See

discussions, stats, and author profiles for this publication at: http://www.researchgate.net/publication/266657438

Ultra-Low Power Design of Wearable Cardiac

Monitoring Systems
ARTICLE JUNE 2014
DOI: 10.1145/2593069.2596691

CITATIONS

7 AUTHORS, INCLUDING:
H. Mamaghanian

David Atienza

cole Polytechnique Fdrale de Lausanne

cole Polytechnique Fdrale de Lausanne

11 PUBLICATIONS 183 CITATIONS

268 PUBLICATIONS 2,388 CITATIONS

SEE PROFILE

SEE PROFILE

Francisco J. Rincn

Srinivasan Murali

cole Polytechnique Fdrale de Lausanne

cole Polytechnique Fdrale de Lausanne

16 PUBLICATIONS 76 CITATIONS

68 PUBLICATIONS 2,626 CITATIONS

SEE PROFILE

SEE PROFILE

Available from: David Atienza

Retrieved on: 30 August 2015

Ultra-Low Power Design of Wearable Cardiac

Monitoring Systems
Ruben Braojos, Hossein Mamaghanian,
Alair Dias Junior*, Giovanni Ansaloni and David Atienza

Francisco J. Rincon,
Srinivasan Murali

Embedded Systems Laboratory, EPFL, Switzerland

Email: {ruben.braojoslopez}, {hossein.mamaghanian},
{alair.diasjunior}, {giovanni.ansaloni}, {david.atienza}@ep.ch

SmartCardia, Switzerland
Email: [email protected],
[email protected]

AbstractThis paper presents the system-level architecture

of novel ultra-low power wireless body sensor nodes (WBSNs)
for real-time cardiac monitoring and analysis, and discusses the
main design challenges of this new generation of medical devices.
In particular, it highlights rst the unsustainable energy cost
incurred by the straightforward wireless streaming of raw data
to external analysis servers. Then, it introduces the need for new
cross-layered design methods (beyond hardware and software
boundaries) to enhance the autonomy of WBSNs for ambulatory
monitoring. In fact, by embedding more onboard intelligence
and exploiting electrocardiogram (ECG) specic knowledge, it
is possible to perform real-time compressive sensing, ltering,
delineation and classication of heartbeats, while dramatically
extending the battery lifetime of cardiac monitoring systems. The
paper concludes by showing the results of this new approach to
design ultra-low power wearable WBSNs in a real-life platform
commercialized by SmartCardia. This wearable system allows a
wide range of applications, including multi-lead ECG arrhythmia
detection and autonomous sleep monitoring for critical scenarios,
such as monitoring of the sleep state of airline pilots.
Index TermsWearable Embedded Systems, Bio-Medical Signal Processing, Wireless Body Sensor Nodes.

I. I NTRODUCTION AND M OTIVATION

Embedded cardiac monitors are wearable and miniaturized
devices, providing the acquisition, on-board processing and
wireless transmission of cardiac bio-signals for prolonged
periods of time. These wireless body sensor nodes (WBSNs)
allow non-intrusive and long-term monitoring of cardiac parameters of patients, such as electrocardiogram (ECG) and

pulse oximetry (SpO2). They represent novel solutions in

the healthcare domain [1], both for the prevention of acute
episodes (e.g., strokes) and for the assessment of chronic
conditions (e.g., sleep disorders, stress-related pathologies).
The design of cardiac monitoring platforms is applicationdriven, because relevant data for medical examination strongly
depends on the considered scenario. As an example, sleep
monitoring applications involve the analysis of heart rate
variability over a time window of the acquired bio-signal
while the assessment of the recovery after a stroke instead
requires the detailed analysis of the morphology of heartbeats.
Minimizing the bandwidth on the energy-hungry wireless link
[2] according to the application requirements is therefore an
effective strategy for the design of ultra-low power cardiac
monitors. To this end, on-node signal processing can be
employed to derive the data of interest at different levels of
abstraction before transmission, as illustrated in Figure 1.
This paper investigates how the energy efciency of cardiac
monitors can be enhanced by leveraging embedded processing
and hardware-software optimization methodologies. We address its design space in a top-down fashion. First, cardiac applications and their requirements are introduced in Section II.
Then, in Section III we detail how signal processing algorithms
can support such applications and, in Section IV the challenges



 "

* Alair Dias Junior is also with FUMEC University, Brazil.

