor are quantized by design. the classified quality and corresponding SNR (signal to noise
However, the range of values changes. Assume that the sensing ratio) and PRD ranges.
matrix has d ones per column. Then it has a total of nd ones.
Thus, each row will have on average nd 3
m ones. If we assume TABLE I
the input data to be in the range of [−1024, 1023] (under 11- Q UALITY OF R ECONSTRUCTION [38]
bit), then in the worst case, the range of output values may
Quality PRD SNR
1 For MIT-BIH Arrhythmia database, r = 11. Very good < 2% > 33 dB
2 This includes the overhead bits required for the stream header and frame Good 2-9% 20-33 dB
headers to be explained later. Undetermined ≥ 9% ≤ 20 dB
3 Since the ones are randomly placed, hence we won’t have same number
of ones in each row.
 4

III. P ROPOSED C ODEC A RCHITECTURE TABLE III

F RAME HEADER
The ECG signal is split into windows of n samples each.
The windows of ECG signal are further grouped into frames Parameter Description Bits
of w windows each. The last frame may have less than w µy Mean value 16
windows. The encoder compresses the ECG signal frame by σy Standard deviation 16
frame. The encoder converts the ECG signal into a bitstream q Frame quantization parameter 3
r Frame range parameter 4
and the decoder reconstructs the ECG signal from the bit- nw Windows in frame 8
stream. nc Words of entropy coded data 16
The encoder first sends the encoder parameters in the form
of a stream header table II. Then for each frame of ECG signal,
Digital Compressive
it sends a frame header table III followed by a frame payload Samples
Windowing
Sensing
Flattening

consisting of the entropy coded measurements for the frame.

The encoding algorithm is presented in algorithm 1. The
Adaptive Adaptive Entropy Compressed
decoding algorithm is presented in algorithm 2. Quantization Clipping Coding Bits

A. Stream Header Fig. 1. Digital Compressive Sensing Encoder

The stream header table II consists of all necessary informa-

tion required to initialize the decoder. In particular, it contains Algorithm 1: Encoder algorithm
the pseudo-random generator key that can be used to reproduce Send stream header ;
the sparse binary sensing matrix used in the encoder by the Build sensing matrix Φ;
decoder, the number of samples per window (n), the number foreach frame of ECG signal as x do
X ← window(x) ;
of measurements per window (m), the number of ones per nw ← number of windows in the frame ;
column in the sensing matrix (d), and the maximum number // Sense
of windows in each frame of ECG signal (w). It contains a Y ← ΦX;
y ← flatten(Y);
flag indicating whether adaptive or fixed quantization will be // Adaptive quantization
used. It contains the limits on the normalized root mean square for q = qmax : qmin : −1 do
error for the adaptive quantization (ρ) and adaptive clipping ȳ ← 21q y ;
ỹ ← 2q ȳ;
(γ) steps. If fixed quantization is used, then it contains the ρ ←;
fixed quantization parameter value (q). Both ρ and γ in the if N RMSE(y, ỹ) ≤ ρ then
stream header are encoded using a simple decimal format break ;
end
a1e−b where a and b are encoded as 4-bit integers. end
Following [5], we construct a sparse binary sensing matrix // Quantized Gaussian model parameters
Φ of shape m × n. Each column of Φ consists of exactly µy ≤ round(mean(ȳ)) ;
σy ≤ ceiling(std(ȳ)) ;
d ones and m − d zeros, where the position of ones has σy ← max(σy , 1);
been randomly selected in each column. Identical algorithm // Adaptive range adjustment
is used to generate the sensing matrix using the stream for r = 2 : 8 do
ymin ← µy − rσy ;
header parameters in both the encoder and the decoder. The ymax ← µy + rσy ;
randomization is seeded with the parameter key sent in the ŷ ← clip(ȳ, ymin , ymax );
stream header. if N RMSE(y, ŷ) ≤ γ then
break ;
end
TABLE II end
S TREAM HEADER c ← ans code(ŷ, µy , σy , ymin , ymax ) ;
nc ← number of words in c;
Parameter Description Bits Send frame header(µy , σy , q, r, nw , nc );
Send frame payload(c);
key PRNG key for Φ 64
end
n Window size 12
m Number of measurements per window 12
d Number of ones per column in sensing matrix 6
w Number of windows per frame 8
adaptive Adaptive or fixed quantization flag 1 The input to the encoder is a frame of digital ECG samples at a
if (adaptive )
ρ N RMSE limit for adaptive quantization 8
given sampling rate fs . In the MIT-BIH database, the samples
else are digitally encoded in 11 bits per sample (unsigned). Let the
q Fixed quantization parameter 4 frame of ECG signal be denoted by x. The frame is split into
γ N RMSE limit for clipping 8
non-overlapping windows of n samples each. We shall denote
each window of n samples by vectors xi .
1) Windowing: A frame of ECG consists of multiple such
B. Encoding windows (up to a maximum of w windows per frame as per
Here we describe the encoding process for each frame. the encoder configuration). Let the windows by denoted by
Figure 1 shows a block diagram of the frame encoding process. x1 , x2 , . . . . Let there be w such windows. We can put them

together to form the (signed) signal matrix 4 :
 
