NETWORK-AWARE BIOMEDICAL SIGNAL PROCESSING

Loss Concealment or Loss Awareness

Beatriz Barros

, Ana Aguiar

and Daniel E. Lucani

Instituto de Telecomunicac¸

oes, Departamento de Engenharia e Gest

ao Industrial, Faculdade de Engenharia da

Universidade do Porto, Rua Dr. Roberto Frias s/n, Porto, Portugal

Instituto de Telecomunicac¸

oes, Departamento de Engenharia Electrot

ecnica e de Computadores, Faculdade de

Engenharia da Universidade do Porto, Rua Dr. Roberto Frias s/n, Porto, Portugal

Keywords:

Biomedical signal processing, Networking, Loss concealment.

Abstract:

The development of biomedical signal processing algorithms typically assumes that the data can be sampled

at an uniform rate and without loss of samples. Although this is a valid assumption for Holter applications

or clinical testing, these assumptions become questionable in the presence of remote monitoring of patients

through inherently lossy communication networks. The task for the networking engineers has been to create

better, more reliable protocols to avoid packet losses from affecting the signal processing algorithms. However,

inherent constraints from resource-constrained devices and lossy networks used for remote monitoring make

this objective infeasible in many situations. Given irreparable losses due to data transmission, this paper

poses the following questions: (i) how would the current algorithms react to losses, and (ii) what alternatives

are available to still guarantee reliable monitoring and detection of emergency events. For the latter, we

consider two options: the use of current algorithms after a loss concealment stage, and the design of loss aware

algorithms. We argue that a joint design of network protocols and signal processing algorithms is instrumental

for providing reliable biomedical monitoring. We propose a simple, yet powerful model of the network under

a variety of packet loss channels as well as data packetization mechanisms. Extensive numerical results are

provided for addressing question (i), focusing on the sensitivity and positive predictivity of standard ECG

algorithms under a variety of network scenarios. We use the MIT-BIH arrhythmia database and simple loss

concealment mechanisms and show that even small percentages of packet losses can have a signiﬁcant impact

on a algorithm’s performance.

1 INTRODUCTION

Conventional wisdom for biomedical signal process-

ing algorithms says that the data, although noisy, is re-

ceived without losses and typically at an uniform sam-

pling rate. Although a valid assumption for the case

of Holter devices and clinical observation, a stronger

challenge is faced when considering transmission of

the data through inherently lossy networks. The de-

mands on the networking community have been to

provide protocols and mechanisms that support a loss-

less transmission of the data. For example, Ref. (Ale-

sanco and Garca, 2010) provides analysis and simula-

tion results using the real-time ECG protocol (RETP),

an alternative to TCP/IP for reliable transmission of

all data samples, studying networks with losses and

its effects upon delay and mean opinion score (MOS)

value for clinical assessment. In this sense, the devel-

opment of signal processing algorithms for feature ex-

traction and communication protocols for data gather-

ing are typically considered as separate tasks.

Although valid in Holter devices and during clini-

cal observation, remote and online monitoring of pa-

tients introduces new challenges and constraints to the

system. A closer inspection reveals that biomedical

devices for remote data gathering rely on resource-

constrained, bandwidth-limited, wireless devices and

unreliable connectivity, e.g., (Hu et al., 2009) (Pan-

dian et al., 2008). These limitations may prevent the

system from providing perfect reliability on the data

transmission process, which translates in samples be-

ing lost irreparably.

Although remote and automatic monitoring of pa-

tients is usually seen as a clear option to reducing

medical costs, if packet losses produce an increase of

false positives in emergency scenarios i) the econom-

ical beneﬁts may vanish as personnel and equipment

may be mobilized unnecessarily, or ii) the system may

318

Barros B., Aguiar A. and E. Lucani D..

NETWORK-AWARE BIOMEDICAL SIGNAL PROCESSING - Loss Concealment or Loss Awareness.

DOI: 10.5220/0003876503180325

In Proceedings of the International Conference on Health Informatics (HEALTHINF-2012), pages 318-325

ISBN: 978-989-8425-88-1

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

not be deployed in practice due to lack of reliability.

