Professional Documents
Culture Documents
Abstract—Wireless body sensor networks (WBSN) hold the promise to to be further compressed using non-linear digital techniques. CS
enable next-generation patient-centric tele-cardiology systems. A WBSN- is a methodology that has been recently proposed to address this
enabled electrocardiogram (ECG) monitor consists of wearable, minia-
problem. The main idea behind CS is relatively simple. Suppose
turized and wireless sensors able to measure and wirelessly report cardiac
signals to a WBSN coordinator, which is responsible for reporting them the considered signal has a (potentially large) bandwidth Ω but
to the tele-health provider. However, state-of-the-art WBSN-enabled ECG has a sparse approximation, i.e., it can be represented by a linear
monitors still fall short of the required functionality, miniaturization and superposition of K elements of a dictionary. CS states that you only
energy efficiency. Among others, energy efficiency can be significantly need to collect roughly K samples (<< the number of samples
improved through embedded ECG compression, which reduces airtime
over energy-hungry wireless links. In this paper, we quantify the potential corresponding to Nyquist-rate sampling) using simple measurement
of the emerging compressed sensing (CS) signal acquisition/compression waveforms, thus sensing/sampling and compressing at the same time.
paradigm for low complexity energy-aware ECG compression on the To the best of our knowledge, CS has never been applied to
state-of-the-art ShimmerTM WBSN mote. Interestingly, our results show ECG. It has, however, been recently considered for efficient EEG
that CS represents a competitive alternative to state-of-the-art digital
wavelet transform (DWT)-based ECG compression solutions in terms of
acquisition and compression [8], [9], [10]. Accordingly, the present
overall energy efficiency and ShimmerTM node lifetime extension. work introduces 2 main contributions: (1) It is the first to thor-
oughly quantify the potential of CS for low-complexity energy-
I. I NTRODUCTION aware ECG compression on resource-constrained WBSN platforms;
The increasingly prevalent cardiac diseases are requiring escalating (2) To quantify this potential, it provides an exhaustive system-level
levels of supervision and medical management, which are contribut- comparison between CS-based and state-of-the-art digital wavelet
ing to skyrocketing health care costs and, more importantly, are transform (DWT)-based embedded ECG compression. This system-
unsustainable for traditional health care infrastructures. Wireless body level comparison is based on the embedded implementations of
sensor networks (WBSN) promise to allow inexpensive, continuous the two considered ECG compression algorithms, which have been
and remote health monitoring for next-generation of ambulatory per- optimized for real-time implementation on a representative state-
sonal tele-cardiology or mobile-cardiology systems. While the resting of-the-art WBSN mote, namely ShimmerTM [2]. In particular, our
electrocardiogram (ECG) monitoring is standard practice in hospitals, comparative study of these 2 algorithms covers signal reconstruction
its ambulatory counterpart is still facing many technical challenges. quality, embedded memory usage, CPU execution time and energy
Recently, the realization of wireless-enabled ultra-low-power ECG consumption.
monitors for ambulatory use has been receiving significant industrial The rest of this paper is organized as follows. Section II introduces
and academic interest [1], [2], [3]. Still, today’s state-of-the-art the CS-based ECG compression and reconstruction scheme investi-
ambulatory ECG monitors fall short in terms of either clinical gated in this work, while Section III details the performance metrics
relevance and/or autonomy figure, primarily because they transmit and the ECG database used for its performance evaluation. Then,
uncompressed ECG data over power-hungry wireless links. It is today Section IV details our embedded implementation of CS-based ECG
widely acknowledged that the achievement of truly WBSN-enabled compression. In Section V, its performance is compared to that of
personal ECG monitoring systems requires more breakthroughs not state-of-the-art DWT-based compression. Finally, Section VI draws
only in terms of ultra-low-power read-out electronics and radios, but the concluding remarks.
also in terms of embedded ECG compression and feature extraction Notation: In all the following, normal letters designate scalar
algorithms, and assorted ultra-low-power dedicated digital processors. quantities, boldface lower-case letters indicate column vectors, and
ECG compression relies on the sparse (and thus compressible) boldface capitals represent matrices. Moreover, mi and Mi,j are the
nature of the ECG, as it can be approximated by a compact repre- ith entry of vector m and the (i, j)th entry of matrix M, respectively.
sentation in the wavelet domain [4]. Capitalizing on this sparsity, we Finally, (.)H and ||.||p denote the conjugate transpose, and the lp -
propose to apply the emerging compressed sensing (CS) approach [5], norm of a vector, respectively.
