VLSI implementation of low-power costefficient lossless ECG encoder design for

wireless healthcare monitoring application

diff_1
x(n)
x(n3)

diff_3

S.-L. Chen and J.-G Wang

An efcient VLSI architecture of a lossless ECG encoding circuit is
proposed for wireless healthcare monitoring applications. To reduce
the transmission and storage data, a novel lossless compression algorithm is proposed for ECG signal compression. It consists of a novel
adaptive rending predictor and a novel two-stage entropy encoder
based on two Huffman coding tables. The proposed lossless ECG
encoder design was implemented using only simple arithmetic units.
To improve the performance, the proposed ECG encoder was designed
by pipeline technology and implemented the two-stage entropy encoder
by the architecture of a look-up table. The VLSI architecture of this
work contains 3.55 K gate counts and its core area is 45987 m2
synthesised by a 0.18 m CMOS process. It can operate at 100 MHz
processing rate with only 36.4 W. The data compression rate
reaches an average value 2.43 for the MIT-BIH Arrhythmia
Database. Compared with the previous low-complexity and high performance techniques, this work achieves lower hardware cost, lower
power consumption, and a better compression rate than other lossless
ECG encoder designs.

Introduction: The electrocardiogram (ECG) is an important biomedical

signal and widely used to establish clinical diagnosis of heart diseases.
The traditional ECG recorder transmits the ECG data by cables, which is
incompetent for real-time monitoring in mobile scenarios or 24-hour
healthcare. In recent years, wireless homecare services have been
widely discussed and are of interest in academic and commercial
researches [1]. Similar to other wireless mobile systems, power saving
is necessary for the long-term usage of wireless ECG recorders. For
this reason, a data compression technique can be used to not only
reduce transmission energy but also transmit more data by the RF
module with limited bandwidth. Manikandan and Dandapat [2] proposed high-performance wavelet-based algorithms for ECG compression. Bendifallah et al. [3] present high-compression-rate ECG
compression methods based on discrete cosine transform (DCT) transform. These algorithms have the benet of high compression rate.
However, these algorithms are lossy compression algorithms, which
are unsuitable for developing healthcare monitoring systems.
To t the requirement of lossless ECG compression, Arnavut [4] proposed a high performance lossless ECG compression algorithm based
on the Burrows-Wheeler transform and MTF coding. Although this
method reaches high compression for lossless ECG compression, it is
not easy to implement into VLSI circuits. In our previous work [1], a
low-complexity lossless data encoder for a wireless body sensor
network had been proposed. Chua and Fang [5] proposed an efcient
lossless data compressor based on Golomb-Rice coding. Although
these designs provided efcient VLSI architectures, it is necessary to
reduce hardware cost and power consumption more to develop a
low-cost and long-term usage wireless monitoring system. In this
Letter, a novel low-complexity and high-compression-rate lossless
ECG encoder is proposed.

diff_2

x(n1)
x(n2)

diff_3

diff_1
diff_2

x(n)

x(n1)

x(n2)

x(n3)

diff_1=diff_2

diff_1=2 diff_2 diff_3

Fig. 1 First-order and second-order prediction

a First-order prediction
b Second-order prediction

To be able to improve the accuracy of prediction, the adaptive trending prediction methodology had been developed. By this prediction
method, the prediction strategy will be adaptively selected from the rstorder and second-order depending on the trends of the signals. For
example, if the trend of the last three values are linear, as shown in
Fig. 1b, the fourth value would probably match the slope trend and it
should be predicted by the second-order prediction method. On the
other hand, if only the last two values can be found as linear, as
shown in Fig. 1a, the fourth value would probably match the linear
trend and it should be predicted by the rst-order prediction method.
By this adaptive trending prediction, the accuracy of prediction can be
efciently improved.
Two-stage entropy coding: This is based on Huffman coding since it
gives variable-length codes (VLCs) for source xed-length codes
(FLCs). The Huffman coding is based on a Huffman coding tree
which has the feature that one branch should be added to represent a
new VLC code for each different value. The coding tree would grow signicantly if the input values vary in a large range. It is difcult to
develop a low-cost and high-performance VLSI chip which has a
block circuit with a huge depth of tree. For this reason, a novel two-stage
entropy coding technique has been developed. The prediction difference
(PD) values which most often appear are encoded by the rst Huffman
table as shown in Fig. 2. The rst Huffman table has the range of 5 to
+5 and one extending code for indicating out of range. If the prediction
difference values are over the range of the rst Huffman table, the prediction difference value will be modied from diff_1 to diff_1-diff_2
and then check the absolute value of diff_1-diff_2 is less or greater
than 20. If the absolute value is less than 20, the value of diff_1-diff_2
would be encoded by the second Huffman table. Otherwise, if the absolute value is greater than 20, the value of diff_1-diff_2 would be directly
sent by 9-bit without Huffman encoding. By composing the adaptive
trending prediction and two-stage entropy coding techniques, the compression rate of MIT-BIH Arrhythmia data reaches up to an average
value of 2.43.
adaptive trending predictor
input

