Page 1
Loss Aware Sample Packetization Strategy for
Improvement of Body Sensor Data Analysis
Ming Li1 Yu Cao
2 and B Prabhakaran
3
1Department of Computer Science California State University Fresno CA 93619 USA
2Department of Computer Science University of Massachusetts Lowell MA 02148 USA
3Department of Computer Science The University of Texas at Dallas Richardson TX 75080 USA
Email minglicsufresnoedu ycaocsumledu prabautdallasedu
Abstractmdash With limited bandwidth and high channel loss in
Wireless Body Area Networks (WBANs) it remains a
challenging issue to transmit time series body sensor data for
satisfactory data analytics In this paper we investigate the
interesting problem of sample packetization ie assembling
multiple samples in each packet before transmission over the
wireless link We first illustrate the ineffectiveness of the
default sequential packetization under channel loss Then we
propose Loss Aware Sample Packetization (LASP) a heuristic
loss aware sample packetization strategy to improve the data
quality for improvement of the sensor data analysis The core
idea of LASP is to minimize contiguous sample loss during the
data transmission thereby significantly increasing the accuracy
of data recovery at the aggregator side Extensive simulations
are conducted to evaluate the effectiveness of the proposed
approach on data recovery quality as well as data analysis
accuracy Results show that LASP yields a nicer sample loss
pattern therefore significantly improve the data recovery
quality Index TermsmdashData analysis sample packetization time series
data body area networks
I INTRODUCTION
Recently there has been increasing investigation on a
new type of network architecture generally known as
Body Sensor Networks (BSNs [1]) or Wireless Body
Area Networks (WBANs [2]) attributed to the exciting
advances in design of lightweight small size ultra-low
power and intelligent monitoring wearable sensors
Compared with existing technologies such as WLANs
BSNs enable the wireless communications inside or
around the vicinity of human body therefore further
extending the desirable concept of pervasive wireless
computing into a completely new level Key applications
that may benefit from this new technique include remote
health monitoring sports training and entertainments
Extensive research [3]-[12] has been conducted on data
analytics of body sensor data However due to the
limited capacity of body sensor networks and the channel
fading packet loss occurs frequently during data
transmission which in turn imposes significant impact on
Manuscript received March 1 2015 revised November 18 2015
This work was supported by the US National Science Foundation
under Grant No 1229213 Corresponding author email minglicsufresnoedu
doi1012720jcm1011851-858
the accuracy of data analysis Existing approaches to
address network loss include ensuring Quality of Service
(QoS) support for specific data streams [13]-[15] as well
as increasing network throughput [16]-[18] These
strategies although effective in general have not fully
considered the impact of packet loss on analysis accuracy
of BSN sensor data
Basically the accuracy of data analysis under the
scenario of wirelessly transmitted sensor data depends on
two factors 1) network delivery ratio ie how many
samples are successfully delivered through the network to
the receiver which is in turn determined by network
capacity channel stability and traffic load 2) data
recovery ratio ie among all lost samples how many of
them can be recovered to a level that approximates the
original samples While network delivery ratio can be
largely improved by QoS strategies data recovery has to
be addressed by closely looking at data characteristics
and specific analytics algorithms
In this paper we investigated the critical issue of
improving data recovery ratio to ensure the success of
data analysis First we analyzed the factors that affect
data recovery under BSN channel and determined the
appropriate system model and assumptions Then we
focused on the largely overlooked difference between
sample loss and packet loss and identified the problem of
sample packetization ie how to group samples in
different packets We found that packetization strategies
have significant impact on data recovery Given the
sample packet loss ratio it is possible to improve data
recoverability by avoiding certain ldquoworst caserdquo sample
grouping patterns For example the natural packetization
or the default sequential packetization which group
contiguous packets together has the high likelihood of
losing high importance samples Further we formulated
the problem of improving data recovery as finding a
pattern of sample packetization such that the number of
continuous sample pairs between any two packets is
minimum This way lost samples can be largely
recovered with higher accuracy through estimation from
received neighboring samples Meanwhile we introduced
the concept of critical samples and defined it as samples
that cannot be recovered with satisfactory accuracy even
if immediate neighboring samples are received
Considering the challenge of time complexity of the
algorithm as well as the reassembly overhead of received
Journal of Communications Vol 10 No 11 November 2015
851copy2015 Journal of Communications
out of order samples we propose LASP (Loss Aware
Sample Packetization) an efficient heuristic approach
where a fixed pattern is followed for each sample group
of a specific size The fixed pattern has the advantage of
low number contiguous sample pairs and can be easily
applied at the receiver side for reassembling of samples
Then an additional reliability enhancement with
approach similar to selective repeat [19] for critical
samples was proposed for further improvement
II RELATED WORKS
Quality of Service (QoS) scheduling for streaming
body sensor data has been well investigated Zhou and Lu
[13] proposed BodyQoS a virtual MAC for quality of
service scheduling in BSNs The approach basically
measures the effective bandwidth and adaptively
allocates remaining resources to meet the QoS
requirements of applications A desirable feature of
BodyQoS is that it does not require the modification of
the underlying MAC layer implementation Fulford-Jones
and Malan [20] proposed CodeBlue an ad hoc
infrastructure for emergent medical care In this project
several types of body sensors (eg pulse oximeter
ECGEKG sensor) are individually connected to Zigbee
enabled radio transmitters which communicate with
access points directly Due to the ad hoc architecture and
the capability of self-organizing CodeBlue yields
scalability for network expanding and flexibility to
connect various wireless devices Jiang and Cao [21]
proposed CareNet an integrated wireless environment
used for remote health care systems CareNet offers
features such as high reliability and performance
scalability security and integration with web based portal
systems High reliability is achieved by using two-tier
architecture Younis and Akkaya [22] proposed
Distributed Queuing Body Area Network a MAC
protocol aiming at providing better QoS It uses a cross-
layer fuzzy rule based scheduling algorithm to optimize
MAC layer performance in terms of QoS and energy
efficiency
There have been considerable standardization efforts
during recent years Among many potential technologies
Zigbee and Bluetooth are most widely deployed ZigBee
is a very low power collision avoidance protocol
optimized for lower power sensors It has developed a
health care specific protocol and are compliant with all
IEEE 11073 devices as well as most other IEEE 802154
[23] wireless devices Bluetooth supports high-bandwidth
and many several existing devices and has a health care
compliant version defined but has very high power
requirements and uptime for the radios Bluetooth Low
Energy [24] is a new proposed system from Bluetooth
which will have lower energy requirements and still be
interoperable with Bluetooth Classic but details are still
forthcoming at this time Being aware of the unique
requirement of supporting a wide range of applications by
body area networks an IEEE 802156 [25] task force has
Channel is unreliable and may drop data
With multiple body sensors communicating with the
same aggregator channel may reach saturation status
and therefore dropping packets in the queue
Reliable transmission of all samples incurs high
protocol overhead for relatively small body sensor
samples
A Data Recovery and Critical Samples
With quite strict requirement on the completeness of
sensor samples by data analysis algorithms missing
samples at data aggregator should be recovered at the
maximally possible level Itrsquos possible to use selective
repeat [19] to send back to specific sensors and request
retransmission of all lost samples However this
approach has the disadvantage of higher overhead and is