[6], [7] for a low-complexity, real-time and energy-aware ECG signal
compression on WBSN motes. This is motivated by the observation II. CS-BASED ECG C OMPRESSION
that this new signal acquisition and compression paradigm could be Figure 1 depicts the block diagram of CS-based ECG encoding.
well suited for low-power implementations as it dramatically reduces Due to the limited on-chip memory and real-time constraints, the
the need for resource- (both processing and storage) intensive DSP digitized ECG signal is processed in non-overlapping windows of N
operations on the encoder side. samples. Let x be the real-valued N -dimensional ECG signal vector
CS is a new sensing and processing paradigm, which challenges corresponding to such a window (i.e., x ∈ RN ). As aforementioned,
the traditional analog-to-digital conversion based on the Nyquist- x has a sparse approximation, i.e., it can be represented by a
Shannon sampling theorem. For sparse signals such as the ECG, linear superposition of S elements of an orthonormal wavelet basis
(above) Nyquist-rate sampling produces a large amount of redundant Ψ = [ψ 1 |ψ 2 | · · · |ψ N ], as follows x = ΨαS , where αS represents
digital samples, which are costly to wirelessly transmit and require the N -dimensional coefficient vector with only S non-zero entries
1745
50
20 CR = 50 45 CS (before)
18 40 DWT (before)
16 CR = 60 35 DWT (after)
14 30
CR = 65
12
25
10 CR = 70
20
8 CR = 75
15
6
CR = 80 10
4
5
2
0
0 0 20 40 60 80 100
5 10 15 20
Compression Ratio (CR)
Number of nonzero elements in sensing matrix Φ
Fig. 2. Output SN R vs. d for different compression ratios (CR)
Fig. 4. Output P RD vs. CR for CS and DWT before and after inter-packet
redundancy removal and Huffman coding
22
Sparse sensing
cardiology, we limited our review to those state-of-the-art algorithms
Output SNR (averaged over all Data)
20 Gaussian sensing
that were shown to achieve competitive compression ratios while be-
18
ing amenable to efficient, fast and low-memory-footprint implemen-
16
tation on our target platforms. Interestingly, these algorithms are all
14 based on the digital wavelet transform (DWT), namely, the embedded
12 zerotree wavelet (EZW) [22], the set partitioning in hierarchical trees
10 (SPIHT) [23] and thresholding-based algorithms [24]. Furthermore,
8 the latter thresholding-based algorithm by Benzid et al. [24], where
6
a fixed percentage of wavelet coefficients are zeroed, was shown
to outperform EZW and SPIHT at a lower computational cost.
4
50 55 60 65 70 75 80 Therefore, we will herein use it as a baseline algorithm to benchmark
Compression Ratio (CR) the performance, memory footprint, computational complexity and
Fig. 3. Performance comparison between Gaussian and sparse sensing energy consumption of our CS-based ECG compression scheme.
non-zero elements d in the sparse binary sensing matrix Φ. Clearly, B. Reconstruction quality
the output SN R saturates after d = 12 non-zero elements, which is
Figure 4 compares the output P RD, averaged over all database
the value retained for the rest of our hardware implementation. Note
records, for CS and DWT-based compression before and after inter-
that, in all our experiments, the sparse recovery problem of (3) was
packet redundancy removal and Huffman coding for different com-
solved using the SPGL1 solver [16] in Matlab c
.
pression ratios. It confirms the crucial role of the redundancy removal
Figure 3 shows the average output SN R vs. different compression
module and the careful design of the Huffman encoding. On average,
ratios (CR) for the two different approaches to the implementation
”very good” quality of reconstruction for DWT-based compression
of the random sensing matrix Φ explored in this section, namely,
is achieved up to CR = 73%, and up to CR = 51% for CS.
(1) Gaussian random sensing: the double-precision Matlab version;
Moreover, ”good” reconstruction quality is reached on average for
(2) Sparse binary sensing:
√ the version with d = 12 and all non-zero
up to CR = 90% and CR = 71% for DWT and CS, respectively.
entries equal to 1/ 12 implemented on the MSP430. These results
were obtained for an input vector of N = 512 samples and a 12- C. Memory usage
bit resolution for the input vector x and the measurement vector y.
The complete implementation of the thresholding-based DWT
Interestingly, the obtained results validate that there is no meaningful
compression requires 4.6 kB of RAM memory and 10 kB of Flash
performance difference between these two approaches.
memory, 5 kB of which is used for Huffman codebook storage. For
The use of a fixed binary sensing matrix, combined with the quasi-
CS, the complete implementation requires 6.5 kB of RAM and 7.5
periodic nature of the ECG signal, yields very similar consecutive
kB of Flash, 1.5 kB of which are for Huffman codebook storage.