This work has been partially supported by the Spanish Government Research Grant TEC2012-33892, the EC FP7 FET Phidias project (Grant agreement no. 318013.), a CAPES Scholarship (grant no. 2171-13-9), and the ObeSense (no. 20NA21 143081) and BodyPoweredSenSE (no. 20NA21 143069)
RTD projects evaluated by the Swiss NSF and funded by Nano-Tera.ch with
Swiss Confederation nancing.
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are not
made or distributed for prot or commercial advantage and that copies bear
this notice and the full citation on the rst page. Copyrights for components
of this work owned by others than ACM must be honored. Abstracting with
credit is permitted. To copy otherwise, or republish, to post on servers or to
redistribute to lists, requires prior specic permission and/or a fee. Request
permissions from [email protected].
DAC 14, June 01 - 05 2014, San Francisco, CA, USA
Copyright 2014 ACM 978-1-4503-2730-5/14/06...$15.00.

 

 

  

  
! 

 
 





 


$ 




%

Fig. 1. On-node digital signal processing increases the energy efciency

of cardiac monitoring by rising the abstraction level and decreasing the
bandwidth of transmitted data.

involved in embedding them in resource-constrained cardiac

monitors. Section V reports the performance of embedded
cardiac monitors in real-world scenarios and nally section VI
outlines promising future directions in this eld and concludes
the paper.
II. W IRELESS B ODY S ENSOR N ODES A PPLICATIONS
While traditional cardiac monitoring equipment are bulky
and target patient monitoring for short periods of time, recent
advances in technology have enabled the development of ultrasmall and wearable monitoring devices. With the development
of embedded cardiac monitors, we can envision a new generation of wearable systems where different sensors are integrated
directly in the users dress, processing and streaming the vital
signs information to a mobile phone or a cloud server.
This opens up new application domains for cardiac monitoring, ranging from extracting behavioral information of the
user (mental and emotional state), sleep/fatigue information
and early detection and prevention of diseases.
The digital signal processing required for such different
application scenarios varies widely. Applications that extract
behavioural information typically only require processing of
beat-to-beat intervals, while the diagnosis of heart problems
requires the processing of more detailed morphological information of each heartbeat. Applications such as driver sleep
monitoring require classiers that are trained over different
population sets and driving scenarios, and ported onto embedded micro controllers. The noise level of the signal and the
required ltering algorithms also vary based on the application.
For a cardiac system embedded in vehicles that use noncontact sensors, a major challenge is to remove the common
mode electrical noise and obtain a clean signal, while for
ambulatory monitoring of stroke patients, the challenge is to
remove muscular and motion artifacts. Towards this end, in
this paper, we present an application driven methodology for
cardiac and bio-signal processing.
III. C ARDIAC SIGNAL PROCESSING METHODS
As illustrated in Figure 1, digital signal processing on
embedded cardiac monitors typically follows a number of
phases. A ltering stage is mandatory before performing
further processing steps. To further counteract noise, different
signals with the same modality (e.g.: multi-lead ECGs) can
be combined after ltering. Afterwards, characteristic features
are derived from the the cardiac signals with processes such
as ECG delineation, and classication is performed to detect
abnormalities in each heart-beat. Information at each level
is instrumental to perform automated diagnosis for different
cardiac-related pathologies.
An alternative solution, also effective in minimizing transmission bandwidth, is to compress the signal on the cardiac
monitor, and reconstruct it for further analysis on the receiver
side. The two approaches are not mutually exclusive. In [3], in
fact, a classication methodology on compressed ECG signals
has been proposed.
This section illustrates promising solutions for the automation of the different phases.