Our take is that a design of signal processing algo-

rithms that are aware of lossy transmissions is funda-

mental for guaranteeing reliable monitoring and be-

come pivotal towards a joint design of network proto-

cols and processing algorithms. We propose two dif-

ferent approaches to the problem from the signal pro-

cessing viewpoint: loss concealment and loss aware-

ness. The former relies on using loss concealment

mechanisms, e.g., prediction algorithms to determine

missing samples, as well as appropriate interleaving

and packetization of the samples in order to hide the

losses from current algorithms. The beneﬁt of this ap-

proach is that current algorithms may be kept without

change. The main drawback of this approach is that

the signal processing algorithms do not differentiate

between actual samples and predicted samples, i.e.,

they are not able to exploit this knowledge.

Although previous work has considered loss con-

cealment as an option, e.g., (Theera-Umpon et al.,

2008), (Prieto-Guerrero et al., 2007), it has done

so without exploiting additional dimensions pro-

vided by the communication architecture. For exam-

ple, (Theera-Umpon et al., 2008) considers that losses

happen during a given interval, which could be a valid

assumption considering that a data packet is likely

to carry several samples. However, interleaving the

samples prior to generating the data packets, i.e., re-

arranging the samples across several packets so that

they are non-contiguous, provides a powerful option

to spread the lost samples. This can allow the con-

cealment mechanisms to operate more efﬁciently as

the number of contiguous samples missing may de-

crease dramatically.

Loss awareness calls for a re-design of biomedial

signal processing algorithms to consider one or sev-

eral of the following: i) variable (and random) sam-

pling, ii) conﬁdence values for samples providing full

conﬁdence for actual samples and lower conﬁdence

for predicted/estimated samples, iii) incorporate net-

work statistics as key parameters for the algorithms,

iv) control of the data source to request speciﬁc lost

samples (but possibly not all) or reduce sampling rate.

To attain optimal performance of the two ap-

proaches, a joint design of network protocols and sig-

nal processing algorithm is required. In other words, a

joint design of network protocols that can adapt to re-

quirements of the algorithms and algorithms that can

adapt to network characteristics and effectively con-

trol the protocols. This new paradigm for network-

aware biomedical algorithm design shall be instru-

mental to making effective, remote, and low cost

biomedical monitoring a reality.

Aiming at a deeper understanding of how packet

losses affect vital signs processing algorithms as well

as how concealing packet losses can improve this be-

havior, we make the following contributions:

• Requirement and Challenges: We provide a dis-

cussion of requirements and challenges of remote,

online biomedical monitoring emphasizing dif-

ferent possible sources of sample losses, which

may not only be limited to network congestion or

losses in wireless channels but that can be gener-

ated due to active security attacks.

• Evaluation Framework: Based on key communi-

cations, network, and loss concealment parame-

ters and building blocks, we propose a simple yet

detailed model to characterize algorithm perfor-

mance. This evaluation model is adaptive, in the

sense that different versions of the proposed build-

ing blocks can be incorporated seamlessly.

• Numerical Analysis: We evaluate ECG algo-

rithms in terms of sensitivity and positive predic-

tivity under a variety of network and loss con-

cealment scenarios. The MIT-BIH data base and

the ecgpuwave algorithm are used as an interest-

ing example. The results clearly illustrate that ad-

vanced loss concealment mechanisms or, alterna-

tively, loss-aware vital signs algorithms are a must

in networks that cannot guarantee delivery of ev-

ery sample. This case is extremely relevant for re-

mote monitoring of patients using simple wireless

devices, e.g., wireless sensor networks, wireless

body-area networks. Our preliminary results con-

ﬁrm that concealment of lost samples is possible

in a limited number of scenarios (low packet loss

rates, low number of samples per packet) even

with simple loss concealment algorithms, which

implies that no modiﬁcation of current algorithms

is needed after an initial loss concealment stage.

However, loss concealment becomes insufﬁcient

in more typical wireless network scenarios.

• Alternative Algorithms: We provide a discus-

sion of alternative approaches to joint network

protocol-signal processing algorithm design as

well as loss concealment mechanisms that may be

promising in our applications.

2 REQUIREMENTS AND

CHALLENGES

Remote and automatic monitoring of patients sets a

series of design requirements and inherent challenges

driven from them.