not efficient for real-time analysis Instead it is desirable
to recover samples using received ones and try to achieve
satisfactory approximation Usually lost samples can be
recovered by estimation (eg linear interpolation) at the
Journal of Communications Vol 10 No 11 November 2015
852copy2015 Journal of Communications
been working on finalizing a Wireless Body Area
Network (WBAN) standard The new standard aims at
providing flexible and configurable energy efficient
MAC operations
Plenty of research has been conducted using body
sensors to monitor Activity of Daily Living (ADL) Two
survey papers [3] [4] give details on the state-of-art-of
research in the area of sensor network with inertial
sensors and their applications in healthcare and wellbeing
Due to the growing interest on body sensor-based ADL
recognition several sensor-based human activity datasets
[10] [11] were introduced in the last few years A most
recent paper published in 2014 ACM Computing Survey
[12] discuss the key research challenges that human
activity recognition shares with general pattern
recognition and identify those challenges that are specific
to human activity recognition This paper also describes
the concept of an Activity Recognition Chain (ARC) as a
general-purpose framework for designing and evaluating
activity recognition systems Most of the techniques in
the current state of the art are focusing on the data
analysis and pattern recognition using the raw sensor data
Very little research in data analytics area considers the
issues of developing new networking strategy to improve
the accuracy and speed of sensor data analytics
III SYSTEM MODEL AND ASSUMPTIONS
In a typical medical application multiple sensors such
as ECG EKG EMG EEG motion sensors and blood
pressure sensors send multimodal time series data to
nearby Data Aggregators (DAs) which can be a cell
phone watch headset PDA laptop or robot based on
the application needs Then through BluetoothWiFi
these data can be delivered remotely to physician side for
real time diagnosis or to medical database for record
keeping or to request for emergency For such systems
the following assumptions are made
data aggregator In addition for most sample recovery
methods estimation accuracy of a missing sample is
significantly higher when the immediate previous and
next samples are received
For formal representation we define data recovery
ratio as the percentage of lost samples that can be
estimated using received samples with satisfactory
approximation from the data analysis requirement In
addition we define critical samples as samples that
cannot be recovered with satisfactory accuracy even if
immediate neighboring samples are received Obviously
it is ineffective to recover critical samples with received
ones for data analysis Instead selective repeat can be
adopted
B Sample Loss vs Packet Loss
Existing efforts on QoS support help address packet
loss However from data analysis aspect it is sufficient
as long as enough samples are received or recovered
Therefore understanding the relationship between packet
loss and data recovery ratio is critical For this purpose it
is important to consider the largely overlooked difference
between packet loss and sample loss Given specific
packet loss ratio it is highly desirable that sample loss
pattern yields optimal data recovery ratio
Since samples are usually assembled in packets it is
common that multiple samples are lost at the same time
In FIFO queuing based transport protocols such as UDP
samples are packetized through sequential order by
default ie samples i i+1 i+2 hellip i+L
the sample packet with L being the packet size However
if a packet is dropped then all the L continuous samples
are lost In general recovering L continuous samples
incurs significant error especially when L is not small
As illustrated in Fig 1 it will be better off to distribute
adjacent samples in different packets such that if one
packet is lost each missing sample can still be estimated
by received adjacent samples with higher accuracy It
should be noted that how samples are assembled can be
largely manipulated without affecting network operation
correctness and efficiency We name this process sample
packetization [26] The issue is therefore how to identify
an appropriate pattern for sample packetization such that
the effect of loss on data analysis is minimized
Fig 1 Comparison of two basic packetization approaches
IV LOSS RESILIENT SAMPLE PACKETIZATION (LASP)
Understanding the potential of sample packetization on
improving data recovery ratio we can formulate the
problem statement as follows given a set of N collected
samples channel loss condition and a specific estimation
technique how to packetizesamples such that the overall
chance of missing continuous samples is minimized
Obviously with minimum continuous sample loss the
data sample recovery accuracy can be significantly
improved Then the solution for the aforementioned
problem is to find a pattern of sample packetization such
that the number of continuous sample pairs between any
two packets is minimal
A Possible Solutions and Challenges
To better understand the idea letrsquos take 16 samples
with index from 0 to 15 Fig 2 shows several possible
packetization patterns with corresponding numbers of
continuous sample pairs For simplicity it is assumed that
four samples are assembled in each packet naming P0 P1
P2 and P3 N1 represents the number of continuous
samples in the same packet N2 and N3 represent the
average and maximum numbers of contiguous sample
pairs between any two packets It is clear that approach (a)
is worst since it incurs N1 of 3 which indicates difficulty
of recovering samples even if one packet is lost Approach
(c) works slightly better than (b) since its N2 and N3 are
smaller indicating better performance when multiple
packet loss occurs
Basically given N samples finding the optimal
packetization pattern is computationally expense
especially when N is not small For low energy consuming
body sensors it is not acceptable to enumerate all
possibility for comparison Approach (c) in Fig 2 is
slightly better than (b) but finding such pattern at real-
time incurs overhead On the other hand one may think
that random pattern works well and may be able to
generate small number of continuous sample pairs
However even if it works well the reassembly of samples
according to their original order is quite difficult In order
to restore the order significant effort must be made to
deliver the pattern therefore incurring high network traffic
and leading to a waste of network bandwidth Therefore
we decide to take a heuristic solution based on approach
(b) described in the next subsection
Fig 2 Comparison of packetization approaches of 16 samples
B Proposed Approach
We proposed to design a heuristic packetization pattern
based on the approach (b) in Fig 2 Fig 3 illustrates the
For a total of 8 samples two packetization approaches are
considered
Sequential Packet 1 (0 1 2 3) Packet 2 (4 5 6 7) Alternate Packet 1 (0 2 4 6) Packet 2 (1 3 5 7)
Assuming the first packet gets lost the ldquoalternaterdquo approach yields better data analysis performance since it is easier
to recover recover samples 2 4 6 with good accuracy However
ldquosequentialrdquo approach will have difficulty since the lost samples
are too far away from received ones
Journal of Communications Vol 10 No 11 November 2015
853copy2015 Journal of Communications
-1 are put into
procedure of the proposed approach which is described
as follows
Step 1 for each collected sample group packetization
is performed using alternate pattern Assuming N samples
and M packets for each sample i (0ltiltN) the ID of
corresponding packet that it should be inserted into is
iM
Step 2 for the same group all critical samples are
identified A sample i is considered critical if |estimatedi-
originali| gt α originali Here the estimated value depends
on the specific estimation method and the values of the
neighboring samples α is a constant and is fixed as 025
in this paper
Step 3 all packets are sent to the aggregator At the
same time a special packet containing the IDs of all
critical samples is also sent
Step 4 upon the receipt of packets data aggregator re-
assembles received samples based on the same pattern
Therefore the kth sample in jth packet has the sample ID
of j+kN
Step 5 aggregator identifies lost critical samples by
comparing the IDs of the lost samples and the IDs of all
critical samples
Step 6 aggregator requests the sensor to retransmit
critical samples and ensure its successful delivery
Step 7 aggregator performs data recovery for other
lost samples before performing analysis
Fig 3 Comparison of packetization approaches of 16 samples
It should be noted that the proposed strategy is an
application layer approach and therefore is different from
the physical layer network coding [27] which aims at