A. ECG compression
Todays state-of-the-art WBSN-enabled ambulatory ECG
monitors still fall short of the required energy efciency and
longevity. This is mainly because of the raw data transfer over
energy-hungry wireless links. It is today acknowledged that
the achievement of truly WBSN-enabled ambulatory monitoring systems requires more breakthroughs not only in terms
of ultra-low-power read-out electronics and radios, but also
in terms of dedicated digital processors, which execute the
associated embedded feature extraction and data compression
algorithms in order to reduce airtime over wireless links.
In [4], the potential of the compressed sensing (CS) signal
acquisition/compression paradigm for low-complexity energyefcient ECG compression has been investigated. It is shown
that CS could outperform its state-of-the-art counterpart compression algorithms, thanks to its low complexity and CPU
execution time in terms of overall energy efciency. With more
advances on the reconstruction algorithm, in [5], the possibility
of a real-time CS decoder running on an iPhone (acting as a
WBSN coordinator) has been demonstrated.
Building on these initial works which proves the suitability
of using CS for ECG compression on the resource- and
energy-aware WBSN, the technique has been extended to
fully leverage and exploit underlying structural information,
like any state-of-the-art compression technique. More specic
recovery algorithms for single lead and multi-lead ECG has
been presented in [6].
Beyond all these works, CS is usually used as a very low
cost and easy to implement compression technique. Signals
should be acquired with the traditional limitations on the
bandwidth (BW) and after the major portion of redundant
data should be discarded. The main challenges are then taking
the whole design of the front-end and read-out devices to the
next level based on the promises of the CS and merging the
sampling and compression steps. This removes a large part of
the digital architecture and considerably simplies analog-toinformation (A2I) conversion devices.
This so-called analog CS, where compression occurs directly in the analog sensor readout electronics prior to analogto-digital conversion, could thus be of great importance for
applications where bandwidth is moderate, but computationally complex, and power resources are severely constrained.
Different realization of the CS-based analog-to-information
readout devices has been introduced in the literature [7], [8],
although designing a truly CS-based A2I still remains as a
challenge.
B. Noise ltering and source combination
Cardiac bio-signals are usually affected by noise, which
must be ltered before relevant features can be retrieved
from the acquisitions. Different sources of noise range from
those coming from environmental factors (e.g. electromagnetic
interference) to others of biological nature (e.g. muscular
activity). The authors of [9] propose a ltering technique based
on the application of two morphological operators (erosion
and dilation), which removes unwanted components from


 


 










Fig. 2.







 






 




Delineated normal sinus beat.

the input signal. Alternatively, [10] proposes a technique for

removing low-frequency components causing the so called
baseline wandering. This method, based on cubic splines,
searches for knots in a characteristic silent region of the
acquired signal (before each QRS complex), and interpolates
three consecutive knots to estimate the baseline.
In [11], it is shown that the effect of noise can be also
reduced by combining different ECG leads before the analysis
and/or delineation phase. Simple root mean square (RMS)
aggregation of inputs, is presented as a light-weight, yet
effective, implementation strategy.
C. ECG Delineation
ECG delineation is the process of identifying the ducial
points (start, peak and end) of the characteristic waves composing each heart-beat (Figure 2). The information enables
the diagnosis of a large set of cardiac conditions, such as
arrhythmias. For this purpose, [12] propose a method based
on wavelet decomposition, relying on the fact that different
waves present distinct frequency components.
An alternative strategy, presented in [13], proposes to use
a morphological transform of the ECG to automate the delineation of ducial points. The method gives an effective
solution to the problem of delineation, as minima in the
transformed signal indicates the presence of peaks in the
original wave, while maxima (or sudden changes in slope)
delimit the start and end point of each wave.
D. Embedded classication
Higher-level ECG processing, after the delineation phase,
can further reduce the amount of data to be transmitted, while
also adding further information about the processed signal. The
work presented in [14] proposes the utilization of a classier
to identify abnormal heartbeats and trigger delineation only in
those cases. The authors describe a methodology to implement
a fuzzy network into a state-of-the-art WBSN meeting realtime and sensitivity constraints.
An important problem in classication is to identify a set
of representative features for each heartbeat. The random projection approach introduced in [15] is particularly promising
in this context. It allows the minimization of the number of
features (and therefore a simplied classication process) as
well as the efcient computation of the feature set.
IV. E MBEDDED CARDIAC MONITORING PLATFORMS
The design of a truly effective embedded cardiac monitor
has to overcome diverse challenges. The most important design