• Data Rates. Depend considerably on the vital

signs that are being collected and linked to the

NETWORK-AWARE BIOMEDICAL SIGNAL PROCESSING - Loss Concealment or Loss Awareness

319

sampling rates. For example, a 3-lead ECG mon-

itor sending 2 channels sampled at 250 Hz with

8-bit samples has a raw data rate of 4 kbps. How-

ever, a 12-lead ECG monitor sending 8 channels

sampled at 1000 Hz with 12-bit samples, gener-

ates a raw data rate of 96 kbps. Although, the

former is a much more common case for remote

monitoring, we emphasize that this is the source

data rate. The actual requested data rate to the net-

work shall depend on a variety of factors, of which

packetization (i.e., how many and which samples

are sent in each packet) plays the key role. Com-

munication protocols at different layers add head-

ers to each packet traversing the network for con-

trol and identiﬁcation purposes. Depending on the

protocols used, this header’s size could be signif-

icant. In many scenarios, it is in the order of tens

of bytes. Clearly, if only one 8-bit sample is sent

per packet in the 3-lead case and if overall header

is 20 bytes per packet, the rate from the perspec-

tive of the network is actually 84kbps. Sending

several samples per packet reduces the overhead,

but at the cost of i) additional delay in the trans-

mission of the samples, ii) higher impact to the

system due to a single packet loss.

• Time Criticality. Remote monitoring of patients is

time critical in the sense that the samples should

be received within a certain time frame to be able

to predict, prepare, and/or react to critical situa-

tions. However, a delay of several seconds is still

acceptable from this perspective (Alesanco and

Garca, 2010).

• Light-weight, Cost-effective Solutions. One of the

key arguments for remote, online monitoring is

to reduce the economical and qualiﬁed human re-

sources cost of monitoring patients. Devices and

services provided require a cost-effective design,

which may limit the storage and processing capa-

bilities of these devices. These limitations con-

stitute one of the key motivations for considering

network-aware algorithms which can leverage re-

sources in an effective fashion and where more

computationally intensive algorithms are not im-

plemented in the end devices. This requirement

also motivates the use of off-the-shelf network de-

vices with standard protocols, e.g., Bluetooth or

ZigBee, which introduce limitations of their own

in terms of maximum data rate per user, reliability,

transmission range, protocol overhead, or trans-

mission band.

• Reliable Monitoring. Although the objective of

reliable monitoring of patients has been translated

thus far into providing reliable, lossless transmis-

sion of data, this need not be the case. The ulti-

mate objective of the system is to deliver a reli-

able monitoring service to patients and medical

personnel. In fact, this ultimate objective may

be compromised if the system is unable to adapt

to varying network conditions, which may ren-

der present sampling settings unserviceable, or if

the loss of samples compromises the accuracy of

event detection algorithms and/or triggers unnec-

essary alerts.

• Security and Privacy. Vital signs monitoring deal

with sensitive and private information, which im-

pose stringent requirements in terms of cyber-

security and consumer privacy. However, current

network architectures consider devices with lim-

ited processing capabilities close to the individ-

ual (data source), which calls for light-weight se-

curity solutions (Stuart et al., 2008). The archi-

tecture must ensure that unauthorized agents are

unable to access the data collected and that ac-

tive attacks capable of disrupting data collection

are mitigated. From the perspective of design of

network-aware signal processing algorithms, ac-

tive attacks can become a source of packet losses

in the system and one that could be relevant even

if the architecture has enough resources to guaran-

tee delivery of data packets in normal operation.

3 SYSTEM MODEL

We model the transmission of a vital signs data stream

across a network as depicted in Figure 1. The raw

data stream received from the vital signs monitor is

fed into a transmitter, which in its simplest form pre-

pares the data for sending across the network by split-

ting it into packets, containing one or more samples

each. These packets are passed through the network,

which causes some packets to be lost. At the receiver,

the stream is re-built from the data in the received

packets, whereby the samples corresponding to lost

packets are replaced. The re-built stream is then fed

into a signal processing algorithm that extracts use-

ful information from the data. For our evaluation,

we use the well-established heart beat detection al-

gorithm ecgpuwave (Pan and Tompkins, 1985). This

model enables us to assess the impact of the most rele-

vant parameters of the transmission of a signal across

a packet data network on the performance of a widely

known signal processing algorithm that assumes by

design that no data losses occur.