X = x1 x2 . . . xw .
2) Compressive Sensing: The digital compressive sensing
is represented as
yi = Φxi (13)
where yi is the measurement vector for the i-th window.
Combining the measurement vectors for each window, the
measurement matrix for the entire record is given by
Y = ΦX.
Note that by design, the sensing operation can be implemented
using just lookup and integer addition. The ones in each row
of Φ identify the samples within the window to be picked up
and summed. Consequently, Y consists of only signed integer
values.
3) Flattening: Beyond this point, the window structure of
signal is not relevant for quantization and entropy coding
purposes in our design. Hence, we flatten it (column by
column) into a vector y of mw measurements.
y = flatten(Y).
4) Quantization: Next, we perform a simple quantization of
measurement values. If fixed quantization has been specified
in the stream header, then for each entry y in y, we compute
ȳ = y/2q .
For the whole measurement vector, we can write this as
1 integer arithmetic.
ȳ = q y. (17)
2
This can be easily implemented in a computer as a signed right
shift by q bits (for integer values). This reduces the range of
values in y by a factor of 2q .
If adaptive quantization has been specified, then we vary the
quantization parameter q from a value qmax = 6 down to a
value qmin = 0. For each value of q, we compute ȳ = 21q y. We
then compute ỹ = 2q ȳ. We stop if N RMSE(y, ỹ) reaches a
limit specified by the parameter ρ in the stream header.
5) Entropy Model: Before we explain the clipping step, we
shall describe our entropy model. We model the measurements
as samples from a quantized Gaussian distribution which can
only take integral values. First, we estimate the mean µy and
standard deviation σy of measurement values in ȳ. We shall
need to transmit µy and σy in the frame header. We round the
values of µy and σy to the nearest integer so that it can be
transmitted efficiently as integers.
Entropy coding works with a finite alphabet. Accordingly,
the quantized Gaussian model requires specification of the
minimum and maximum values that our quantized measure-
ments can take. The range of values in ȳ must be clipped to
this range.
4 PhysioNet provides the baseline values for each channel in their ECG
records. Since the digital samples are unsigned, we have subtracted them by
the baseline value (1024 for 11-bit encoding). 11 bits mean that unsigned
values range from 0 to 2047. The baseline for zero amplitude is digitally
represented as 1024. After baseline adjustment, the range of values becomes
[−1024, 1023].
5
6) Adaptive Clipping: We clip the values in ȳ to the range
[µy − rσy , µy + rσy ] where r is the range parameter estimate
(12)
from the data. This is denoted as
ŷ = clip(ȳ, µy − rσy , µy + rσy ). (18)
Similar to adaptive quantization, we vary r from 2 to 8 till we
have captured sufficient variation in ȳ and N RMSE(y, ŷ) ≤
γ.
7) Entropy Coding: We then model the measurement values
in ŷ as a quantized Gaussian distributed random variable with
(14) mean µy , standard deviation σy , minimum value µy −rσy and
maximum value µy + rσy . We use the ANS entropy coder to
encode ŷ into an array c of 32-bit integers (called words).
This becomes the payload of the frame to be sent to the
decoder. The total number of compressed bits in the frame
payload is the length of the array nc times 32. Note that we
have encoded and transmitted ŷ and not the unclipped ȳ. ANS
entropy coding is a lossless encoding scheme. Hence, ŷ will be
reproduced faithfully in the decoder if there are no bit errors
involved in the transmission 5 .
8) Integer Arithmetic: The input to digital compressive
(15) sensing is a stream of integers. The sensing process with
sparse binary sensing matrix can be implemented using integer
sums and lookup. It is possible to implement the computation
of approximate mean and standard deviation using integer
arithmetic. We can use the normalized mean square error based
thresholds for adaptive quantization and clipping steps. ANS
(16)
entropy coding is fully implemented using integer arithmetic.
Thus, the proposed encoder can be fully implemented using
C. Decoding
The decoder initializes itself by reading the stream header
and creating the sensing matrix to be used for the decoding
of compressive measurements frame by frame.
Compressed
Bits
Entropy
Decoding
Inverse
Quantization
Windowing
Reconstruction Decoded
Algorithm
Flattening
Samples
Fig. 2. Digital Compressive Sensing Decoder
Here we describe the decoding process for each frame.
Figure 2 shows a block diagram for the decoding process.
Decoding of a frame starts by reading the frame header which
provides the frame encoding parameters: µy , σy , q, r, nw , nc .
1) Entropy Decoding: The frame header is used for build-
ing the quantized Gaussian distribution model for the decoding
of the entropy coded measurements from the frame payload.
The minimum and maximum values for the model are com-
puted as:
ymin , ymax = µy − rσy , µy + rσy . (19)
nc tells us the number of words (4nc bytes) to be read from
the bitstream for the frame payload. The ANS decoder is used
5 We assume that appropriate channel coding mechanism has been used.