goals are to require low maintenance and allow long-term

execution, at the same time causing as little discomfort as
possible to subjects. Moreover, accurate diagnostic data should
be retrieved for medical evaluation.
These conicting goals require a careful system-wide evaluations in the platform design. Energy efciency is a key
objective, as it allows to minimize the battery size and weight
while retaining continuous operations for extended periods of
time. This section describe several effective strategies (both
at the hardware and software levels) allowing ultra-low power
regimes while performing embedded processing.
Moreover, physical constrains impede the continuos acquisition of important cardiac parameters. We show how parameters
such as Blood Pressure (BP) can instead be estimated from
other ready-obtainable quantities.
A. Software optimizations
The ltering, delineation and classication algorithms proposed in Section III must be carefully tailored to be executed
on resource-constrained embedded cardiac monitors. Typically, these platforms (like the SmartCardia device presented
in Figure 4) operate at a clock frequency of few MHz and
only support integer arithmetic operations.
Computationally-intensive parts of applications must therefore be analyzed to derive light-weight implementations. This
approach is exemplied in [12], where a proper choice of
the lter bank coefcients leads to an efcient execution of
the wavelet-based delineation on a state-of-the-art embedded
device. In the cases of morphological ltering and delineation,
if a at structuring element is employed, the computational
demands of the morphological operations can be drastically
reduced by keeping track of only the center value, maximum
and minimum in a sliding window of the input signal.
The approximation of complex functions is also as an effective strategy to reduce algorithmic complexity. For instance,
in the scenario of heartbeat classication, which usually involves the evaluation of many gaussian functions, a foursegments linearization is shown to achieve close-to-optimal
results [14], while vastly simplifying the computational requirements. Moreover, by using random projections (described
in Section III-D) memory usage can be minimized considering
a projection matrix only composed by elements of value 0,
1 and -1, which can be represented using only two bits per
component.
Sparsity considerations are important to minimize computation and memory usage. In [16], it is shown how few nonzero elements in the sensing matrix sufce to achieve close-tooptimal results when performing compressive sensing, while
minimizing the run-time workload.
Along the same lines, the knowledge of the structure of
coefcient vectors can be exploited to increase the quality of
compressed sensing by differentiating signal information from
recovery artifacts. A rst approach stems from the observation
that wavelet coefcients are naturally organized into a tree
structure, and the largest coefcients cluster along the branches
of this tree. A CS reconstruction algorithm based on the


'


&


&


'







"#




&

%






"#

!


'



%





"

%

%

 



&

!



 


!


"

!




Fig. 3. Hardware architecture of a typical multi-core WBSN. In red, hardware

support for synchronization.