The packetization step plays a relevant role, since

for a same packet loss rate in the network more sam-

ples are lost when each packet carries a larger number

of samples. The drawback of transmitting a single

HEALTHINF 2012 - International Conference on Health Informatics

320

Erasure'

Gilbert-Elliot'

Interpolate'

“0”'

Repeat'Last'

ecgpuwave'

Packe=sa=on'

Interleaving'

Sender'

Error'Concealment'

Heart''

wave'

Receiver'

Network'

Heart''

beat'

Figure 1: System Model.

sample per packet is efﬁciency, since each data packet

sent on the network carries overhead due to several

layers of in-band signaling. Additionally, some net-

work may require some overhead per packet for ac-

cessing the resources. Hence, there is a trade-off be-

tween efﬁcient network utilization and the effects of

eventual packet losses.

The pattern of network error occurrence can also

signiﬁcantly inﬂuence the impact of network losses.

The packet losses may occur randomly, as in Blue-

tooth links, or correlated in time, as in WiFi or cellu-

lar links. The effects of correlated packet losses can

be more damaging than the effects of random losses,

because correlated losses can cause the deletion of a

signiﬁcant part of the signal that is to be processed.

In the speciﬁc case of heart wave and heart beat de-

tection, the loss of various packets containing eventu-

ally more than one sample can cause a QRS complex

to disappear. While this may not be very serious for

heart beat detection, it may have signiﬁcant impact in

subsequent algorithms, like arrhythmia detection.

Finally, the receiver must decide how to replace

the samples of the missing packets when rebuilding

the data stream, since periodic samples are expected.

The simplest approach is replacing the missing sam-

ples with zero. The research ﬁeld of audio stream-

ing offers a good starting point to look for more in-

telligent ways to deal with packet losses (Perkins

et al., 1998). Several loss concealment techniques

have been studied over the years, of which we ad-

dress one sender-based technique, interleaving, and

two receiver-based techniques, sample repetition and

interpolation of transmitted state, described in detail

below.

4 PACKETIZATION OVERHEAD

The trade-off for risking the loss of multiple samples

per packet is that it is more efﬁcient to transmit mul-

tiple samples in a packet in terms of network resource

usage, with side effects on the energy required for

transmission. We deﬁne as overhead the ratio of ap-

plication data, in this case bits of ECG samples, to

the total transmitted data, i.e. including all the in-

band signaling of the various protocol layers involved

in the communication. To gain a better perception of

the values involved, Table 1 shows the overhead for

several packetizations used for transmission of MIT-

BIH 11 bit samples using UDP (8 Byte header) over

IPv4 over WiFi, or Bluetooth, both commonly used

technologies for the wireless transmission of ECG

data. For the Bluetooth calculations, we consider

SCO packet types and HV1 proﬁle for 1 sample per

packet, and multiple HV2 packets for all other pack-

etizations. The values in the table clearly show the

advantage of transmitting multiple samples in each

network packet, and only for more than 20 samples

in each network packet does the overhead decrease

below 50%.

NETWORK-AWARE BIOMEDICAL SIGNAL PROCESSING - Loss Concealment or Loss Awareness

321

Table 1: Network overhead for the chosen packetizations

and commonly used network technologies.

IP over WiFi Bluetooth

1 92.6 88.8

2 86.3 87.6

5 71.5 68.9

10 55.6 67.5

20 38.5 34.9

50 20.1 16.4

5 NETWORK LOSS MODELS

The packet loss models of a network are characterized

by the average packet loss rate and the time correla-

tion between those losses. These models depend on a

variety of parameters, ranging from the medium used

(wireless vs. wired) to the topology and size of the

network. Hence different types of networks show dif-

ferent packet loss behaviours.

Two models are commonly used for the generic

characterization of network packet losses, namely the

random independent loss model and the Gilbert-Elliot

model. The ﬁrst models packet losses as independent

events, and is characterized by the average packet loss

rate. The latter models packet losses as a two-state

markov chain, assuming a dependency between loss

events of consecutive packets. It is characterized by

the loss probabilities in each state and the transition

probabilities between the states.