Algorithm 2: Decoder algorithm
Read stream header ;
Build sensing matrix Φ;
while there is more data do
µy , σy , q, r, nw , nc ← read frame header ;
c ← read frame payload (nc );
// Entropy model parameters
ymin ← µy − rσy ;
ymax ← µy + rσy ;
ŷ ← ans decode(c, µy , σy , ymin , ymax ) ;
// Inverse quantization
ỹ ← 2q ŷ;
Ỹ ← window(ỹ) ;
X̃ ← reconstruct(Ỹ) ;
x̃ ← flatten(X̃) ;
end
to extract the encoded measurement values ŷ from the frame
payload.
2) Inverse Quantization: We then perform the inverse quan-
tization as
ỹ = 2q ŷ. (20)
Next, we split the measurements into measurement windows
of size m corresponding to the signal windows of size n.
Ỹ = window(ỹ). (21)
We are now ready for the reconstruction of the ECG signal
for each window.
3) Reconstruction: The architecture is flexible in terms of
choice of the reconstruction algorithm.
X̃ = reconstruct(Ỹ). (22)
Each column (window) in Ỹ is decoded independently. In our
experiments, we built a BSBL-BO (Block Sparse Bayesian
Learning-Bound Optimization) decoder [24], [26], [37]. Our
implementation of BSBL is available as part of CR-Sparse
library [39]. As Zhang et al. suggest in [24], block sizes
are user defined and are identical and no pruning of blocks
is applied. Our implementation has been done under these
assumptions and is built using JAX so that it can be run
on GPU hardware easily to speed up decoding. The only
configurable parameter for this decoder is the block size which
we shall denote by b in the following. Once the samples
have been decoded, we flatten (column by column) them to
generate the sequence of decoded ECG samples to form the
ECG record.
x̃ = flatten(X̃). (23)
4) Alternate Reconstruction Algorithms: It is entirely pos-
sible to use a deep learning based reconstruction network
like CS-NET [27] in the decoder. We will need to train
the network with ỹ as inputs and x as expected output.
However, we haven’t pursued this direction yet as our focus
was to study the quantization and entropy coding steps in this
work. One concern with deep learning based architectures is
that the decoder network needs to be trained separately for
each encoder configuration and each database. The ability to
dynamically change the encoder/decoder configuration param-
eters is severely restricted in deep learning based architectures.
This limitation doesn’t exist with BSBL type algorithms.
6
D. Discussion
1) Measurement statistics: Several aspects of our encoder
architecture are based on the statistics of the measurements
y. We collected the summary statistics including mean µ,
standard deviation σ, range of values in y, skew and kurtosis
for the measurement values for each of the ECG records in
the MIT-BIH database. These values have been reported for
one particular encoder configuration in table V. In addition,
we also compute the range divided by standard deviation rng σ
and the iqr (inter quantile range) divided by standard deviation
iqr
σ .
2) Gaussianity: The key idea behind the our entropy coding
design is to model the measurement values as being sampled
from a quantized Gaussian distribution. Towards this, we
measured the skew and kurtosis for the measurements for each
record as shown in table V for the sensing matrix configuration
of m = 256, n = 512, d = 4. For a Gaussian distributed
variable, the skew should be close to 0 while kurtosis should
be close to 3. While the skew is not very high, Kurtosis does
vary widely. Figure 3 shows the histograms of measurement
values for 6 different records. Although the measurements are
not Gaussian distributed, it is not a bad approximation. The
quantized Gaussian approximation works well in the entropy
coding. It is easy to estimate from the data.
The best compression can be achieved by using the empir-
ical probabilities of different values in y in entropy coding.
However, doing so would require us to transmit the empirical
probabilities as side information. This may be expensive. We
can estimate the improvement in compression overhead by use
of the quantized Gaussian approximation. Let P denote the
empirical probability distribution of data and let Q denote the
corresponding quantized Gaussian distribution. Bamler in [40]
show empirically that the overhead of using an approximation
distribution Q in place of P in ANS entropy coding is close
to the KL divergence KL(P||Q) which is given by
 
X P(y)
KL(P||Q) = P(y) log2 . (24)
Q(y)
y
We computed the empirical distribution for y for each record
and measured its KL divergence with the corresponding quan-
tized Gaussian distribution. It is tabulated in the kld column in
table V. It varies around 0.11±0.07 bits across the 48 records.
Thus, the overhead of using a quantized Gaussian distribution
in place of the empirical probabilities can be estimated to be
4 − 18%. One can see that the divergence tends to increase as
the kurtosis increases. We determined the Pearson correlation
coefficient between kurtosis and kld to be 0.67.
Figure 4 shows an example where the empirical distribution
is significantly different from the corresponding quantized
Gaussian distribution due to the presence of a large number of
0 values in y. Note that this is different from the histograms
in fig. 3 where the y values have been binned into 200 bins.
Also, fig. 3 suggests that the empirical distributions vary
widely from one record to another in the database. Hence
using a single fixed empirical distribution (e.g. the Huffman
code book preloaded into the device in [5]) may lead to lower
compression.

(a) Record 100
(d) Record 202
Fig. 3. Histograms of measurement values with the sparse binary sensing matrix with m = 256, n = 512, d = 4 in 200 bins
Fig. 4. Empirical and quantized Gaussian distributions for measurement
values y in record 234
3) Clipping: An entropy coder can handle with a finite set
of symbols only. Hence, the range of input values [measure-
ments coded as integers] must be restricted to a finite range.
This is the reason one has to choose a distribution with finite
support (like quantized Gaussian). From table V one can see
that while the complete range of measurement values can be
up to 40-50x larger than the standard deviation, the iqr is
less than 1.5 times the standard deviation. In other words, the
measurement values are fairly concentrated around its mean.
This can be visually seen from the histograms in Figure 3 also.
4) Quantization: Figure 5 demonstrates the impact of the
quantization step on the reconstruction quality for record 100
under non-adaptive quantization. 6 windows of 512 samples
were used in this example. The first panel shows the original
(digital) signal. Remaining panels show the reconstruction at
different quantization parameters. The reconstruction visual
quality is excellent up to q = 5 (PRD below 7%), good at
q = 6 (PRD at 9%) and clinically unacceptable at q = 7
(with PRD more than 15%). One can see significant waveform
distortions at q = 7. Also note how the quality score keeps
7
(b) Record 102 (c) Record 115
(e) Record 208 (f) Record 234
increasing till q = 4 and starts decreasing after that with a
massive drop at q = 7.
Fig. 5. Reconstruction of a small segment of record 100 for different values
of q = 0, 2, 4, 5, 6, 7 with m = 256, n = 512, d = 4 under non-adaptive
quantization. Block size for the BSBL-BO decoder is 32.
IV. R ESULTS
A. Impact of quantization parameter
In fig. 6, we study the impact of the quantization parameter
q on different compression statistics at different values of
measurement ratio mr = m n for record 100. The number
of measurements (m) have been chosen to vary from 20%
to 60% of the window size (n) in different configurations.
The quantization parameter varies from 0 to 7. Nonadaptive
quantization was used for this experiment.
(a) shows that reconstruction quality doesn’t change much
from q = 0 till q = 4 after which it starts degrading. As
expected, the SNR/PRD degrades as m is reduced. At 30%
and below, the PRD is above 9% which is unacceptable.
 8