connected tree model has been proposed in [17]. Second, in

case of the multi-lead ECG compression, there is a strong
correlation between the sparsity structure among the leads,
each lead therefore conveying useful information about other
leads. In particular, non-zero coefcients are partitioned in
subsets or groups, and this information can be employed to
enhance the compression performance across all leads [6].
B. Ultra-low-power architectures for cardiac monitoring
Cardiac monitoring applications present a high degree of
parallelism, either because multiple sources are processed by
the same algorithms (e.g., ltering of multiple ECG inputs)
or because processing is divided into consecutive different
phases of computation (as introduced in Figure 1). These
characteristics allow the parallelization of the workload on a
multi-core core architecture by exploiting voltage scaling to
achieve substantial energy savings.
In [18] we propose a platform interfacing multiple processors to independent multi-bank program and data memories (cf. Figure 3). Low-overhead mechanisms are introduced
to synchronize code execution and enable single-instruction
multiple-data (SIMD) operations, resulting in a decrease of
the energy consumption of the instruction memory subsystem. To this end, the employed broadcasting mechanism,
which is implemented by the interconnect networks, merges
multiple identical read requests from different cores into a
single memory access. Moreover, this architecture includes
a software technique based in barrier insertion to maintain
cores in lock-step and recover from de-synchronization after
data dependent branches. This technique requires a reduced
instruction set extension that annotates the synchronization
status on dedicated data words stored in the shared memory to
allow the synchronization hardware to properly orchestrate the
execution ow. This architecture enables the use of producerconsumer relationships among computational stages and the
application of the proposed methodology allows to correctly
map bio-medical applications onto the multi-core platform
in order to avoid program memory conicts, and therefore
unnecessary stalls and performance degradation. Interestingly,
ne-tuned load balancing is not a necessary precondition for
energy efciency in cardiac monitoring systems.
A complementary approach for minimizing power consumption in this kind of devices is to include application-specic

accelerators. In the case of compressed sensing, the authors of

[19] highlight that a minimal hardware support accompanied
by a specic instruction set extension of a RISC core can
achieve more than ten-fold power saving with respect to a
baseline implementation while performing compressed sensing
over an ECG signal.
C. Real-time estimation of multiple cardiac parameters
In clinical scenarios, information extracted from the ECG
is analyzed along with other cardiac bio-signals, such as
blood pressure (BP) and pulse oximetry, to assess the global
health status of the cardiovascular system. The realization of
multi-modal cardiac monitors integrating diverse sensors in
a wearable device is, therefore, desirable but not straightforward. As an example, sensors capable of directly measuring
blood pressure are either cumbersome (e.g., inating cuffs) or
extremely complex (e.g., arterial tonometry).
An interesting approach for multi-modal analysis is to
combine available information to estimate parameters which
cannot be easily measured. For instance, the pulse arrival time
(PAT), calculated using ECG and a simple and inexpensive
photoplethysmograph (PPG) nger probe, can be used to
estimate the pulse wave velocity (PWV), which is a surrogate
marker for arterial stiffness and BP [20].
In addition, when signals of multiple modalities are considered, the correlation between the different inputs can be used to
facilitate signal processing. Most cardiac bio-signals originate
from the response to the bioelectric stimuli reected in the
ECG. The signals are, therefore, time-locked to these stimuli.
This information can be used to remove noise (which is instead
uncorrelated to the stimuli) with different techniques such
as ensemble averaging (EA) and adaptive impulse correlated
ltering (AICF). In [21] it is shown that ECG information can
be employed to calculate, among other parameters, the EA of
the pulse oximetry. Also, AICF can be used to lter the ECG
signal [22] and to de-noise PPG signals [23]. The disadvantage
of using EA is that the beat-to-beat variation of the signals
is lost after the processing. AICF, on the other hand, is also
capable of tracking dynamic changes in the signal.
V. E XPERIMENTAL RESULTS
Embedded cardiac monitors enable a new family of application scenarios in the healthcare domain. Although they
are computationally- and energy- constrained, these devices
are able to perform the advanced signal processing algorithms
described in Sections III and IV, allowing novel breakthroughs
in the healthcare monitoring eld. As an example, the SmartCardia device [24] depicted in Figure 4 is able to perform
3-lead ECG monitoring and execute a set of algorithms for
real-time ECG ltering and arrhythmia detection. In particular, it allows the automatic detection of abnormal cardiac
events and the remote notication of such phenomena to a
centralized server infrastructure, thanks to an on-board radio
transceiver. Compressed Sensing is employed to efciently
transmit excerpts of the acquired signals, periodically or when
an abnormality is detected. This SmartCardia device embeds

Sampling

Radio

Comp.

Multilead CS
SingleLead CS
No Comp.
0

Fig. 6.

Fig. 4.