6 LOSS REPAIR AND

CONCEALMENT TECHNIQUES

At the receiver, the arriving packets are buffered to re-

move variations in the transmission latency across the

network (jitter in networking jargon). Then, the data

in the packets is used to rebuild a stream of periodic

heart wave samples. When packets are lost, the cor-

responding missing samples must be replaced, so that

the timing of the rebuilt stream is not disrupted and

it contains periodic samples as expected by the sub-

sequent algorithms. The approach taken to deal with

missing samples at this step is called error conceal-

ment, a name that we borrow from the vocabulary of

voice streaming. Some of these mechanisms rely on

actions on the part of the sender to better empower

the receiver at this step, and some others are purely

receiver-based.

In sender-based techniques, the sender processes

the information to be sent such that the transmission

becomes resilient to some amount of errors. An alter-

native is that the sender re-transmits missing data trig-

gered by a request from the receiver, although this in-

volves at least one full round-trip delay and additional

buffering at the sender. The sender-based technique

that we will focus on is called interleaving, although

additional techniques can be tested in our framework.

Interleaving consists in sending non-adjacent samples

(in time) in each data packet and separating time-

adjacent samples across several data packets. On the

loss of a packet that carries more than one sample, the

missing samples are not successive, thus transforming

the loss of a large chunk of data (a packet) into the loss

of the same amount of samples but spread across the

number of packets involved in the interleaving pro-

cess. The drawback of this technique is the delay that

it introduces and the corresponding buffering require-

ments both at the sender and the receiver: enough data

must be stored at the sender before generating a batch

of packets.

As receiver-based error concealment techniques

we illustrate two insertion techniques, substitution

with ”0” and with the previous sample, and one in-

terpolation technique. In the ﬁrst case, any missing

sample is replaced with a value equivalent to 0 or with

the last correctly received sample, respectively (”0” or

”Repeat Last” in Figure 1). In the case of interpola-

tion, each sample is replaced with the linear interpo-

lation between the previous and following correctly

received samples. Linear interpolation is more costly

in terms of computation than simple insertion tech-

niques.

7 METHODOLOGY AND

PARAMETERS

We carried out extensive simulations to assess the im-

pact of network transmission losses on the accuracy of

the well-known heart beat detection algorithm, ecg-

puwave (Pan and Tompkins, 1985). We used various

packetizations and varied the quality of network trans-

mission, i.e., the average packet loss rate. For each

case, we used one of the three error concealment tech-

niques described above. Additionally, we used inter-

leaving with the basic error concealment case (”0”).

We set off simulating the case of uncorrelated trans-

mission losses, which is the best case scenario. That

is, correlated losses will cause more damage to the

rebuilt data stream, so that the results shown here rep-

resent the best expected performance in case of a spe-

ciﬁc amount of network losses. The parameters used

in the study are summarized in the Table 2.

We evaluated the impact on the performance of

the heart beat detection algorithm by comparing the

sensitivity and the positive predictive value (PPV)

HEALTHINF 2012 - International Conference on Health Informatics

322

(a) ”0”. (b) ”0” with interleaving.

Figure 2: Sensitivity and PPV with interleaving for random transmission losses.

Table 2: Parameter space explored for the results.

Packetization [# samples] {1, 2, 5, 10, 20, 50}

Interleaving On/Off

Packet loss rate ∈ [0.1; 50]%

Error concealment ”0”, ”Last”, ”Interpolation”

obtained for an ECG stream rebuilt after suffering

packet losses with the values obtained for the same

ECG stream that did not suffer losses. The sensitiv-

ity expresses the rate of all existing QRS complexes

that are detected, whereas the PPV expresses the rate

of detected QRS complexes that do not exist in the

original ECG stream. We use the MIT-BIH arrhyth-

mia database (Moody and Mark, 1990) and the ecg-

puwave implementation in the PhisioToolkit (Moody

et al., 2000).

8 RESULTS

The reference values for comparing the impact of

samples losses are the speciﬁcity and PPV of the heart

beat detection algorithm without any losses, which

are 99.1% and 99.33%, respectively. Figure 2 shows

the speciﬁcity and PPV for simple replacement of

missing samples with ”0”, without and with interleav-

ing, in the presence of uncorrelated network errors.