(b) shows that PSS (percentage space saving) increases
linearly with q. As q increases, the size of the alphabet for
entropy coding reduces and this leads to increased space
savings.
(c) shows the variation of the quality score with q and we
can clearly see that quality score generally increases till q = 4
after which it starts decreasing.
The key result is shown in panel (d) which shows the
variation of PSS with PMS at different values of q. PSS
increases linearly with PMS. Also, PSS is much higher than
PMS at higher quantization levels.
We have also listed the numerical values in table IV. The
first column for q = 0 depicts the case where no quantization
has been applied, we are only doing clipping and entropy
coding. Since PSS is consistently higher than PMS, it is clear
that we need less than 11 bits per sample on an average
after entropy coding of the measurement values. Increasing
quantization linearly increases the PSS. Since up to q = 4,
there is no noticeable impact on PRD, as seen in panel (a), it
is safe to enjoy these savings in bit rate. At 40% PMS, one
can attain up to 69.3% PSS without losing any reconstruction
quality while one can push further up to q = 6 to the PSS
of 79.9% with some degradation in quality while still staying
under acceptable PRD.
TABLE IV
PSS AT DIFFERENT VALUES OF q AND PMS FOR RECORD 100
q 0 1 2 3 4 5 6
PMS
V. C ONCLUSION AND F UTURE W ORK
A. Contributions
The key contributions of our encoder architecture are the
following:
• We have presented a digital compressive sensing based
encoder design including adaptive digital quantization
and entropy coding which can be implemented entirely
using integer arithmetic.
• We have demonstrated that ANS based arithmetic coding
can lead to significant space savings.
• We have demonstrated that we don’t need a hard-coded
code-book for entropy coding. We can simply model the
measurement values distributed as quantized Gaussian
values for effective entropy coding.
• We use simple mean and standard deviation as the entropy
model parameters which can be dynamically adapted. The
overhead of the side information is negligible.
• We have demonstrated up to 16% of additional space sav-
ings contributed by the quantization and entropy coding
steps.
• We have described a complete bitstream specification for
carriage of all the side information needed to decode the
entropy coded compressive measurements.
The software code implementing this codec and all scripts
for the experimental studies conducted in this work have been
released as open source software on GitHub [35].
7

40.0 47.5 53.0 58.4 63.8 69.3 74.6 79.9 84.8

B. Future Work
50.0 55.7 60.3 64.8 69.3 73.8 78.3 82.6 87.0
55.0 59.8 63.9 67.9 72.0 76.1 80.1 84.0 88.0 The reconstruction algorithm in the decoder is replaceable.
60.0 63.6 67.2 70.8 74.5 78.1 81.7 85.2 88.6
65.0 68.0 71.1 74.3 77.5 80.6 83.8 86.8 89.7 Deep learning based architectures can provide further boost
70.0 71.8 74.6 77.3 80.0 82.7 85.5 88.1 90.7
75.0 76.5 78.7 81.0 83.2 85.5 87.7 90.0 92.1 to the quality of reconstruction. The impact of quantization
80.0 81.0 82.8 84.6 86.4 88.2 90.0 91.8 93.6
and clipping on deep learning based reconstruction needs to
be further studied.
The quantization, clipping and entropy coding blocks in our
codec design are general purpose and should be useful in other
B. Compression across different records CS based applications (e.g., EEG, EMG, FECG etc.).
During the course of long term ECG monitoring, one may
Table VI shows the compression statistics for the encoder
need to change the encoder parameters (m, n, d) to respond
configuration of m = 256, n = 512, d = 4 and the decoder
to changing characteristics of the ECG data. Enhancements in
configuration of b = 32 (block size). The statistics are
the bitstream specification can be done accordingly.
further summarized in table VII. The column OH reports
the overhead in percentage of the bits needed to encode the One advantage of using fixed length coding for measure-
stream header and frame headers. The column Q SNR reports ments is random access to any part of the bitstream. This is
the quantization noise introduced due to the quantization and not straightforward in stream codes. ANS however makes it
clipping steps in the encoder. We can see that this configuration easy to build a seakable decoder by introducing checkpoints
achieves PSS between 66.05±2.78%. We achieve an additional [40]. Such checkpoints may be communicated as in the frame
16% of space savings over and above the PMS (50%). This headers.
additional space savings is the contribution due to adaptive
quantization and entropy coding. The PRD is 5.6 ± 1.8%. Out A PPENDIX
of the 48 records, only 4 records have a PRD greater than 9%.
A. Compressive Sensing
Maximum PRD is 10.2%. The quality score is 57.8 ± 18.6.
The header bits overhead is 0.21 ± 0.01% which is negligible. Under the digital CS paradigm, we assume that the signals
This configuration achieves of 3.7 ± 0.2 bits per sample. have passed through the analog to digital converter (ADC) and
Quantization SNR is 37.3 ± 0.9 dB. Our adaptive quantization are represented using signed integers at a specific resolution.
algorithm ensures that the quantization SNR doesn’t vary CS is an emerging signal compression paradigm that relies
much. on the sparsity of signals (in an appropriate basis) to compress
 9

25.0
MR
0.2
22.5 90 70 0.25
0.3
20.0 0.35
0.4
80 60 0.45
17.5 0.5
0.6
15.0

PSS (%)
PRD %

70 50

QS
12.5
MR MR
0.2 0.2
10.0 0.25 0.25 40
0.3 60 0.3
7.5 0.35 0.35
0.4 0.4
0.45 0.45 30
5.0 0.5 50 0.5
0.6 0.6
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
q q q

(a) q vs PRD (b) q vs PSS (c) q vs Quality Score

6
MR MR
0.2 10 0.2
90 0.25 0.25
5 0.3 0.3
0.35 9 0.35
0.4 0.4
80 0.45 0.45
0.5 8 0.5
4
0.6 0.6
7
PSS (%)

bpm
bps
70
3
q 6
0
60 1 5
2 2
3
4 4
5
50 6 1
7 3

40 45 50 55 60 65 70 75 80 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
PMS (%) q q

(d) PMS vs PSS (e) q vs bps (f) q vs bpm

m
Fig. 6. Variation of compression statistics vs quantization parameter at different measurement ratios n
for record 100