x(n)

x(n3)

x(n2)

x(n1)

PD(n)

diff_3

diff_2
<<

x(n)

2*dIff_2-diff_3

two-stage entropy encoder

Lossless encoding algorithm: The proposed novel lossless encoding

algorithm consists of an adaptive trending prediction and a two-stage
entropy coding process. The adaptive trending prediction is used to
improve the coding efciency of the two-stage entropy coding. The
details of these two parts are described below.
Adaptive trending prediction: Since the biomedical signals change
slowly and have predicable distribution, the prediction methodology can
be used to improve the performance of the proposed lossless encoding
algorithm. In our previous work [1], a rst-order prediction methodology based on linear prediction was used to forecast the present
value of the biomedical signal with two passed values. As shown in
Fig. 1a, the present value x(n) can be obtained by passing two values
of x(n1) and x(n2) with the relationship that diff_1 is equal to
diff_2. To improve the performance of prediction, a second-order prediction method based on slope prediction was added into the proposed
lossless compression algorithm. As shown in Fig. 1b, the present
value x(n) can be obtained by passing three values of x(n1), x(n2)
and x(n3) with the relationship that diff_1 is two times of diff_2
minus diff_3.

PD(n)

extend
1
1

10
111
0000

first Huffman
table

extend

PD(n)

extend
1
1

11
0111
0110

second Huffman
table

00110

20

010...000

00111

19

010...001

PD(n1)

output

PD(n1)

Fig. 2 Architecture of proposed lossless ECG encoder

VLSI architecture: The lossless ECG encoder consists of two parts

including an adaptive trending predictor and a two-stage entropy
encoder. Fig. 2 shows the architecture of the proposed lossless ECG
encoder. The details of the adaptive trending predictor and a two-stage
entropy encoder are described below.

ELECTRONICS LETTERS 17th January 2013 Vol. 49 No. 2

Adaptive trending predictor: This consists of ve registers, one

adder, four subtractors, one shifter, and one multiplexer. Four of the ve
registers are used to store the input values of x(n), x(n 1), x(n 2),
and x(n 3), and the other one is used to store the value of prediction
difference. The register PD(n) is also a pipeline register for improving
the performance of the proposed encoder design. The values of diff_2
and diff_3 can be obtained by two substractors. The value of 2*diff_2diff _3 can be calculated by a shifter and a substractor with the obtained
values of diff_2 and diff_3. The predicted value of diff_1 can be selected
adaptively from the values of diff_2 and 2*diff_2-diff _3 according to
the trend of the signal. Finally, the value of the prediction difference
can be produced for entropy coding.
Two-stage entropy encoder: This consists of two Huffman tables,
one register, and one substractor. The two Huffman tables were
implemented by the architecture of look-up tables. The two-stage
encoder receives the input signal PD from the predictor and then
sends it to the rst Huffman table for entropy encoding. If the value
of PD is in the range of 5 to 5, the value of PD would be encoded
by the rst Huffman table and the encoded data would be sent to
output directly. Otherwise, if the PD value is out of range in the rst
table, an extending code would be produced and the value of PD
(n)-PD(n1) would be sent to the second Huffman table for entropy
coding. The results of these two tables are variable length codes
(VLCs). Finally, The VLC codes will be collected and sent to output
for transmission.
Results and discussion: To be able to analyse the performance of the
proposed ECG lossless compression algorithm, the MIT-BIH
Arrhythmia Database was used as test patterns. After encoding by the
proposed lossless compression algorithm, the average compression
rate of all patterns in the MIT-BIH Arrhythmia Database is 2.43 as
shown in Table 1. To verify the proposed lossless ECG encoder
design, a FPGA board with an Altera EP4CE115F29C7N FPGA core
was used. Also, the VLSI technique was used to realise this design. It
was implemented by Verilog code, simulated by the ncverilog tool,
and synthesised by design compiler with TSMC 0.18 m COMS
generic logic process technology. Synthesised results show that this
encoder chip contains 3.57-K NAND-equivalent gate counts. The auto
placement and routing tool IC Compiler was used to generate the
layout of the proposed encoder design. The width and length of the proposed encoder chip are 213.52 m and 215.38 m and the chip area is
45987 m2. The power consumption of this design is 36.4 W operating
at 100 MHz 1 V supply voltage.