SmartCardia 3-lead ECG monitoring device.

an ultra-low-power micro-controller for digital processing, an

acquisition front-end and a wireless transmission stage, all in
a compact form factor. The mean time between charges is
typically one week.
This level of functionality is achieved by an optimized
implementation of hardware and software components, such
as the ones described in this paper. Moreover, these optimizations do not have major impact on the obtained quality of
the retrieved data. As an example, the performance of the
illustrated ECG delineation algorithms are in line with the
results reported by computing-demanding off-line variants,
while requiring only a fraction of the resources (7% of the
duty cycle and 7.2kB of memory [12]). For this application,
the measured sensitivity and specicity of retrieved ducial
points are above 90% in all cases, which is at the target level
for medical use in this eld.
Figure 5 compares the averaged signal-to-noise ratio (SNR)
results over different compression ratios (CR) for singlelead and multi-lead CS compression [6]. These results show
that an averaged SNR over 20 dB (corresponding to good
reconstruction quality [16]) is reached for CR = 65.9% and
CR = 72.7% for single and multi-lead CS, respectively. To
characterize the power gure, the compression algorithms are
implemented on our target WBSN running FreeRTOS on a
16-bit processor and simple medium access control (MAC)
scheme for wireless communication (IEEE 802.15.4) between
the node and the base station. Figure 6 depicts the share of
radio, OS and compression in the total node power consumpSingleLead CS
MultiLead CS

30
25
G
15

40
Energy [J]

60

80

Breakdown of energy consumption for our target platform.

tion for raw data streaming and two CS-based compression for
good quality of reconstruction. The average power reduction
estimates are 44.7% and 56.1% compared to raw-data streaming for single-lead and multi-lead CS compression. This proves
the suitability of CS as a promising low-power compression
technique for wearable cardiac monitoring systems.
In addition, similar results have been obtained for applications performing a diagnosis at a higher level of abstraction,
such as Atrial Fibrillation (AF) detection [25]. This cardiac
monitoring application uses the results of the ECG delineation
to analyze the regularity of the heart beat rate as well as
the shape of the P wave, which constitute two characteristic
irregularities of AF episodes. The results generated after the
observation of these irregularities can be subsequently analyzed in real-time using a fuzzy classier. This low-complexity
approach achieves 96% sensitivity and 93% specicity, which
are comparable gures to state-of-the-art off-line AF detection
algorithms while operating in real-time on an embedded
device.
Finally, the use of multi-core computing architectures in
cardiac monitoring systems (cf. Section IV-B) increases further
the energy efciency of the processing stage. In particular,
Figure 7 shows how the ltering (3L-MF), delineation (3LMMD) and classication (RP-CLASS) applications described
in Section III can be efciently executed on the presented
multi-core platform of Figure 3 (MC), reducing up to 40%
the global power consumption with respect to a single-core
variant (SC).
)((

 

%" &

Averaged SNR over all record

35
VG

20

+$

+$

$




$

$


'



#

#
#

-,

,(

*,

10
65.9
5

20

72.7

40
60
80
Compression Ratio (CR)

100

Fig. 5. Output averaged SNR over all records over different compression
ratios for single-lead and multi-lead compression.

(













Fig. 7.
Average power consumption decomposition of a synchronized
multicore (MC) system and an equivalent single-core (SC) architecture.