TABLE V TABLE VI
M EASUREMENT STATISTICS AT m = 256, n = 512, d = 4 C OMPRESSION STATISTICS AT m = 256, n = 512, d = 4

rng iqr record CR PSS bpm bps OH Q SNR SNR PRD QS

record µy σy iqr rng skew kurtosis kld
σ σ
100 3.3 69.3 6.7 3.4 0.2 36.1 25.9 5.1 64.2
100 -490.0 224.4 293.0 2562 -0.5 3.7 0.049 11.4 1.3
101 3.1 68.2 7.0 3.5 0.2 37.0 25.9 5.0 62.4
101 -454.8 287.1 323.0 9647 -0.4 15.4 0.13 33.6 1.1
102 3.2 68.6 6.9 3.5 0.2 35.9 24.4 6.0 52.8
102 -393.4 197.0 257.2 2035 -0.6 3.7 0.053 10.3 1.3
103 3.0 66.7 7.3 3.7 0.2 38.2 26.1 5.0 60.5
103 -370.3 285.9 302.0 8824 -0.8 15.9 0.16 30.9 1.1
104 3.0 67.0 7.3 3.6 0.2 36.9 20.4 9.6 31.6
104 -360.1 231.5 284.0 4204 -0.6 5.0 0.071 18.2 1.2
105 3.0 66.3 7.4 3.7 0.2 36.5 24.8 5.8 51.5
105 -360.2 346.7 326.0 13704 -0.4 22.5 0.24 39.5 0.9
106 2.8 64.7 7.8 3.9 0.2 37.8 24.4 6.0 47.1
106 -284.7 270.6 314.0 4135 -0.1 4.4 0.046 15.3 1.2
107 2.7 62.5 8.3 4.1 0.2 38.5 26.8 4.6 58.2
107 -372.2 624.6 747.0 9639 -0.5 4.6 0.062 15.4 1.2
108 2.9 66.0 7.5 3.7 0.2 37.6 24.4 6.0 48.8
108 -365.9 422.3 336.0 12986 -0.2 22.1 0.36 30.7 0.8
109 3.0 66.3 7.4 3.7 0.2 37.5 27.3 4.3 68.6
109 -368.5 335.5 389.0 7487 0.2 7.3 0.064 22.3 1.2
111 2.8 64.9 7.7 3.9 0.2 38.3 24.6 5.9 48.0
111 -262.1 307.7 306.0 11998 -0.9 23.1 0.2 39.0 1.0
112 3.2 68.3 7.0 3.5 0.2 35.9 28.7 3.7 85.6
112 -1315.8 538.9 724.0 7839 -0.6 4.1 0.088 14.5 1.3
113 2.8 63.9 7.9 4.0 0.2 38.0 26.0 5.0 55.0
113 -248.3 336.7 379.0 6121 -0.1 5.2 0.064 18.2 1.1
114 3.0 66.2 7.4 3.7 0.2 37.6 22.5 7.5 39.6
114 -248.7 219.4 235.0 8756 1.2 30.4 0.14 39.9 1.1
115 3.1 67.4 7.2 3.6 0.2 36.3 27.5 4.2 72.6
115 -780.8 461.1 559.0 9715 -0.8 6.7 0.094 21.1 1.2
116 3.0 66.5 7.4 3.7 0.2 36.2 27.6 4.2 71.6
116 -1498.1 874.9 930.0 35083 -0.3 22.7 0.22 40.1 1.1
117 3.2 68.3 7.0 3.5 0.2 35.6 28.8 3.6 87.0
117 -1363.0 560.1 748.0 9179 -0.7 4.4 0.098 16.4 1.3
118 3.1 67.4 7.2 3.6 0.2 38.3 27.1 4.4 69.4
118 -1373.1 613.9 809.0 10935 -0.5 4.4 0.1 17.8 1.3
119 3.0 67.1 7.2 3.6 0.2 37.5 29.0 3.6 85.6
119 -1378.1 629.0 825.0 8363 -0.6 3.8 0.074 13.3 1.3
121 3.1 67.6 7.1 3.6 0.2 37.1 30.7 2.9 106.5
121 -1295.5 635.2 763.0 17012 -1.4 13.1 0.16 26.8 1.2
122 3.1 68.0 7.0 3.5 0.2 35.6 27.9 4.0 78.0
122 -1349.7 555.2 747.0 7870 -0.6 3.7 0.079 14.2 1.3
123 3.2 68.6 6.9 3.5 0.2 35.6 28.3 3.8 82.8
123 -1274.2 514.1 699.0 6603 -0.6 3.8 0.084 12.8 1.4
124 3.0 67.2 7.2 3.6 0.2 37.3 30.0 3.2 96.7
124 -1293.1 678.9 829.0 11886 -0.6 5.4 0.1 17.5 1.2
200 2.8 63.7 8.0 4.0 0.2 37.5 22.2 7.8 35.3
200 -168.8 261.8 288.0 6285 -0.3 7.7 0.08 24.0 1.1
201 3.0 67.0 7.3 3.6 0.2 38.1 26.2 4.9 61.8
201 -255.9 189.5 196.0 9986 0.7 55.1 0.16 52.7 1.0
202 2.9 65.1 7.7 3.8 0.2 38.4 25.8 5.1 55.9
202 -270.8 262.1 274.0 11421 -1.0 33.0 0.16 43.6 1.0
203 2.8 64.0 7.9 4.0 0.2 38.2 22.7 7.3 38.0
203 -271.1 471.7 455.0 13371 -0.1 15.1 0.2 28.3 1.0
205 3.2 69.1 6.8 3.4 0.2 36.2 25.8 5.1 62.9
205 -489.1 233.6 301.0 3257 -0.6 4.1 0.059 13.9 1.3
207 2.8 64.6 7.8 3.9 0.2 37.4 25.5 5.3 53.4
207 -273.5 339.1 344.0 11400 -1.2 18.1 0.22 33.6 1.0
208 2.8 63.9 8.0 4.0 0.2 37.9 24.7 5.8 47.7
208 -264.6 427.3 410.0 12449 0.4 16.5 0.18 29.1 1.0
209 2.8 64.5 7.8 3.9 0.2 38.4 20.3 9.7 29.0
209 -263.6 249.6 292.0 4731 -0.5 4.9 0.065 19.0 1.2
210 2.9 65.3 7.6 3.8 0.2 38.3 24.4 6.1 47.7
210 -250.9 228.3 249.0 7909 0.5 17.2 0.087 34.6 1.1
212 2.8 64.3 7.9 3.9 0.2 37.8 20.1 9.9 28.2
212 -249.7 296.9 336.0 7232 -0.5 7.3 0.088 24.4 1.1
213 2.7 63.6 8.0 4.0 0.2 38.6 22.7 7.3 37.6
213 -353.5 523.1 635.0 6528 -0.2 4.0 0.039 12.5 1.2
214 2.8 64.0 7.9 4.0 0.2 37.6 26.7 4.6 60.0
214 -257.8 315.3 369.0 4248 0.2 4.1 0.04 13.5 1.2
215 2.9 65.3 7.6 3.8 0.2 37.5 19.8 10.2 28.2
215 -243.0 208.7 256.0 3544 -0.3 4.2 0.039 17.0 1.2
217 2.8 64.3 7.8 3.9 0.2 37.1 27.0 4.5 62.5
217 -264.2 458.7 528.0 11928 -0.3 8.5 0.089 26.0 1.2
219 3.0 66.7 7.3 3.7 0.2 36.9 29.0 3.6 84.4
219 -938.8 687.6 811.0 10727 -0.4 4.6 0.074 15.6 1.2
220 3.3 69.6 6.7 3.3 0.2 35.1 26.6 4.7 69.9
220 -864.8 378.5 503.0 4731 -0.5 3.7 0.059 12.5 1.3
221 2.9 65.1 7.7 3.8 0.2 38.9 24.9 5.7 50.5
221 -262.4 228.0 270.0 3000 -0.1 4.1 0.036 13.2 1.2
222 2.9 65.6 7.6 3.8 0.2 37.6 22.0 7.9 36.8
222 -250.4 212.1 252.0 4327 -0.8 5.7 0.085 20.4 1.2
223 3.1 68.0 7.0 3.5 0.2 36.3 27.9 4.0 77.8
223 -818.4 419.1 529.0 7449 -0.6 5.0 0.068 17.8 1.3
228 2.8 64.5 7.8 3.9 0.2 36.9 22.9 7.1 39.4
228 -221.5 356.9 294.0 11962 -1.0 24.0 0.35 33.5 0.8
230 2.8 64.6 7.8 3.9 0.2 38.2 24.7 5.9 48.2
230 -271.8 285.3 308.0 7044 -0.4 8.6 0.1 24.7 1.1
231 2.9 65.6 7.6 3.8 0.2 38.2 25.5 5.3 54.8
231 -253.1 228.9 278.0 3250 -0.3 4.2 0.042 14.2 1.2
232 3.0 66.8 7.3 3.7 0.2 37.9 22.8 7.3 41.4
232 -242.2 191.9 243.0 2951 -0.4 4.0 0.039 15.4 1.3
233 2.8 63.8 8.0 4.0 0.2 37.9 24.9 5.7 48.8
233 -245.9 427.2 480.0 9636 -0.1 6.5 0.083 22.6 1.1
234 2.8 64.6 7.8 3.9 0.2 37.3 25.2 5.5 51.6
234 -257.2 345.1 316.0 12002 -0.7 19.8 0.25 34.8 0.9