coordination of multiple node transmission in order to
achieve higher network throughput especially in wireless
networks However the system performance will
potentially benefit from physical layer network coding if
implemented
C Data Quality Model
In this section we present a variance based data
quality model for the analysis of data recovery ratio We
define estimation variance (EV) as the relative difference
between the estimated sample values and original values
which can calculate it as
(1)
where di is the original value of the sample i and esti is the
estimated value of sample i from received samples
Further the total estimation variance is calculated as
(2)
(3)
where K is the total number of estimated samples in any
evaluation group Clearly smaller EV and AEV indicates
better data recovery which is critical for the success of
data analysis
Now letrsquos assume that a group of M independent
packets are transmitted For each packet j its total
estimation variance of all samples in the packet EV(j)
depends on the specific characteristics of the data samples
in the packet Thus for a given packet loss ratio of l the
expected total estimation variance is calculated as
If EV(j) is the same value (EV) for all packets which is
not usual in real scenarios due to the statistic dropping of
packets we then have that
Clearly the total EV depends on total number of
packets loss ratio and data characteristics Further
average expected estimation variance is now
(6)
That means if packets are independently recovered
then loss ratio does not affect the level of data recovery
However when two or more adjacent packets are lost it is
difficult to have independent data recovery therefore
higher loss ration does decrease the data recovery ratio
TABLE I EV VS DATA CHARACTERISTICS
Strategy Data Characteristics
sqrt linear square cubic quartic
Sequential 0075 0 039 090 140
Alternate 0003 0 002 006 012
To illustrate the effect of the packetization on data with
different characteristics we take a data segment that
follows different patterns and calculate the total EV for
the segment with two different packetization schemes
Each data segment contains 8 samples which are
packetized into two different packets with either the
ldquosequentialrdquo or ldquoalternaterdquo approach under the condition
that the second packet is lost Table I shows the result It
can be seen that as the curve deviates more from the linear
pattern higher variance is yielded for both ldquosequentialrdquo
and ldquoalternaterdquo approach However ldquoalternaterdquo approach
shows a much slower increase compared with ldquosequentialrdquo
V PERFORMANCE EVALUATION
We performed extensive simulations to study the
performance of the proposed packetization strategy For
simplicity we label the proposed strategy as LASP and
sequential packetization as ldquoExistingrdquo The following four
strategies were compared
Existing samples are packetized in order of creation
ie sequential
Collect
Samples
Perform
PacketizationSample
Assembling
Identify Critical
Samples
Request Lost
Critical Samples
Resend Critical
Samples
Data
Recovery
Sensor Side Aggregator
iiii ddestEV
i iEVEV
KEVAEV i i
M
m
m
j
mMm jEVllm
MEV
0 1
)()1(
EVlMmllm
MEVEV
M
m
mMm
0
)1(
EVlMEVlMAEV
Journal of Communications Vol 10 No 11 November 2015
854copy2015 Journal of Communications
(4)
(5)
Finally the Average Estimation Variance (AEV) is
calculated as
Existing+cs existing approach with lost critical
samples retransmitted
LASP proposed approach described in section 42
LASP-cs proposed approach but lost critical samples
are not retransmitted
We use the Body Sensor Network (BSN) Development
Kit v3 [28] developed by Prof Guang-Zhong Yang of
Imperial College to collect the accelerometer data from
users The BSN development kit v3 includes a USB
programming board a sensor board a battery board a
prototype board and a pair of the new BSN v3 nodes The
sensor board has a 3D accelerometer and a temperature
sensor In our experiments we recruit three subjects and
attach the sensor board to the arm of the subject Each
subject is asked to perform three types of Activity of Daily
Living (ADL) jogging sitting and walking The duration
of each activity is around 3 minutes The 3D
accelerometer data was collected when the subjects
performed the ADL activity We chose to use the
sampling rate as 10 Hz We also implemented linear
interpolation to calculate a sample directly between any
two neighboring ones the interpolation effectively yields
a 33 Hz sampling rate with the smoothness of the 16 Hz
rate Using the 10 Hz sampling rate with linear
interpolation yielded a rather smooth graph without a lot
of noise and yet it was fast enough that the peaks and
troughs of ADL were all present
Our evaluation consists of two parts In part 1 received
samples with different approaches are compared in term
of several key metrics In part 2 recovered data are
analyzed by various representative analysis algorithms for
comparison of the accuracy
Fig 4 Comparison of Average Estimation Variance (AEV)
A Sample Quality Analysis
For each received samples we evaluate how
satisfactory the data is recovered by comparing the
estimated values with original ones Fig 4 shows the
comparison of average estimation variance of lost samples
for x values of jogging data It can be seen that LASP has
only about 6-9 normalized variance on average while
existing approach yields an average of 15-17 Also
retransmission of critical samples does not help much for
existing approach mainly due to the high estimation
inaccuracy when a contiguous sample train gets lost Fig 5
compares number of lost samples with normalized
variance higher than 25 Itrsquos clear that LASP is much
better than existing approach even if critical samples are
not retransmitted This is due to the fact that LASP creates
a much nicer sample loss pattern as described earlier and
therefore minimizes the variance for lost samples
Fig 5 Number of samples with low recovery ratio
Fig 6 Comparison of data recovery
To provide a visual representation of how well LASP
improves data recovery 100 samples are extracted from a
sequence of 1043 samples for jogging Itrsquos clear that
LASP while having some variance creates a good
approximation of the original values However existing
approach easily mis- estimate many samples
B Data Analysis
In order to determine whether the proposed
packetization strategy could contribute to improved
accuracy of data analysis we design our experiments as
follow First we employ the sliding window techniques
with dynamic window length For each sliding window
we extract high level features (eg average absolute
difference time between peaks and etc) instead of using
the raw sensor value Second we choose two widely used
data mining and machine learning classification
algorithms Naive Bayes Classifier and Support Vector
Machines (SVM) Classifier Next we train two models
using the aforementioned classifiers We define these two
models as ldquoNBCrdquo (which is trained using Naive Bayes
Classifier) and ldquoSVMrdquo (which is trained using Support
Vector Machines) Once the two models were trained we
will apply each model to the data sets generated by the
four strategies (introduced in the beginning of Section 5)
For each testing six accuracy measurements are used
True Positive Rate (TP Rate) False Positive Rate (FPR)
Precision Recall F-Measurement and Area Under an
ROC Curve (ROC Area)
We will first introduce our feature extraction
techniques for generating semantic features from sliding
windows In our previous work [29] we mainly employ
raw time-series data In the experiments we extend our
previous work by employing six types of features which
are listed as follows The first and second type of feature
is Average and Standard Deviation respectively The
third type of feature is called Average Resultant
Acceleration [30] [31] which is defined as the average
of the square roots of the summation of the sensor
Journal of Communications Vol 10 No 11 November 2015
855copy2015 Journal of Communications
Journal of Communications Vol 10 No 11 November 2015
856copy2015 Journal of Communications
readings within each sliding window The fourth type of
feature is named as Average Absolute Difference [30]
[31] which is defined as the average absolute difference
between the value of each of the raw sensor reading in the
sliding window and the mean value over of all the sensor
reading The fifth category of features is called Time
Between Peaks [30] which is defined as time differences
(calculated in milliseconds) between peaks in the
sinusoidal waveform from the raw sensor readings within
the sliding window
1) Comparison with various models
Table II shows the detailed results using the NBC
model As a representative scenario the results for a loss
rate of 20 are studied As we can see from this table the
proposed enhancement in this paper does contribute to the