Simply substituting the missing samples causes an er-

ratic behavior of the heart beat detector in terms of

sensitivity and the PPV degrades to values below 90%

even for low amount of losses. This large amount of

falsely detected beats is caused by the large amount of

steep slopes introduced when replacing missing sam-

ples with values close to 0. Spreading the missing

samples in time, achieved by interleaving samples be-

fore packetization at the sender, maintains the sensi-

tivity and PPV above 90% for random network losses

up to 1%, demonstrating the potential of this tech-

nique.

Figure 3 shows the results for insertion with pre-

vious correct sample and linear interpolation without

interleaving. Oppositely to replacing missing samples

with ”0”, replacing each sample with the last correctly

received value shows negligible performance degra-

dation (sensitivity and PPV larger than 99%) of the

heart beat detection algorithm for up to 1 sample per

packet and up to 50% network packet losses. On the

other hand, as expected, larger packetizations tolerate

less network losses, namely 2% for 10 sample pack-

etization and only 0.5% for 20 samples per packet.

Finally, replacing missing samples with a value re-

sulting from interpolation using the nearby correctly

received samples provides only 0.5% better sensitiv-

ity and PPV than insertion with ”Last” for more than

5% losses, despite the additional computational com-

plexity involved in the latter concealment technique.

Table 3 shows the maximum usable packetization

NETWORK-AWARE BIOMEDICAL SIGNAL PROCESSING - Loss Concealment or Loss Awareness

323

(a) ”Last”. (b) Interpolation.

Figure 3: Sensitivity and PPV for random transmission losses.

Table 3: Maximum packetization to guarantee sensitivity

and PPV above 99%.

Packet Loss Rate [%] ”0” ”Last” Interpolation

0.1 - 50 50

0.2 - 20 20

0.5 - 20 10

1 - 10 10

2 - 10 5

5 - 2 5

10 - 1 2

20 - 1 2

50 - 1 1

to guarantee sensitivity and PPV above 99% for each

error concealment technique used, showing that none

of the studied techniques performs better than the

other in all situations. Speciﬁcally, using ”Last” for

error concealment is more efﬁcient for networks with

low packet losses whereas the additional complexity

of linear interpolation pays off for less reliable net-

works. From the perspective of network efﬁciency,

however, the combinations of packetizations and er-

ror concealment techniques for which the well-known

ecgpuwave algorithm would perform acceptably im-

ply transmission overheads of more than 50% (see ta-

ble 1), and are extremely inefﬁcient and energy costly.

9 DISCUSSION

Although our results show that it is feasible to trans-

mit ECG data across wireless links with currently

available technologies for post-processing by signal

processing algorithms, e.g., heart beat or arrhythmia

detection, they do not provide efﬁcient and reliable

operation. Due to practical considerations, like pa-

tient comfort and usability, ECG monitoring devices

are usually resource constrained both in terms of pro-

cessing and available energy. An efﬁcient use of these

resources is then critical in ECG monitoring system

design. Our evaluation clearly shows that this is

not possible today with state-of-the-art technologies.

Next, we describe three ﬁelds that we believe should

be explored targeting speciﬁcally the transmission of

ECG data for automatic post-processing solutions.

Firstly, the framework that we present and the as-

sociated parameter space needs to be further evalu-

ated, and the fundamental trade-offs involved in the

choice of transmission technologies, packetizations,

loss concealment techniques, etc, must be studied for

varied network scenarios. Emphasis shall be put in

a joint optimization of the parameters involved in the

transmission of ECG data to achieve highest network

and energy efﬁciency, conditioned on guaranteeing

minimum performance levels of the signal processing

algorithms.

HEALTHINF 2012 - International Conference on Health Informatics

324

Another line of research that will produce relevant

insights is the development of loss concealment tech-

niques to efﬁciently repair the ECG stream at the re-

ceiver, with or without the cooperation of the sender.

We speciﬁcally envision the use of linear prediction,

Kalman ﬁlters or other adaptive ﬁlters to reconstruct

the ECG signal. Although this may not provide a

strong improvement over simpler techniques for the

purpose of heart beat detection, it shall play a sig-

niﬁcant role in the performance of subsequent, more

ellaborate algorithms, like arrhythmia detection.

Taking a holistic view of the problem, we fur-

ther propose the development of network-aware sig-

nal processing algorithms that are either resilient or

can adapt to certain levels of sample loss. We envision

the application of non-uniform sampling mechanisms

and results from the ﬁeld of compressed sensing.