using incoherent measurements. The basic model can be linear measurements. Each row of Φ represents one linear
expressed as: measurement as a linear functional on x. y is a measurement
y = Φx + e (25) vector consisting of m distinct measurements done on x. By
design Φ is a full rank matrix. Hence every set of m columns
where x ∈ Rn is a signal to compress (of length n). of Φ is linearly independent. e ∈ Rm is the error/noise
Φ ∈ Rm×n is a sensing matrix which compresses x via introduced during the measurement process. In our digital CS
 10

TABLE VII to provide the value of noise variance as input. It is able to

S UMMARY COMPRESSION STATISTICS AT m = 256, n = 512, d = 4 estimate λ within a algorithm.
CR PSS bpm bps OH Q SNR SNR PRD QS
The estimate of x under Bayesian learning framework is
mean 3.0 66.0 7.5 3.7 0.2 37.3 25.4 5.6 57.8
given by the posterior mean of x given the measurements y.
std 0.2 1.8 0.4 0.2 0.0 0.9 2.6 1.8 18.6
min 2.7 62.5 6.7 3.3 0.2 35.1 19.8 2.9 28.2
50% 2.9 66.1 7.5 3.7 0.2 37.5 25.7 5.2 54.9
max 3.3 69.6 8.3 4.1 0.2 38.9 30.7 10.2 106.5 C. Entropy Coding
From the perspective of entropy coding, a symbol is a unit
of information to be transmitted. For us, each value in a
paradigm, the noise is introduced by the quantization step in
measurement vector is a symbol. An alphabet is the set of
our encoder. In our case x is a window of a raw ECG record
all symbols that can be transmitted. For us, it is the set of
from one channel/lead. Often x is not sparse by itself, but is
integral values that the entries in a measurement vector can
sparse in some orthonormal basis Ψ expressed as x = Φα and
take. A message is a sequence of symbols to be transmitted.
the representation α is sparse. ECG signals exhibit sparsity in
For us, it is the sequence of all measurement values. There
wavelet bases.
are two primary types of coding schemes: symbol codes and
Most natural signals have richer structures beyond sparsity. stream codes. A symbol code (e.g., Huffman code) assigns
A common structure is natural signals is a block/group struc- a unique bit string to each symbol. A stream code maps a
ture [41]. We introduce the block/group structure on x as message (a sequence of symbols) to a (large) integer. Symbol
codes approach optimal compression only when the Shannon