improvement of the accuracy For example if we compare
the ldquoF-Measurementrdquo (row six of Table II) values without
lost critical samples retransmitted and with lost critical
samples retransmitted (+cs) we will find that the values
have been improved more than 5
TABLE II ACCURACY COMPARISON AMONG DIFFERENT STRATEGIES
USING NBC MODEL
Strategy Name
Existing Existing+cs LASP-cs LASP
TP Rate 0791 0803 0824 0837
FP Rate 0115 0112 0090 0089
Precision 0809 0814 0831 0847
Recall 0793 0904 0815 0817
F-Measure 0785 0801 0821 0824
ROC Area 0924 0927 0946 0949
TABLE III ACCURACY COMPARISON AMONG DIFFERENT STRATEGIES
USING SVM MODEL
Strategy Name
Existing Existing+cs LASP-cs LASP
TP Rate 0633 0645 0693 0699
FP Rate 0174 0169 0147 0145
Precision 0836 0839 0849 0851
Recall 0635 0645 0692 0699
F-Measure 0631 0630 0689 0695
ROC Area 0728 0737 0783 0782
We can draw similar conclusions from Table III if we
employ the second classification model (SVM
classification model) Again we only show the results
when the loss rate is 20 and the results under other loss
rate are similar As shown in Table III the average
performance improvements are around 10 We also
performs two sample t tests Results show that the p-value
was less than 5 which means the performance
improvements are statistically significant
2) Accuracy under various loss rates
We then study whether or not the proposed LASP
strategy is robust under different loss rates To do so we
evaluated the classification accuracy under different loss
rate within LASP Specifically we evaluated the accuracy
(precision and recall) under three loss rates 10 20
and 30 using the first classification model (NBC
model) Fig 7 shows the precision and recall of our NBC
classification algorithms As shown in this figure while
the classification performance does suffering when the
loss rate increases the reduction of classification
performance is minor For example when the loss rate is
increased for 20 (from 10 to 30) the precision and
recall are reduced by less than 3 These results are
strong indication that our proposed LASP strategy are
very robust for sample loss
Fig 7 Accuracy of LASP under different loss rate with NBC
classification model
3) Comparison with various sample lengths
Finally we conducted a third experimentin order to
determine whether the proposed approach can reduce the
detection time For example given two groups of time
series T1 and T2 we apply the same classification model
to both of them If the classification accuracy for T1 is
better than the result for T2 and the difference between the
accuracy is statistically significant we can claim that
strategy for T1 is more effective (in terms of data analytics)
than the strategy for T2 In our experiments we choose
three lengths 15 seconds (L1) 30 seconds (L2) and 45
seconds (L3) For each length we generate two data
streams the first data stream is generated by sequential
packetization (ldquoExistingrdquo) and the second data stream is
generated by the proposed strategy (LASP) As shown in
Table IV each column represents the results of data
stream with different length For example the second row
(ldquoExisting-L1rdquo) indicates the data stream whose length is
15 seconds (L1) using the sequential packetization
(ldquoExistingrdquo) strategy Each row represents different
accuracy measurements which are introduced in the
beginning of section VB From Table IV we can see that
given the same length of the data stream the accuracy
measurement for data stream from LASP is substantially
higher than the accuracy measurement for data stream
from ldquoExistingrdquo strategy This verifies that if we employ
the proposed approach we could achieve faster detection
speed In another world LASP has a good potential in
applications where real-time classification is desired
TABLE IV ACCURACY COMPARISON AMONG DIFFERENT LENGTHS
GENERATED BY DIFFERENT STRATEGIES
Strategy Existing LASP
L1 L2 L3 L1 L2 L3
TP Rate 0582 0609 0611 0655 0677 0681
FP Rate 0228 0206 0201 0188 0164 0167
Precision 0789 0811 0815 0804 0828 0827
Recall 0570 06 0611 0645 0679 0688
F-Measure 0558 0581 0587 0645 0673 0677
ROC Area 0679 0699 0701 0729 0755 0759
VI SUMMARY
Existing strategies for transmission of time series body
sensor data assumes sequential sample packetization
Journal of Communications Vol 10 No 11 November 2015
857copy2015 Journal of Communications
which leads to difficulty of data recovery in the high loss
ratio WBANs In this paper we investigated the issue of
sample packetization pattern and proposed a heuristic
approach to minimize the effect of channel loss on success
of data analysis The proposed approach is very efficient
for sensor data processing incurs no extra overhead and
does not require modification of the transport protocol
In the future we will implement the packetization
scheme in real sensor network platform such as TinyOS
[32] and experiment with more biomedical sensor data to
validate its effectiveness
ACKNOWLEDGMENT
The authors wish to thank ICNC 2015 reviewers for
their valuable comments that significantly improved the
quality of the paper This work was partially supported by
the US National Science Foundation under Grant No
1229213
REFERENCES
[1] G Z Yang Body Sensor Networks New York USA Springer
Science+Business Media LLC 2006
[2] C Cordeiro and M Patel ldquoBody area networking standardization
present and future directionsrdquo presented at the Proceedings of the
ICST 2nd International Conference on Body Area Networks
Florence Italy 2007
[3] J Ko C Lu M B Srivastava J A Stankovic A Terzis and M
Welsh ldquoWireless sensor networks for healthcarerdquo Proceedings of
the IEEE vol 98 pp 1947-1960 2010
[4] O Amft H Junker P Lukowicz G Troster and C Schuster
ldquoSensing muscle activities with body-worn sensorsrdquo in Proc
International Workshop on Wearable and Implantable Body
Sensor Networks 2006
[5] B French A Smailagic D Siewiorek V Ambur and D
Tyamagundlu ldquoClassifying wheelchair propulsion patterns with a
wrist mounted accelerometerrdquo presented at the Proceedings of the
ICST 3rd International Conference on Body Area Networks
Tempe Arizona 2008
[6] L Jones N Deo and B Lockyer ldquoWireless physiological sensor
system for ambulatory userdquo in Proc International Workshop on
Wearable and Implantable Body Sensor Networks 2006 pp 4
pp-149
[7] K V Laerhoven H W Gellersen and Y G Malliaris ldquoLong
term activity monitoring with a wearable sensor noderdquo in Proc
International Workshop on Wearable and Implantable Body
Sensor Networks 2006
[8] U Maurer A Smailagic D P Siewiorek and M Deisher
ldquoActivity recognition and monitoring using multiple sensors on
different body positionsrdquo in Proc International Workshop on
Wearable and Implantable Body Sensor Networks 2006
[9] M Quwaider and S Biswas ldquoBody posture identification using
hidden Markov model with a wearable sensor networkrdquo presented
at the Proceedings of the ICST 3rd International Conference on
Body Area Networks Tempe Arizona 2008
[10] R Chavarriaga H Sagha A Calatroni S T Digumarti G Tr et
al ldquoThe opportunity challenge A benchmark database for on-
body sensor-based activity recognitionrdquo Pattern Recogn Lett vol
34 pp 2033-2042 2013
[11] M Zhang and A A Sawchuk ldquoUSC-HAD A daily activity
dataset for ubiquitous activity recognition using wearable sensorsrdquo
presented at the Proceedings of the 2012 ACM Conference on
Ubiquitous Computing Pittsburgh Pennsylvania 2012
[12] A Bulling U Blanke and B Schiele ldquoA tutorial on human
activity recognition using body-worn inertial sensorsrdquo ACM
Comput Surv vol 46 pp 1-33 2014
[13] G Zhou J Lu C Wan M Yarvis and J Stankovic ldquoBodyqos
Adaptive and radio-agnostic qos for body sensor networksrdquo in
Proc IEEE INFOCOM Phoenix AZ 2008 pp 565-573
[14] G Zhou C Y Wan M D Yarvis and J A Stankovic
ldquoAggregator-centric QoS for body sensor networksrdquo presented at
the Proceedings of the 6th International Conference on
Information Processing in Sensor Networks Cambridge
Massachusetts USA 2007
[15] B Otal L Alonso and C Verikoukis ldquoNovel QoS scheduling
and energy-saving MAC protocol for body sensor networks
optimizationrdquo presented at the Proceedings of the ICST 3rd
International Conference on Body Area Networks Tempe
Arizona 2008
[16] J J Garcia and T Falck ldquoQuality of service for IEEE 802154-
based wireless body sensor networksrdquo in Proc 3rd International
Conference on Pervasive Computing Technologies for Healthcare
2009 pp 1-6
[17] F Gengfa and E Dutkiewicz ldquoBodyMAC Energy efficient
TDMA-based MAC protocol for wireless body area networksrdquo in
Proc 9th International Symposium on Communications and
Information Technology 2009 pp 1455-1459
[18] N Read M Li Y Cao S H Liu T Wilson and B Prabhakaran
ldquoLoss resilient strategy in body sensor networksrdquo presented at the
Proc of ACMIEEE International Conference on Body Area