Table 1: Compressions with previous designs

CR

[1]
1.9

[5]
2.38

Process
0.18 m 65 nm
Operation frequency (Hz) 100 M
24 M
Core voltage
1.8 V
1V
Simulation power
150 W 170 W
Gate counts
13.4 K 53.9 K
Core area (m2)
Normalised area

134 K
2.91

58 K
9.67

This work
2.43
0.18 m
100 M
1V
36.4 W
3.57 K

Table 1 lists the specications of previous VLSI designs [1, 5] with

this work. The compression rate (CR) of the MIT-BIH Arrhythmia
Database in this work is 2.43, which is much better than 1.9 or 2.38
of the previous design [1] or [5]. The operation clock frequency of
this design is 100 MHz, which is higher than 27 MHz of the previous
design [5]. Compared in terms of the hardware cost, this work reduces
at least 73.3 % or 93.3 % gate counts than that of the previous design
[1] or [5]. Moreover, the power consumption of this work is 36.4 W
which is much less than 150 W or 170 W of the previous design
[1] or [5]. It means that at least 75.7 % or 78.5 % power consumption
can be saved in this design. The normalised area of the previous
design [1] or [5] is 2.91 or 9.67 times of this work, which means that
at least 65.6 % or 89.7 % chip area is efciently reduced in this design.
Conclusion: In this Letter, a cost-efcient and high-compression-rate
lossless ECG compressor is proposed. A novel low-complexity lossless
compression algorithm based on an adaptive trending prediction and a
two-stage Huffman coding technique is developed. The VLSI architecture of this design has the benets of low cost, low power consumption,
high performance, and a high compression rate.
Acknowledgment: This work was supported by the National Science
Council, Taiwan, under grant numbers NSC-100-2218-E-033-007 and
NSC-101-2221-E-033-006.
The Institution of Engineering and Technology 2013
9 October 2012
doi: 10.1049/el.2012.3505
S.-L. Chen (Department of Electronic Engineering, Chung Yuan
Christian University, Chung Li City 320, Taiwan)
E-mail: chrischen@cycu.edu.tw
J.-G Wang (Department of Information and Electronics, Beijing
Institute of Technology, Beijing 100081, Peoples Republic of China)
References
1 Chen, S.L., Lee, H.Y., Chen, C.A., Huang, H.Y., and Luo, C.H.:
Wireless body sensor network with adaptive low power design for biometrics and healthcare applications, IEEE Syst. J., 2009, 3, (4), pp.
398409
2 Manikandan, M.S., and Dandapat, S.: Effective quality-controlled
SPIHT-based ECG coding strategy under noise environments,
Electron. Lett., 2008, 44, (20), pp. 11821183
3 Bendifallah, A., Benzid, R., and Boulemden, M.: Improved ECG compression method using discrete consine transform, Electron. Lett., 2011,
47, (2), pp. 8789
4 Arnavut, Z.: ECG signal compression based on Burrows-Wheeler transformation and inversion ranks of linear prediction, IEEE Trans. Biomed.
Eng., 2007, 54, (3), pp. 410418
5 Chua, E., and Fang, W.C.: Mixed bio-signal lossless data compressor for
portable brain-heart monitoring systems, IEEE Trans. Consum.
Electron., 2011, 57, (1), pp. 267273

46 K
1

Note: The area is normalised to 0.18 m process by areaprocess(0.18/Process)2

ELECTRONICS LETTERS 17th January 2013 Vol. 49 No. 2