x = x1 x2 . . . xg (26)
information content of each symbol happens to be close to
where each xi is a block of b values. The signal x consists of powers of 2. In particular, Huffman coding is not optimal since
g such blocks/groups. Under the block sparsity model, only a it encodes each symbol using an integral number of bits. It is
few k ≪ g blocks are nonzero (active) in the signal x however, not capable of encoding symbols using fractional number of
the locations of these blocks are unknown. We can rewrite the bits.
sensing equation as: In contrast, stream codes can achieve near optimal com-
g
X pression for any probability distribution. Arithmetic Coding
y= Φ i xi + e (27) (AC) [42], [43] encodes the entire message of symbols into
i=1 a single (large) integer enabling fractional bits for different
by splitting the sensing matrix into blocks of columns appro- symbols in a message. A recent family of entropy coders called
priately. Asymmetric Numeral Systems (ANS) [33] allows for faster
implementation thanks to directly operating on a single natural
number representing the current information. It combines the
B. Block Sparse Bayesian Learning
speed of Huffman coding with compression rate of arithmetic
Under the sparse Bayesian framework, each block is as- coding. The constriction library [40] provides a nice way to
sumed to satisfy a parametrized multivariate Gaussian distri- model each symbol in a message as a sample from a given
bution: probability distribution and then encode it using ANS.
P(xi ; γi , Bi ) = N (0, γi Bi ), ∀ i = 1, . . . , g. (28)
R EFERENCES
The covariance matrix Bi captures the intra block correlations.
[1] T. Collins, B. Mikkelsen, and S. Axelrod, “Interact, engage or partner?
We further assume that the blocks are mutually uncorrelated. working with the private sector for the prevention and control of non-
The prior of x can then be written as communicable diseases,” Cardiovascular diagnosis and therapy, vol. 9,
no. 2, p. 158, 2019.
P(x; {γi , Bi }i ) = N (0, Σ0 ) (29) [2] H. Cao, V. Leung, C. Chow, and H. Chan, “Enabling technologies for
wireless body area networks: A survey and outlook,” IEEE Communi-
where cations Magazine, vol. 47, no. 12, pp. 84–93, 2009.
[3] A. Milenković, C. Otto, and E. Jovanov, “Wireless sensor networks for
Σ0 = diag{γ1 B1 , . . . , γg Bg }. (30) personal health monitoring: Issues and an implementation,” Computer
communications, vol. 29, no. 13-14, pp. 2521–2533, 2006.
We also model the correlation among the values within each [4] G. B. Moody and R. G. Mark, “The impact of the mit-bih arrhyth-
active block as an AR-1 process. Under this assumption the mia database,” IEEE Engineering in Medicine and Biology Magazine,
matrices Bi take the form of a Toeplitz matrix vol. 20, no. 3, pp. 45–50, 2001.
[5] H. Mamaghanian, N. Khaled, D. Atienza, and P. Vandergheynst, “Com-
. . . rb−1 pressed sensing for real-time energy-efficient ecg compression on wire-
 
1 r
less body sensor nodes,” IEEE Transactions on Biomedical Engineering,
 r 1 . . . rb−2  vol. 58, no. 9, pp. 2456–2466, 2011.
B= . (31)
 
.. ..  [6] B. Singh, A. Kaur, and J. Singh, “A review of ecg data compression
 .. . .  techniques,” International journal of computer applications, vol. 116,
rb−1 rb−2 ... 1 no. 11, 2015.
[7] S. O. Rajankar and S. N. Talbar, “An electrocardiogram signal compres-
where r is the AR-1 model coefficient. This constraint signif- sion techniques: a comprehensive review,” Analog Integrated Circuits
icantly reduces the model parameters to be learned. and Signal Processing, vol. 98, no. 1, pp. 59–74, 2019.
[8] D. Craven, B. McGinley, L. Kilmartin, M. Glavin, and E. Jones,
Measurement error is modeled as independent zero mean “Compressed sensing for bioelectric signals: A review,” IEEE journal
Gaussian noise P(e; λ) ∼ N (0, λI). BSBL doesn’t require us of biomedical and health informatics, vol. 19, no. 2, pp. 529–540, 2014.
 11