Networks (BodyNets) Beijing China 2011
[19] E Weldon Jr ldquoAn improved Selective-Repeat ARQ strategyrdquo
IEEE Transactions on Communications vol 30 pp 480-486
1982
[20] T Fulford-Jones D Malan M Welsh and S Moulton
ldquoCodeBlue An ad hoc sensor network infrastructure for
emergency medical carerdquo in Proc International Workshop on
Body Sensor Networks 2004
[21] S Jiang Y Cao S Iyengar P Kuryloski R Jafari Y Xue et al
ldquoCareNet An integrated wireless sensor networking environment
for remote healthcarerdquo presented at the Proceedings of the ICST
3rd International Conference on Body Area Networks Tempe
Arizona 2008
[22] M Younis K Akkaya M Eltoweissy and A Wadaa ldquoOn
handling QoS traffic in wireless sensor networksrdquo in Proc 37th
Annual Hawaii International Conference on System Sciences Big
Island HI USA 2004
[23] P Baronti P Pillai V W C Chook S Chessa A Gotta and Y
F Hu ldquoWireless sensor networks A survey on the state of the art
and the 80215 4 and ZigBee standardsrdquo Computer
Communications vol 30 pp 1655-1695 2007
[24] Bluetooth Low Energy Core Specification Version 40 [Online]
AvailablehttpwwwbluetoothcomEnglishTechnologyWorks
PagesBluetooth_low_energy_technologyaspx
[25] K K Sup M A Ameen K Daehan L Cheolhyo and L
Hyungsoo ldquoA study on proposed IEEE 80215 WBAN MAC
protocolsrdquo in Proc 9th International Symposium on
Communications and Information Technology 2009 pp 834-840
[26] M van der Schaar and D S Turaga ldquoCross-Layer packetization
and retransmission strategies for delay-sensitive wireless
multimedia transmissionrdquo IEEE Transactions on Multimedia vol
9 pp 185-197 2007
[27] S Zhang S C Liew and P P Lam ldquoHot topic Physical-layer
network codingrdquo presented at the Proceedings of the 12th annual
International Conference on Mobile Computing and Networking
Los Angeles CA USA 2006
[28] J Ellul B Lo and G Z Yang ldquoThe BSNOS platform A body
sensor networks targeted operating system and toolsetrdquo in Proc
Journal of Communications Vol 10 No 11 November 2015
858copy2015 Journal of Communications
Fifth International Conference on Sensor Technologies and
Applications 2011 pp 381-386
[29] M Li Y Cao and B Prabhakaran ldquoLoss aware sample
packetization strategy for improvement of body sensor data
analysisrdquo in Proc International Conference on Computing
Networking and Communications Anaheim California 2015
[30] L Bao and S S Intille ldquoActivity recognition from user-annotated
acceleration datardquo in Pervasive Computing Springer 2004 pp 1-
17
[31] J R Kwapisz G M Weiss and S A Moore ldquoActivity
recognition using cell phone accelerometersrdquo ACM SigKDD
Explorations Newsletter vol 12 pp 74-82 2011
[32] TinyOS [Online] Available wwwtinyosnet
Ming Li is currently an Associate Professor
and Chair in the Department of Computer
Science California State University Fresno
Prior to that he was an Assistant Professor
from August 2006 to 2012 He received his
MS and PhD degrees in Computer Science
from The University of Texas at Dallas in
2001 and 2006 respectively His research
interests include QoS strategies for IEEE
80211 wireless LANs mobile ad-hoc networks and heterogeneous
wired and wireless networks multimedia streaming over wireless
networks body area networks and robot swarm communications
Yu Cao has been an Assistant Professor at the
Department of Computer Science The
University of Massachusetts Lowell since Aug
2013 From 2010 to 2013 he was a faculty at
The University of Tennessee at Chattanooga
(UTC) From 2007 to 2010 he was a faculty at
California State University Fresno Prior to
that he was a Visiting Fellow of Biomedical
Engineering at Mayo Clinic Rochester
Minnesota He received his MS and PhD degrees in Computer
Science from Iowa State University in 2005 and 2007 respectively He
received the BEng degree from Harbin Engineering University in 1997
the MEng degree from Huazhong University of Science and
Technology in 2000 all in Computer Science His research interests
span a variety of aspects of knowledge discover from complex data
which include the area of biomedical informatics and intelligent system
B Prabhakaran is a Professor of Computer
Science in the University of Texas at Dallas
Prof B Prabhakaran works in the area of
multimedia systems He has published several
research papers in prestigious conferences and
journals in this area Dr Prabhakaran received
the prestigious NSF CAREER Award FY
2003 for his proposal on Animation Databases
Dr Prabhakaran is General Co-Chair of ACM
Multimedia 2011 He is also Technical Program Co-Chair of IEEE
WoWMoM 2012 (World of Wireless Mobile and Multimedia
Networks) He served as the TPC Co-Chair of IEEE ISM 2010
(International Symposium on Multimedia) Dr Prabhakaran is a
Member of the Executive Council of the ACM Special Interest Group
on Multimedia (SIGMM) and is the Co-Chair of IEEE Technical
Committe on Multimedia Computing (TCMC) Special Interest Group
on Video Analytics (SIGVA) Dr Prabhakaran is the Editor-in-Chief of
the ACM SIGMM (Special Interest Group on Multimedia) web
magazine He is Member of the Editorial board of Multimedia Systems
Journal (Springer) and Multimedia Tools and Applications journal
(Springer) He has served as guest-editor (special issue on Multimedia
Authoring and Presentation) for ACM Multimedia Systems journal
Page 2
out of order samples we propose LASP (Loss Aware
Sample Packetization) an efficient heuristic approach
where a fixed pattern is followed for each sample group
of a specific size The fixed pattern has the advantage of
low number contiguous sample pairs and can be easily
applied at the receiver side for reassembling of samples
Then an additional reliability enhancement with
approach similar to selective repeat [19] for critical
samples was proposed for further improvement
II RELATED WORKS
Quality of Service (QoS) scheduling for streaming
body sensor data has been well investigated Zhou and Lu
[13] proposed BodyQoS a virtual MAC for quality of
service scheduling in BSNs The approach basically
measures the effective bandwidth and adaptively
allocates remaining resources to meet the QoS
requirements of applications A desirable feature of
BodyQoS is that it does not require the modification of
the underlying MAC layer implementation Fulford-Jones
and Malan [20] proposed CodeBlue an ad hoc
infrastructure for emergent medical care In this project
several types of body sensors (eg pulse oximeter
ECGEKG sensor) are individually connected to Zigbee
enabled radio transmitters which communicate with
access points directly Due to the ad hoc architecture and
the capability of self-organizing CodeBlue yields
scalability for network expanding and flexibility to
connect various wireless devices Jiang and Cao [21]
proposed CareNet an integrated wireless environment
used for remote health care systems CareNet offers
features such as high reliability and performance
scalability security and integration with web based portal
systems High reliability is achieved by using two-tier
architecture Younis and Akkaya [22] proposed
Distributed Queuing Body Area Network a MAC
protocol aiming at providing better QoS It uses a cross-
layer fuzzy rule based scheduling algorithm to optimize
MAC layer performance in terms of QoS and energy
efficiency
There have been considerable standardization efforts
during recent years Among many potential technologies
Zigbee and Bluetooth are most widely deployed ZigBee
is a very low power collision avoidance protocol
optimized for lower power sensors It has developed a
health care specific protocol and are compliant with all
IEEE 11073 devices as well as most other IEEE 802154
[23] wireless devices Bluetooth supports high-bandwidth
and many several existing devices and has a health care
compliant version defined but has very high power
requirements and uptime for the radios Bluetooth Low
Energy [24] is a new proposed system from Bluetooth
which will have lower energy requirements and still be
interoperable with Bluetooth Classic but details are still
forthcoming at this time Being aware of the unique
requirement of supporting a wide range of applications by
body area networks an IEEE 802156 [25] task force has
Channel is unreliable and may drop data
With multiple body sensors communicating with the
same aggregator channel may reach saturation status
and therefore dropping packets in the queue
Reliable transmission of all samples incurs high
protocol overhead for relatively small body sensor
samples
A Data Recovery and Critical Samples
With quite strict requirement on the completeness of
sensor samples by data analysis algorithms missing
samples at data aggregator should be recovered at the
maximally possible level Itrsquos possible to use selective
repeat [19] to send back to specific sensors and request
retransmission of all lost samples However this
approach has the disadvantage of higher overhead and is