VI. C ONCLUSIONS AND F UTURE A PPLICATIONS

In this paper we have presented the main design challenges and latest approaches targetting next-generation ultralow power wearable cardiac monitors, which need to perform
an energy-efcient real-time acquisition and processing of
ECG, thus providing the possibility of performing on-line
diagnosis and analysis of the cardivascular state of a person.
In order to develop this new cardiac monitoring systems, we
have shown that multiple optimizations exploiting the target
bio-signal features need to be performed at both hardware and
software levels in order to achieve the target ultra-low power
gures needed to extend autonomy of the systems for longterm bio-signals monitoring.
As a result, these new ultra-low power cardiac monitoring
WBSNs will be able to enable novel medical approaches for
the diagnosis of diverse ailments. In particular, a promising
direction to be investigated in the near future is related to
neurodegenerative disorders, such as, Alzheimer Desase (AD).
Even though the mechanisms linking the autonomic regulation
of the heart activity to AD are not fully understood, recent
studies have demonstrated a correlation between neurological
degeneration and aberrant ECG signals, indicating the latter
as a potentially useful bio-marker of AD progression [26].
Moreover, researchers have explored the neural basis of
bodily self-consciousness and have showni in very recent
studies that a tight relationship exists between the brain
processes and the body states, such as, the heartbeat [27].
Therefore, expected development of wearable cardiac monitors
has the potential to open up several new applications in neurorehabilitation by exploiting the relationship between the brain
processes and cardiac factors, as well as enabling effective
mind training and improvement of children with learning
disabilities.
R EFERENCES
[1] Y. Hao and R. Foster, Wireless body sensor networks for healthmonitoring applications, Physiological Measurement, vol. 29, no. 11,
p. R27, Oct. 2008.
[2] F. Chen, F. Lim, O. Abari, A. Chandrakasan, and V. Stojanovic, Energyaware design of compressed sensing systems for wireless sensors under
performance and reliability constraints, Circuits and Systems I: Regular
Papers, IEEE Transactions on, vol. 60, no. 3, pp. 650661, Mar. 2013.
[3] M. Shoaib, K. Lee, N. Jha, and N. Verma, A 0.6 107 uw energyscalable processor for directly analyzing compressively-sensed eeg,
Circuits and Systems I: Regular Papers, IEEE Transactions on, vol. PP,
no. 99, pp. 114, 2014.
[4] H. Mamaghanian, N. Khaled, D. Atienza, and P. Vandergheynst, Compressed sensing for real-time energy-efcient ecg compression on wireless body sensor nodes, Biomedical Engineering, IEEE Transactions
on, vol. 58, no. 9, pp. 2456 2466, sept. 2011.
[5] K. Kanoun, H. Mamaghanian, N. Khaled, and D. Atienza, A realtime compressed sensing-based personal electrocardiogram monitoring
system, in Conference on Design, Automation and Test in Europe
(DATE 11), Sep. 2010.
[6] D. A. Hossein Mamaghanian, Giovanni Ansaloni and P. Vandergheynst,
Power-efcient joint compressed sensing of multi-lead ecg signals,
in IEEE International Conference on Acoustics, Speech, and Signal
Processing (ICASSP), 2014.
[7] Y. Eldar, Analog compressed sensing, in Acoustics, Speech and Signal
Processing, 2009. ICASSP 2009. IEEE International Conference on,
April 2009, pp. 29492952.