[9] S. S. Kumar and P. Ramachandran, “Review on compressive sensing [32] S. A. Chouakri, O. Djaafri, and A. Taleb-Ahmed, “Wavelet transform
algorithms for ecg signal for iot based deep learning framework,” and huffman coding based electrocardiogram compression algorithm:
Applied Sciences, vol. 12, no. 16, p. 8368, 2022. Application to telecardiology,” in Journal of Physics: Conference Series,
[10] H. Al-Nashash, “A dynamic fourier series for the compression of ecg vol. 454, no. 1. IOP Publishing, 2013, p. 012086.
using fft and adaptive coefficient estimation,” Medical engineering & [33] J. Duda, “Asymmetric numeral systems: entropy coding combining
physics, vol. 17, no. 3, pp. 197–203, 1995. speed of huffman coding with compression rate of arithmetic coding,”
[11] L. V. Batista, E. U. K. Melcher, and L. C. Carvalho, “Compression arXiv preprint arXiv:1311.2540, 2013.
of ecg signals by optimized quantization of discrete cosine transform [34] Collet, Yann, “Zstandard,” 2022. [Online]. Available: https://facebook.
coefficients,” Medical engineering & physics, vol. 23, no. 2, pp. 127– github.io/zstd
134, 2001. [35] Kumar, Shailesh, “Compressive sensing based codec for ecg,” 2022.
[12] A. Bendifallah, R. Benzid, and M. Boulemden, “Improved ecg com- [Online]. Available: https://github.com/shailesh1729/ecg
pression method using discrete cosine transform,” Electronics letters, [36] A. L. Goldberger, L. A. Amaral, L. Glass, J. M. Hausdorff, P. C.
vol. 47, no. 2, p. 1, 2011. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, and H. E.
[13] A. Djohan, T. Q. Nguyen, and W. J. Tompkins, “Ecg compression Stanley, “Physiobank, physiotoolkit, and physionet: components of a
using discrete symmetric wavelet transform,” in Proceedings of 17th new research resource for complex physiologic signals,” circulation, vol.
International Conference of the Engineering in Medicine and Biology 101, no. 23, pp. e215–e220, 2000.
Society, vol. 1. IEEE, 1995, pp. 167–168. [37] Z. Zhang, S. Wei, D. Wei, L. Li, F. Liu, and C. Liu, “Comparison
[14] Z. Lu, D. Y. Kim, and W. A. Pearlman, “Wavelet compression of ecg of four recovery algorithms used in compressed sensing for ecg signal
signals by the set partitioning in hierarchical trees algorithm,” IEEE processing,” in 2016 Computing in Cardiology Conference (CinC).
transactions on Biomedical Engineering, vol. 47, no. 7, pp. 849–856, IEEE, 2016, pp. 401–404.
2000. [38] Y. Zigel, A. Cohen, and A. Katz, “The weighted diagnostic distortion
[15] M. Pooyan, A. Taheri, M. Moazami-Goudarzi, and I. Saboori, “Wavelet (wdd) measure for ecg signal compression,” IEEE transactions on
compression of ecg signals using spiht algorithm,” International Journal biomedical engineering, vol. 47, no. 11, pp. 1422–1430, 2000.
of signal processing, vol. 1, no. 3, p. 4, 2004. [39] S. Kumar, “Cr-sparse: Hardware accelerated functional algorithms for
[16] B. S. Kim, S. K. Yoo, and M.-H. Lee, “Wavelet-based low-delay sparse signal processing in python using jax,” Journal of Open Source
ecg compression algorithm for continuous ecg transmission,” IEEE Software, vol. 6, no. 68, p. 3917, 2021.
Transactions on Information Technology in Biomedicine, vol. 10, no. 1, [40] R. Bamler, “Understanding entropy coding with asymmetric nu-
pp. 77–83, 2006. meral systems (ans): a statistician’s perspective,” arXiv preprint
[17] D. L. Donoho, “Compressed sensing,” Information Theory, IEEE Trans- arXiv:2201.01741, 2022.
actions on, vol. 52, no. 4, pp. 1289–1306, 2006. [41] Y. C. Eldar, P. Kuppinger, and H. Bolcskei, “Block-sparse signals:
[18] R. G. Baraniuk, “Compressive sensing [lecture notes],” Signal Process- Uncertainty relations and efficient recovery,” IEEE Transactions on
ing Magazine, IEEE, vol. 24, no. 4, pp. 118–121, 2007. Signal Processing, vol. 58, no. 6, pp. 3042–3054, 2010.
[19] E. J. Candès, “Compressive sampling,” in Proceedings of the Interna- [42] J. Rissanen and G. G. Langdon, “Arithmetic coding,” IBM Journal of
tional Congress of Mathematicians: Madrid, August 22-30, 2006: invited research and development, vol. 23, no. 2, pp. 149–162, 1979.
lectures, 2006, pp. 1433–1452. [43] I. H. Witten, R. M. Neal, and J. G. Cleary, “Arithmetic coding for data
[20] E. J. Candès and M. B. Wakin, “An introduction to compressive compression,” Communications of the ACM, vol. 30, no. 6, pp. 520–540,
sampling,” Signal Processing Magazine, IEEE, vol. 25, no. 2, pp. 21–30, 1987.
2008.
[21] E. J. Candes and T. Tao, “Near-optimal signal recovery from random
projections: Universal encoding strategies?” Information Theory, IEEE
Transactions on, vol. 52, no. 12, pp. 5406–5425, 2006.
[22] L. F. Polania, R. E. Carrillo, M. Blanco-Velasco, and K. E. Barner,
“Compressed sensing based method for ecg compression,” in 2011 IEEE
international conference on acoustics, speech and signal processing
(ICASSP). IEEE, 2011, pp. 761–764.
[23] J. Zhang, Z. Gu, Z. L. Yu, and Y. Li, “Energy-efficient ecg compression
on wireless biosensors via minimal coherence sensing and weighted l1
minimization reconstruction,” IEEE Journal of Biomedical and Health
Informatics, vol. 19, no. 2, pp. 520–528, 2014.
[24] Z. Zhang, T.-P. Jung, S. Makeig, and B. D. Rao, “Compressed sensing
for energy-efficient wireless telemonitoring of noninvasive fetal ecg
via block sparse bayesian learning,” IEEE Transactions on Biomedical
Engineering, vol. 60, no. 2, pp. 300–309, 2012.
[25] Z. Zhang, T.-P. Jung, S. Makeig, Z. Pi, and B. D. Rao, “Spatiotemporal
sparse bayesian learning with applications to compressed sensing of mul-
tichannel physiological signals,” IEEE transactions on neural systems
and rehabilitation engineering, vol. 22, no. 6, pp. 1186–1197, 2014.
[26] Z. Zhang and B. D. Rao, “Extension of sbl algorithms for the recovery
of block sparse signals with intra-block correlation,” IEEE Transactions
on Signal Processing, vol. 61, no. 8, pp. 2009–2015, 2013.
[27] H. Zhang, Z. Dong, Z. Wang, L. Guo, and Z. Wang, “Csnet: A
deep learning approach for ecg compressed sensing,” Biomedical Signal
Processing and Control, vol. 70, p. 103065, 2021.
[28] M. Mangia, L. Prono, A. Marchioni, F. Pareschi, R. Rovatti, and G. Setti,
“Deep neural oracles for short-window optimized compressed sensing
of biosignals,” IEEE transactions on biomedical circuits and systems,
vol. 14, no. 3, pp. 545–557, 2020.
[29] L. F. Polanı́a and R. I. Plaza, “Compressed sensing ecg using re-
stricted boltzmann machines,” Biomedical Signal Processing and Con-
trol, vol. 45, pp. 237–245, 2018.
[30] F. Picariello, G. Iadarola, E. Balestrieri, I. Tudosa, and L. De Vito,
“A novel compressive sampling method for ecg wearable measurement
systems,” Measurement, vol. 167, p. 108259, 2021.
[31] K. Luo, J. Li, and J. Wu, “A dynamic compression scheme for energy-
efficient real-time wireless electrocardiogram biosensors,” IEEE Trans-
actions on Instrumentation and Measurement, vol. 63, no. 9, pp. 2160–
2169, 2014.