measurement vectors y. So, there is a large inter-packet redundancy
that must be removed prior to encoding and wireless transmission. D. Power and energy consumption
Exploiting this fact, the redundancy removal module in Figure 1
To measure the power consumption during operation, we place
computes the difference between consecutive vectors, and only this
a 10.3Ω resistor in the power path of the platform. The voltage is
difference is further quantized. At the end, a state-of-the-art entropy
then readily measured using an oscilloscope, and the corresponding
coding module is used for further compression. Since the range of
current and power consumption are calculated. Our experimental
the difference signal just before encoding was found to be between
set-up consists of a Shimmer node implementing embedded ECG
[−256 : 255], a Huffman codebook of size 512 is needed with a
compression and wirelessly transmitting the compressed data using
maximum codeword length of 16 bits, for a given compression ratio.
its IEEE 802.15.4-compliant radio to a remote base station (a desktop
V. P ERFORMANCE A SSESSMENT AND B ENCHMARKING PC to which a second Shimmer node is connected). For simplicity,
a simple medium access control (MAC) scheme is implemented
A. Baseline thresholding-based DWT Compression between the node and the base station in FreeRTOS: the base station
Since our focus is on ECG compression techniques implementable periodically sends a beacon to maintain node synchronization, and the
in real-time on WBSN nodes for future ”personal” or ”wearable” tele- ECG sensing node sends its data in its guaranteed time slots (GTS),
1746
TABLE I
N ODE LIFE TIME FOR ” GOOD ” RECONSTRUCTION QUALITY WITH motes.
EMBEDDED COMPRESSION
VII. ACKNOWLEDGEMENT
DWT CS No Comp.
This research has been partially funded by the Nano-Tera.ch NTF
Compression Ratio (%) 90 71 0
Code execution time (ms) 580 99.5 0 Project BioCS-Node, which is financed by the Swiss Confederation.
Packet Ready every ... (ms) 2968.8 1023.7 296.9
R EFERENCES
Beacon Interval (ms) 11875 4094 1187
Energy Consumption (mJ) 8.78 7.46 8.37 [1] Toumaz Technology, 2009. [Online]. Available: http://www.toumaz.
Life time (h) (280 mAh@3.7 V ) 113.76 133.95 119.4 com/public/news.php?id=92
[2] Shimmer Research. [Online]. Available: http://shimmer-research.com
which is assigned by the base station and notified in the beacon. This [3] R. F. Yazicioglu and et al., “Ultra-low-power wearable biopotential
simple MAC is in fact compatible with the beacon-enabled mode of sensor nodes,” in Proc. of the IEEE EMBC’09, Sep. 2009.
[4] L. Sörnmo and P. Laguna, Bioelectrical Signal Processing in Cardiac
the IEEE 802.15.4 MAC protocol [25]. and Neurological Applications. Elsevier Academic Press, 2005.
To compress an ECG data vector of 512 samples (corresponding [5] E. Candès, J. Romberg, and T. Tao, “Robust uncertainty principles: Exact
to a duration of 2 seconds at sampling rate of 256 Hz), the signal reconstruction from highly incomplete frequency information,”
embedded CS code executes in 99.5 ms, whereas the DWT-based IEEE Trans. on Inform. Theory, vol. 52, no. 2, pp. 489–509, Feb. 2006.
[6] D. L. Donoho, “Compressed sensing,” IEEE Trans. on Inform. Theory,
code requires 580 ms. This confirms the relevance of CS as a vol. 52, no. 4, pp. 1289–1306, Apr. 2006.
low-complexity encoding scheme. To study the corresponding node [7] E. Candès and T. Tao, “Near optimal signal recovery from random
lifetime, power consumption of the different working modes of the projections: Universal encoding strategies,” IEEE Trans. on Inform.
node was characterized. Our measurements show an average node Theory, vol. 52, no. 12, pp. 5406–5425, Dec. 2006.
[8] S. Aviyente, “Compressed sensing framework for EEG compression,” in
power consumption of 84.6 mW during radio reception, and of Proceedings of the IEEE Workshop on Stat. Signal Proc. (SSP’07), Aug.
64.35 mW during transmission. The radio is only turned on for 2007, pp. 181–184.
beacon reception (i.e., 2.7 ms), and when there is a data packet [9] S. Şenay, L. F. Chaparro, M. Sun, and R. J. Sclabassi, “Compressive
to be transmitted. Moreover, each transmitted packet is 127-Byte sensing and random filtering of EEG signals using slepian basis,” in
long with 13 Byte of MAC overhead (4.2 ms). The rest of time, Proceedings of the EURASIP Signal Proc. Conf. (EUSIPCO’08), Aug.