not efficient for real-time analysis Instead it is desirable
to recover samples using received ones and try to achieve
satisfactory approximation Usually lost samples can be
recovered by estimation (eg linear interpolation) at the
Journal of Communications Vol 10 No 11 November 2015
852copy2015 Journal of Communications
been working on finalizing a Wireless Body Area
Network (WBAN) standard The new standard aims at
providing flexible and configurable energy efficient
MAC operations
Plenty of research has been conducted using body
sensors to monitor Activity of Daily Living (ADL) Two
survey papers [3] [4] give details on the state-of-art-of
research in the area of sensor network with inertial
sensors and their applications in healthcare and wellbeing
Due to the growing interest on body sensor-based ADL
recognition several sensor-based human activity datasets
[10] [11] were introduced in the last few years A most
recent paper published in 2014 ACM Computing Survey
[12] discuss the key research challenges that human
activity recognition shares with general pattern
recognition and identify those challenges that are specific
to human activity recognition This paper also describes
the concept of an Activity Recognition Chain (ARC) as a
general-purpose framework for designing and evaluating
activity recognition systems Most of the techniques in
the current state of the art are focusing on the data
analysis and pattern recognition using the raw sensor data
Very little research in data analytics area considers the
issues of developing new networking strategy to improve
the accuracy and speed of sensor data analytics
III SYSTEM MODEL AND ASSUMPTIONS
In a typical medical application multiple sensors such
as ECG EKG EMG EEG motion sensors and blood
pressure sensors send multimodal time series data to
nearby Data Aggregators (DAs) which can be a cell
phone watch headset PDA laptop or robot based on
the application needs Then through BluetoothWiFi
these data can be delivered remotely to physician side for
real time diagnosis or to medical database for record
keeping or to request for emergency For such systems
the following assumptions are made
data aggregator In addition for most sample recovery
methods estimation accuracy of a missing sample is
significantly higher when the immediate previous and
next samples are received
For formal representation we define data recovery
ratio as the percentage of lost samples that can be
estimated using received samples with satisfactory
approximation from the data analysis requirement In
addition we define critical samples as samples that
cannot be recovered with satisfactory accuracy even if
immediate neighboring samples are received Obviously
it is ineffective to recover critical samples with received
ones for data analysis Instead selective repeat can be
adopted
B Sample Loss vs Packet Loss
Existing efforts on QoS support help address packet
loss However from data analysis aspect it is sufficient
as long as enough samples are received or recovered
Therefore understanding the relationship between packet
loss and data recovery ratio is critical For this purpose it
is important to consider the largely overlooked difference
between packet loss and sample loss Given specific
packet loss ratio it is highly desirable that sample loss
pattern yields optimal data recovery ratio
Since samples are usually assembled in packets it is
common that multiple samples are lost at the same time
In FIFO queuing based transport protocols such as UDP
samples are packetized through sequential order by
default ie samples i i+1 i+2 hellip i+L
the sample packet with L being the packet size However
if a packet is dropped then all the L continuous samples
are lost In general recovering L continuous samples
incurs significant error especially when L is not small
As illustrated in Fig 1 it will be better off to distribute
adjacent samples in different packets such that if one
packet is lost each missing sample can still be estimated
by received adjacent samples with higher accuracy It
should be noted that how samples are assembled can be
largely manipulated without affecting network operation
correctness and efficiency We name this process sample
packetization [26] The issue is therefore how to identify
an appropriate pattern for sample packetization such that
the effect of loss on data analysis is minimized
Fig 1 Comparison of two basic packetization approaches
IV LOSS RESILIENT SAMPLE PACKETIZATION (LASP)
Understanding the potential of sample packetization on
improving data recovery ratio we can formulate the
problem statement as follows given a set of N collected
samples channel loss condition and a specific estimation
technique how to packetizesamples such that the overall
chance of missing continuous samples is minimized
Obviously with minimum continuous sample loss the
data sample recovery accuracy can be significantly
improved Then the solution for the aforementioned
problem is to find a pattern of sample packetization such
that the number of continuous sample pairs between any
two packets is minimal
A Possible Solutions and Challenges
To better understand the idea letrsquos take 16 samples
with index from 0 to 15 Fig 2 shows several possible
packetization patterns with corresponding numbers of
continuous sample pairs For simplicity it is assumed that
four samples are assembled in each packet naming P0 P1
P2 and P3 N1 represents the number of continuous
samples in the same packet N2 and N3 represent the
average and maximum numbers of contiguous sample
pairs between any two packets It is clear that approach (a)
is worst since it incurs N1 of 3 which indicates difficulty
of recovering samples even if one packet is lost Approach
(c) works slightly better than (b) since its N2 and N3 are
smaller indicating better performance when multiple
packet loss occurs
Basically given N samples finding the optimal
packetization pattern is computationally expense
especially when N is not small For low energy consuming
body sensors it is not acceptable to enumerate all
possibility for comparison Approach (c) in Fig 2 is
slightly better than (b) but finding such pattern at real-
time incurs overhead On the other hand one may think
that random pattern works well and may be able to
generate small number of continuous sample pairs
However even if it works well the reassembly of samples
according to their original order is quite difficult In order
to restore the order significant effort must be made to
deliver the pattern therefore incurring high network traffic
and leading to a waste of network bandwidth Therefore
we decide to take a heuristic solution based on approach
(b) described in the next subsection
Fig 2 Comparison of packetization approaches of 16 samples
B Proposed Approach
We proposed to design a heuristic packetization pattern
based on the approach (b) in Fig 2 Fig 3 illustrates the
For a total of 8 samples two packetization approaches are
considered
Sequential Packet 1 (0 1 2 3) Packet 2 (4 5 6 7) Alternate Packet 1 (0 2 4 6) Packet 2 (1 3 5 7)
Assuming the first packet gets lost the ldquoalternaterdquo approach yields better data analysis performance since it is easier
to recover recover samples 2 4 6 with good accuracy However
ldquosequentialrdquo approach will have difficulty since the lost samples
are too far away from received ones
Journal of Communications Vol 10 No 11 November 2015
853copy2015 Journal of Communications
-1 are put into
procedure of the proposed approach which is described
as follows
Step 1 for each collected sample group packetization
is performed using alternate pattern Assuming N samples
and M packets for each sample i (0ltiltN) the ID of
corresponding packet that it should be inserted into is
iM
Step 2 for the same group all critical samples are
identified A sample i is considered critical if |estimatedi-
originali| gt α originali Here the estimated value depends
on the specific estimation method and the values of the
neighboring samples α is a constant and is fixed as 025
in this paper
Step 3 all packets are sent to the aggregator At the
same time a special packet containing the IDs of all
critical samples is also sent
Step 4 upon the receipt of packets data aggregator re-
assembles received samples based on the same pattern
Therefore the kth sample in jth packet has the sample ID
of j+kN
Step 5 aggregator identifies lost critical samples by
comparing the IDs of the lost samples and the IDs of all
critical samples
Step 6 aggregator requests the sensor to retransmit
critical samples and ensure its successful delivery
Step 7 aggregator performs data recovery for other
lost samples before performing analysis
Fig 3 Comparison of packetization approaches of 16 samples
It should be noted that the proposed strategy is an
application layer approach and therefore is different from
the physical layer network coding [27] which aims at