10 CONCLUSIONS

We address the often ignored problem of transmis-

sion of biomedical signal data across networks for re-

mote processing, proposing a framework that models

the relevant building blocks of such a system. We

use the framework to perform an initial numerical

evaluation of the impact of uncorrelated random net-

work packet losses in the performance of the well-

known heart beat detection algorithm ecgpuwave us-

ing the MIT-BIH database. Our results show that

1) packet losses cause signiﬁcant degradation of the

heart beat detection algorithms; 2) simple loss con-

cealment techniques, like insertion of last known sam-

ple and linear interpolation, signiﬁcantly reduce the

impact of network losses, but their performance de-

pends on the packetization used; 3) packetization con-

stitutes an important parameter to choose the trade-

off amongst network and energy efﬁciency and im-

pact of packet losses; 5) there is not one combination

of packetization and loss concealment technique that

performs best for all network scenarios studied.

As a consequence of these ﬁndings, we identify

the need to further research the transmission of data

from biomedical signals across networks and propose

to deepen the understanding of the applicability of

three ﬁelds of research to biomedical signal trans-

mission and processing. Namely, 1) the joint opti-

mization of transmission parameters, 2) the develop-

ment of advanced loss concealment techniques, like

Kalman ﬁlters and linear prediction, and 3) the devel-

opment of loss-resilient signal processing algorithms,

leveraging results from compressed sensing or non-

uniform sampling theory.

ACKNOWLEDGEMENTS

The authors thank Miguel Coimbra and Can Ye for

fruitful discussions.

This work was supported by FCT (Fundac¸

para a Ci

encia e a Tecnologia) through the

VR (Vital Responder) project (within the

CarnegieMellon—Portugal program. ref. CMU-

P/CPS/0046/2008).

REFERENCES

Alesanco, A. and Garca, J. (2010). Clinical assessment

of wireless ecg transmission in real-time cardiac tele-

monitoring. IEEE Transactions on Information Tech-

nology in Biomedicine, 14(5):1144–1152.

Hu, X., Wang, J., Yu, Q., Liu, W., and Qin, J. (2009).

A wireless sensor network based on zigbee for

telemedicine monitoring system. In The 2nd Interna-

tional Conference on Bioinformatics and Biomedical

Engineering, pages 1367–1370.

Moody, G. and Mark, R. (1990). The mit-bih arrhythmia

database on cd-rom and software for use with it. In

Computers in Cardiology, pages 185–188.

Moody, G., Mark, R., and Goldberger, A. (2000). Phys-

ionet: a research resource for studies of complex phys-

iologic and biomedical signals. In Computers in Car-

diology, pages 179–182.

Pan, J. and Tompkins, W. (1985). A real-time qrs detection

algorithm. IEEE Transactions on Biomedical Engi-

neering, 32(3):230 –236.

Pandian, P. S., Safeer, K. P., Gupta, P., Shakunthala, D. T.,

Sundersheshu, B. S., and Padaki, V. C. (2008). Wire-

less sensor network for wearable physiological moni-

toring. Journal of Networks, 3(5).

Perkins, C., Hodson, O., and Hardman, V. (1998). A sur-

vey of packet loss recovery techniques for streaming

audio. IEEE Network, 12(5):40 –48.

Prieto-Guerrero, A., Mailhes, C., and Castani

e, F. (2007).

Lost sample recovering of ecg signals in e-health ap-

plications. In The 29-th Annual International Confer-

ence of the IEEE EMBS, pages 31–34.

Stuart, E., Moh, M., and Moh, T. S. (2008). Privacy and

security in biomedical applications of wireless sen-

sor networks. In First International Symposium on

Applied Sciences on Biomedical and Communication

Technologies, pages 1–5.

Theera-Umpon, N., Phiphatkhunarnon, P., and Auephan-

wiriyakul, S. (2008). Data reconstruction for miss-

ing electrocardiogram using linear predictive coding.

IEEE International Conference on Mechatronics and

Automation (ICMA), pages 638–643.

NETWORK-AWARE BIOMEDICAL SIGNAL PROCESSING - Loss Concealment or Loss Awareness

325