2008.
the radio is off and the processor is idle and only wakes up [10] A. M. Abdulghani, A. J. Casson, and E. Rodriguez-Villegas, “Quantify-
for sampling to preserve energy, which corresponds to an average ing the feasibility of compressive sensing in portable electroencephalog-
power consumption of 6.44 mW . Accordingly, Table I compares raphy systems,” in Proceedings of the Intl. Conf. on Foundations of
the energy consumption and the resulting node lifetime (calculated Augmented Cognition, Neuroergonomics and Operational Neuroscience
(FAC’09), 2009, pp. 319–328.
for a 280 mAh battery at 3.7 V ) of the two algorithms for a [11] E. J. Candès and T. Tao, “Decoding by linear programming,” IEEE
”good” reconstruction quality. It also includes the same figures for the Trans. on Inform. Theory, vol. 51, no. 12, pp. 4203–4215, Dec. 2005.
streaming of uncompressed ECG. Interestingly, the reported results [12] E. van den Berg and et al., “Probing the pareto frontier for basis pursuit
show that CS achieves a relative node lifetime extension of 17.7 % solutions,” SIAM J. on Sci. Computing, vol. 31, no. 2, 2008.
[13] M. Figueiredo and et al., “Gradient projection for sparse reconstruction:
over DWT-based compression. This confirms the superior energy
Application to compressed sensing and other inverse problems,” IEEE
efficiency of CS for a given reconstructed signal quality. This superior Jrl. of Selec. topics in Signal Proc., vol. 1, no. 4, pp. 586–597, 2007.
energy efficiency is achieved thanks to CS’s lower complexity and [14] I. Daubechies, M. Defrise, and C. D. Mol, “An iterative thresholding
shorter processing time. Furthermore, the results of Table I establish algorithm for linear inverse problems with a sparsity constraint,” Comm.
that DWT-based compression actually fails to achieve any lifetime Pure Appl. Math., vol. 57, pp. 1413–1457, 2004.
[15] J. Tropp, “Greed is good: Algorithmic results for sparse approximation,”
extension, compared to the default alternative of streaming uncom- IEEE Trans. on Inform. Theory, vol. 50, no. 10, pp. 2231–2242, 2004.
pressed ECG data. Alternatively, CS shows a 12.9 % extension in [16] E. v. Berg and M. P. Friedlander, “SPGL1: A solver for large-
node lifetime. In fact, these promising results could have been even scale sparse reconstruction,” Jun. 2007. [Online]. Available: http:
more impressive had the MSP430 been optimized for ultra-low-power //www.cs.ubc.ca/labs/scl/spgl1
[17] G. B. Moody and R. G. Mark, “The impact of the MIT-BIH arrhythmia
DSP operations. Since this generic low-power microcontroller is not, database,” IEEE Eng. in Med. and Bio., vol. 20, no. 3, pp. 45–50, 2001.
it effectively limits the achievable node lifetime extension through [18] “MIT-BIH arrhythmia database.” [Online]. Available: http://www.
embedded signal pre-processing prior to wireless communication. physionet.org/physiobank/database/mitdb/
[19] Y. Zigel, A. Cohen, and A. Katz, “The weighted diagnostic distortion
(WDD) measure for ECG signal compression,” IEEE Trans. on Biomed.
VI. C ONCLUSION Eng., vol. 47, pp. 1422–1430, 2000.
[20] “The GCC toolchain for the Texas Instruments MSP430 MCUs.”
This paper proposed a complete system-level comparison between [Online]. Available: http://mspgcc.sourceforge.net/
a new CS-based and the state-of-the-art DWT-based embedded ECG [21] R. Berinde and et al., “Combining geometry and combinatorics: A
compression algorithms. As expected, nonadaptive CS-based com- unified approach to sparse signal recovery,” in Proc. of Allerton, 2008.
pression was found to exhibit inferior compression performance [22] M. L. Hilton, “Wavelet and wavelet packets compression of electrocar-
compared to its signal-adaptive DWT-based counterpart for a given diogram,” IEEE Trans. on Biomed. Eng., vol. 44, no. 5, pp. 394–402,
1997.
reconstructed signal quality. The presented results, however, were [23] Z. Lu, D. Y. Kim, and W. A. Pearlman, “Wavelet compression of ECG
obtained using the default basis pursuit denoise algorithm for sparse signals by the set partitioning in hierarchical trees algorithm,” IEEE
reconstruction, where no attempt has been made to exploit the highly Trans. on Biomed. Eng., vol. 47, no. 7, pp. 849–856, 2000.
structured nature of the ECG signal. Despite its inferior compression [24] R. Benzid, F. Marir, A. Boussaad, M. Benyoucef, and D. Arar, “Fixed
percentage of wavelet coefficients to be zeroed for ECG compression,”
performance, CS-based compression was found to exhibit the best Electronics Letters, vol. 39, no. 11, pp. 830–831, 2003.
overall energy efficiency thanks to its lower complexity and reduced [25] “Wireless medium access control (MAC) and physical layer (PHY) spec-
CPU execution time for ShimmerTM WBSN platform. As such, ifications for low-rate wireless personal area networks (LR-WPANs),”
our results validate the suitability of compressed sensing for real- IEEE, Tech. Rep. IEEE 802.15.4 Std., 2003.
time energy-aware ECG compression on resource-constrained WBSN
1747