coordination of multiple node transmission in order to
achieve higher network throughput especially in wireless
networks However the system performance will
potentially benefit from physical layer network coding if
implemented
C Data Quality Model
In this section we present a variance based data
quality model for the analysis of data recovery ratio We
define estimation variance (EV) as the relative difference
between the estimated sample values and original values
which can calculate it as
(1)
where di is the original value of the sample i and esti is the
estimated value of sample i from received samples
Further the total estimation variance is calculated as
(2)
(3)
where K is the total number of estimated samples in any
evaluation group Clearly smaller EV and AEV indicates
better data recovery which is critical for the success of
data analysis
Now letrsquos assume that a group of M independent
packets are transmitted For each packet j its total
estimation variance of all samples in the packet EV(j)
depends on the specific characteristics of the data samples
in the packet Thus for a given packet loss ratio of l the
expected total estimation variance is calculated as
If EV(j) is the same value (EV) for all packets which is
not usual in real scenarios due to the statistic dropping of
packets we then have that
Clearly the total EV depends on total number of
packets loss ratio and data characteristics Further
average expected estimation variance is now
(6)
That means if packets are independently recovered
then loss ratio does not affect the level of data recovery
However when two or more adjacent packets are lost it is
difficult to have independent data recovery therefore
higher loss ration does decrease the data recovery ratio
TABLE I EV VS DATA CHARACTERISTICS
Strategy Data Characteristics
sqrt linear square cubic quartic
Sequential 0075 0 039 090 140
Alternate 0003 0 002 006 012
To illustrate the effect of the packetization on data with
different characteristics we take a data segment that
follows different patterns and calculate the total EV for
the segment with two different packetization schemes
Each data segment contains 8 samples which are
packetized into two different packets with either the
ldquosequentialrdquo or ldquoalternaterdquo approach under the condition
that the second packet is lost Table I shows the result It
can be seen that as the curve deviates more from the linear
pattern higher variance is yielded for both ldquosequentialrdquo
and ldquoalternaterdquo approach However ldquoalternaterdquo approach
shows a much slower increase compared with ldquosequentialrdquo
V PERFORMANCE EVALUATION
We performed extensive simulations to study the
performance of the proposed packetization strategy For
simplicity we label the proposed strategy as LASP and
sequential packetization as ldquoExistingrdquo The following four
strategies were compared
Existing samples are packetized in order of creation
ie sequential
Collect
Samples
Perform
PacketizationSample
Assembling
Identify Critical
Samples
Request Lost
Critical Samples
Resend Critical
Samples
Data
Recovery
Sensor Side Aggregator
iiii ddestEV
i iEVEV
KEVAEV i i
M
m
m
j
mMm jEVllm
MEV
0 1
)()1(
EVlMmllm
MEVEV
M
m
mMm
0
)1(
EVlMEVlMAEV
Journal of Communications Vol 10 No 11 November 2015
854copy2015 Journal of Communications
(4)
(5)
Finally the Average Estimation Variance (AEV) is
calculated as
Existing+cs existing approach with lost critical
samples retransmitted
LASP proposed approach described in section 42
LASP-cs proposed approach but lost critical samples
are not retransmitted
We use the Body Sensor Network (BSN) Development
Kit v3 [28] developed by Prof Guang-Zhong Yang of
Imperial College to collect the accelerometer data from
users The BSN development kit v3 includes a USB
programming board a sensor board a battery board a
prototype board and a pair of the new BSN v3 nodes The
sensor board has a 3D accelerometer and a temperature
sensor In our experiments we recruit three subjects and
attach the sensor board to the arm of the subject Each
subject is asked to perform three types of Activity of Daily
Living (ADL) jogging sitting and walking The duration
of each activity is around 3 minutes The 3D
accelerometer data was collected when the subjects
performed the ADL activity We chose to use the
sampling rate as 10 Hz We also implemented linear
interpolation to calculate a sample directly between any
two neighboring ones the interpolation effectively yields
a 33 Hz sampling rate with the smoothness of the 16 Hz
rate Using the 10 Hz sampling rate with linear
interpolation yielded a rather smooth graph without a lot
of noise and yet it was fast enough that the peaks and
troughs of ADL were all present
Our evaluation consists of two parts In part 1 received
samples with different approaches are compared in term
of several key metrics In part 2 recovered data are
analyzed by various representative analysis algorithms for
comparison of the accuracy
Fig 4 Comparison of Average Estimation Variance (AEV)
A Sample Quality Analysis
For each received samples we evaluate how
satisfactory the data is recovered by comparing the
estimated values with original ones Fig 4 shows the
comparison of average estimation variance of lost samples
for x values of jogging data It can be seen that LASP has
only about 6-9 normalized variance on average while
existing approach yields an average of 15-17 Also
retransmission of critical samples does not help much for
existing approach mainly due to the high estimation
inaccuracy when a contiguous sample train gets lost Fig 5
compares number of lost samples with normalized
variance higher than 25 Itrsquos clear that LASP is much
better than existing approach even if critical samples are
not retransmitted This is due to the fact that LASP creates
a much nicer sample loss pattern as described earlier and
therefore minimizes the variance for lost samples
Fig 5 Number of samples with low recovery ratio
Fig 6 Comparison of data recovery
To provide a visual representation of how well LASP
improves data recovery 100 samples are extracted from a
sequence of 1043 samples for jogging Itrsquos clear that
LASP while having some variance creates a good
approximation of the original values However existing
approach easily mis- estimate many samples
B Data Analysis
In order to determine whether the proposed
packetization strategy could contribute to improved
accuracy of data analysis we design our experiments as
follow First we employ the sliding window techniques
with dynamic window length For each sliding window
we extract high level features (eg average absolute
difference time between peaks and etc) instead of using
the raw sensor value Second we choose two widely used
data mining and machine learning classification
algorithms Naive Bayes Classifier and Support Vector
Machines (SVM) Classifier Next we train two models
using the aforementioned classifiers We define these two
models as ldquoNBCrdquo (which is trained using Naive Bayes
Classifier) and ldquoSVMrdquo (which is trained using Support
Vector Machines) Once the two models were trained we
will apply each model to the data sets generated by the
four strategies (introduced in the beginning of Section 5)
For each testing six accuracy measurements are used
True Positive Rate (TP Rate) False Positive Rate (FPR)
Precision Recall F-Measurement and Area Under an
ROC Curve (ROC Area)
We will first introduce our feature extraction
techniques for generating semantic features from sliding
windows In our previous work [29] we mainly employ
raw time-series data In the experiments we extend our
previous work by employing six types of features which
are listed as follows The first and second type of feature
is Average and Standard Deviation respectively The
third type of feature is called Average Resultant
Acceleration [30] [31] which is defined as the average
of the square roots of the summation of the sensor
Journal of Communications Vol 10 No 11 November 2015
855copy2015 Journal of Communications
Journal of Communications Vol 10 No 11 November 2015
856copy2015 Journal of Communications
readings within each sliding window The fourth type of
feature is named as Average Absolute Difference [30]
[31] which is defined as the average absolute difference
between the value of each of the raw sensor reading in the
sliding window and the mean value over of all the sensor
reading The fifth category of features is called Time
Between Peaks [30] which is defined as time differences
(calculated in milliseconds) between peaks in the
sinusoidal waveform from the raw sensor readings within
the sliding window
1) Comparison with various models
Table II shows the detailed results using the NBC
model As a representative scenario the results for a loss
rate of 20 are studied As we can see from this table the
proposed enhancement in this paper does contribute to the