[8] H. Mamaghanian, N. Khaled, D. Atienza, and P. Vandergheynst, Design

and exploration of low-power analog to information conversion based
on compressed sensing, Emerging and Selected Topics in Circuits and
Systems, IEEE Journal on, vol. 2, no. 3, pp. 493501, 2012.
[9] Y. Sun, K. L. Chan, and S. M. Krishnan, ECG signal conditioning by
morphological ltering, Computers in Biology and Medicine, vol. 32,
no. 6, pp. 465479, Sept. 2002.
[10] C. Meyer and H. Keiser, Electrocardiogram baseline noise estimation
and removal using cubic splines and state-space computation techniques, Computers and Biomedical Research, vol. 10, no. 5, pp. 459
470, Oct. 1977.
[11] R. Braojos, G. Ansaloni, D. Atienza, and F. Rincon, Embedded realtime ecg delineation methods: A comparative evaluation, in Bioinformatics Bioengineering (BIBE), 2012 IEEE 12th International Conference on, Nov 2012, pp. 99104.
[12] F. Rincon, N. Boichat, D. Atienza, and N. Khaled, Wavelet-based ECG
delineation on a wearable embedded sensor platform, in Wearable and
Implantable Body Sensor Networks, 2009. BSN 2009. Sixth International
Workshop on, June 2009, pp. 256261.
[13] Y. Sun, K. Luk Chan, and S. Muthu Krishnan, Characteristic wave
detection in ECG signal using morphological transform, BMC Cardiovascular Disorders, vol. 5, pp. 17, Sep 2005.
[14] R. Braojos, G. Ansaloni, and D. Atienza, A methodology for embedded classication of heartbeats using random projections, in Design,
Automation Test in Europe Conference Exhibition (DATE), 2013, March
2013, pp. 899904.
[15] D. Achlioptas, Database-friendly random projections: JohnsonLindenstrauss with binary coins, Journal of Computer and System
Sciences, vol. 66, no. 4, pp. 671 687, June 2003.
[16] H. Mamaghanian, N. Khaled, D. Atienza, and P. Vandergheynst, Compressed sensing for real-time energy-efcient ecg compression on wireless body sensor nodes, Biomedical Engineering, IEEE Transactions
on, vol. 58, no. 9, pp. 24562466, Sept 2011.
[17] M. F. Duarte, M. B. Wakin, and R. G. Baraniuk, Fast reconstruction
of piecewise smooth signals from incoherent projections, in Signal
Processing with Adaptative Sparse Structured Representations (SPARS
05), November 2005.
[18] R. Braojos Lopez, I. Beretta, G. Ansaloni, and D. Atienza Alonso,
Hardware/Software Approach for Code Synchronization in Low-Power
Multi-Core Sensor Nodes, in Proceedings of the IEEE/ACM 2014
Design Automation and Test in Europe (DATE) Conference, vol. 1, no. 1.
IEEE/ACM Press, 2014, pp. 5055.
[19] J. Constantin, A. Dogan, O. Andersson, P. Meinerzhagen, J. Rodrigues,
D. Atienza, and A. Burg, Tamarisc-cs: An ultra-low-power applicationspecic processor for compressed sensing, in VLSI and System-onChip (VLSI-SoC), 2012 IEEE/IFIP 20th International Conference on,
Oct 2012, pp. 159164.
[20] H. Gesche, D. Grosskurth, G. Kuchler, and A. Patzak, Continuous blood
pressure measurement by using the pulse transit time: comparison to a
cuff-based method, European Journal of Applied Physiology, vol. 112,
no. 1, pp. 309315, Jan 2012.
[21] J. M. Sola 12 i Caros, Continuous non-invasive blood pressure estimation, Ph.D. dissertation, ETH Zurich, 2011.
[22] P. Laguna, R. Jane, O. Meste, P. Poon, P. Caminal, H. Rix, and
N. Thakor, Adaptive lter for event-related bioelectric signals using an
impulse correlated reference input: comparison with signal averaging
techniques, Biomedical Engineering, IEEE Transactions on, vol. 39,
no. 10, pp. 10321044, Oct 1992.
[23] K. Pilt, K. Meigas, J. Lass, M. Rosmann, and J. Kaik, Adaptive impulse
correlated lter (aicf) improvement for photoplethysmographic signals,
in Engineering in Medicine and Biology Society, 2008. EMBS 2008. 30th
Annual International Conference of the IEEE, Aug 2008, pp. 273276.
[24] SmartCardia. [Online]. Available: http://smartcardia.com/
[25] F. Rincon, P. Grassi, N. Khaled, D. Atienza, and D. Sciuto, Automated
Real-Time Atrial Fibrillation Detection on a Wearable Wireless Sensor
Platform, in Proceedings of 34th IEEE Annual International Conference
of the Engineering in Medicine and Biology Society (EMBC 2012), 2012,
pp. 24722475.
[26] M. Vilhena Toledo and J. Junqueira, LuizFernando, Cardiac autonomic
modulation and cognitive status in alzheimers disease, Clinical Autonomic Research, vol. 20, no. 1, pp. 1117, Feb. 2010.
[27] O. Blanke, Multisensory brain mechanisms of bodily selfconsciousness, Nature Reviews Neuroscience, vol. 13, no. 8, pp.
556571, 2012.