improvement of the accuracy For example if we compare
the ldquoF-Measurementrdquo (row six of Table II) values without
lost critical samples retransmitted and with lost critical
samples retransmitted (+cs) we will find that the values
have been improved more than 5
TABLE II ACCURACY COMPARISON AMONG DIFFERENT STRATEGIES
USING NBC MODEL
Strategy Name
Existing Existing+cs LASP-cs LASP
TP Rate 0791 0803 0824 0837
FP Rate 0115 0112 0090 0089
Precision 0809 0814 0831 0847
Recall 0793 0904 0815 0817
F-Measure 0785 0801 0821 0824
ROC Area 0924 0927 0946 0949
TABLE III ACCURACY COMPARISON AMONG DIFFERENT STRATEGIES
USING SVM MODEL
Strategy Name
Existing Existing+cs LASP-cs LASP
TP Rate 0633 0645 0693 0699
FP Rate 0174 0169 0147 0145
Precision 0836 0839 0849 0851
Recall 0635 0645 0692 0699
F-Measure 0631 0630 0689 0695
ROC Area 0728 0737 0783 0782
We can draw similar conclusions from Table III if we
employ the second classification model (SVM
classification model) Again we only show the results
when the loss rate is 20 and the results under other loss
rate are similar As shown in Table III the average
performance improvements are around 10 We also
performs two sample t tests Results show that the p-value
was less than 5 which means the performance
improvements are statistically significant
2) Accuracy under various loss rates
We then study whether or not the proposed LASP
strategy is robust under different loss rates To do so we
evaluated the classification accuracy under different loss
rate within LASP Specifically we evaluated the accuracy
(precision and recall) under three loss rates 10 20
and 30 using the first classification model (NBC
model) Fig 7 shows the precision and recall of our NBC
classification algorithms As shown in this figure while
the classification performance does suffering when the
loss rate increases the reduction of classification
performance is minor For example when the loss rate is
increased for 20 (from 10 to 30) the precision and
recall are reduced by less than 3 These results are
strong indication that our proposed LASP strategy are
very robust for sample loss
Fig 7 Accuracy of LASP under different loss rate with NBC
classification model
3) Comparison with various sample lengths
Finally we conducted a third experimentin order to
determine whether the proposed approach can reduce the
detection time For example given two groups of time
series T1 and T2 we apply the same classification model
to both of them If the classification accuracy for T1 is
better than the result for T2 and the difference between the
accuracy is statistically significant we can claim that
strategy for T1 is more effective (in terms of data analytics)
than the strategy for T2 In our experiments we choose
three lengths 15 seconds (L1) 30 seconds (L2) and 45
seconds (L3) For each length we generate two data
streams the first data stream is generated by sequential
packetization (ldquoExistingrdquo) and the second data stream is
generated by the proposed strategy (LASP) As shown in
Table IV each column represents the results of data
stream with different length For example the second row
(ldquoExisting-L1rdquo) indicates the data stream whose length is
15 seconds (L1) using the sequential packetization
(ldquoExistingrdquo) strategy Each row represents different
accuracy measurements which are introduced in the
beginning of section VB From Table IV we can see that
given the same length of the data stream the accuracy
measurement for data stream from LASP is substantially
higher than the accuracy measurement for data stream
from ldquoExistingrdquo strategy This verifies that if we employ
the proposed approach we could achieve faster detection
speed In another world LASP has a good potential in
applications where real-time classification is desired
TABLE IV ACCURACY COMPARISON AMONG DIFFERENT LENGTHS
GENERATED BY DIFFERENT STRATEGIES
Strategy Existing LASP
L1 L2 L3 L1 L2 L3
TP Rate 0582 0609 0611 0655 0677 0681
FP Rate 0228 0206 0201 0188 0164 0167
Precision 0789 0811 0815 0804 0828 0827
Recall 0570 06 0611 0645 0679 0688
F-Measure 0558 0581 0587 0645 0673 0677
ROC Area 0679 0699 0701 0729 0755 0759
VI SUMMARY
Existing strategies for transmission of time series body
sensor data assumes sequential sample packetization
Journal of Communications Vol 10 No 11 November 2015
857copy2015 Journal of Communications
which leads to difficulty of data recovery in the high loss
ratio WBANs In this paper we investigated the issue of
sample packetization pattern and proposed a heuristic
approach to minimize the effect of channel loss on success
of data analysis The proposed approach is very efficient
for sensor data processing incurs no extra overhead and
does not require modification of the transport protocol
In the future we will implement the packetization
scheme in real sensor network platform such as TinyOS
[32] and experiment with more biomedical sensor data to
validate its effectiveness
ACKNOWLEDGMENT
The authors wish to thank ICNC 2015 reviewers for
their valuable comments that significantly improved the
quality of the paper This work was partially supported by
the US National Science Foundation under Grant No
1229213
REFERENCES
[1] G Z Yang Body Sensor Networks New York USA Springer
Science+Business Media LLC 2006
[2] C Cordeiro and M Patel ldquoBody area networking standardization
present and future directionsrdquo presented at the Proceedings of the
ICST 2nd International Conference on Body Area Networks
Florence Italy 2007
[3] J Ko C Lu M B Srivastava J A Stankovic A Terzis and M
Welsh ldquoWireless sensor networks for healthcarerdquo Proceedings of
the IEEE vol 98 pp 1947-1960 2010
[4] O Amft H Junker P Lukowicz G Troster and C Schuster
ldquoSensing muscle activities with body-worn sensorsrdquo in Proc
International Workshop on Wearable and Implantable Body
Sensor Networks 2006
[5] B French A Smailagic D Siewiorek V Ambur and D
Tyamagundlu ldquoClassifying wheelchair propulsion patterns with a
wrist mounted accelerometerrdquo presented at the Proceedings of the
ICST 3rd International Conference on Body Area Networks
Tempe Arizona 2008
[6] L Jones N Deo and B Lockyer ldquoWireless physiological sensor
system for ambulatory userdquo in Proc International Workshop on
Wearable and Implantable Body Sensor Networks 2006 pp 4
pp-149
[7] K V Laerhoven H W Gellersen and Y G Malliaris ldquoLong
term activity monitoring with a wearable sensor noderdquo in Proc
International Workshop on Wearable and Implantable Body
Sensor Networks 2006
[8] U Maurer A Smailagic D P Siewiorek and M Deisher
ldquoActivity recognition and monitoring using multiple sensors on
different body positionsrdquo in Proc International Workshop on
Wearable and Implantable Body Sensor Networks 2006
[9] M Quwaider and S Biswas ldquoBody posture identification using
hidden Markov model with a wearable sensor networkrdquo presented
at the Proceedings of the ICST 3rd International Conference on
Body Area Networks Tempe Arizona 2008
[10] R Chavarriaga H Sagha A Calatroni S T Digumarti G Tr et
al ldquoThe opportunity challenge A benchmark database for on-
body sensor-based activity recognitionrdquo Pattern Recogn Lett vol
34 pp 2033-2042 2013
[11] M Zhang and A A Sawchuk ldquoUSC-HAD A daily activity
dataset for ubiquitous activity recognition using wearable sensorsrdquo
presented at the Proceedings of the 2012 ACM Conference on
Ubiquitous Computing Pittsburgh Pennsylvania 2012
[12] A Bulling U Blanke and B Schiele ldquoA tutorial on human
activity recognition using body-worn inertial sensorsrdquo ACM
Comput Surv vol 46 pp 1-33 2014
[13] G Zhou J Lu C Wan M Yarvis and J Stankovic ldquoBodyqos
Adaptive and radio-agnostic qos for body sensor networksrdquo in
Proc IEEE INFOCOM Phoenix AZ 2008 pp 565-573
[14] G Zhou C Y Wan M D Yarvis and J A Stankovic
ldquoAggregator-centric QoS for body sensor networksrdquo presented at
the Proceedings of the 6th International Conference on
Information Processing in Sensor Networks Cambridge
Massachusetts USA 2007
[15] B Otal L Alonso and C Verikoukis ldquoNovel QoS scheduling
and energy-saving MAC protocol for body sensor networks
optimizationrdquo presented at the Proceedings of the ICST 3rd
International Conference on Body Area Networks Tempe
Arizona 2008
[16] J J Garcia and T Falck ldquoQuality of service for IEEE 802154-
based wireless body sensor networksrdquo in Proc 3rd International
Conference on Pervasive Computing Technologies for Healthcare
2009 pp 1-6
[17] F Gengfa and E Dutkiewicz ldquoBodyMAC Energy efficient
TDMA-based MAC protocol for wireless body area networksrdquo in
Proc 9th International Symposium on Communications and
Information Technology 2009 pp 1455-1459
[18] N Read M Li Y Cao S H Liu T Wilson and B Prabhakaran
ldquoLoss resilient strategy in body sensor networksrdquo presented at the
Proc of ACMIEEE International Conference on Body Area
Networks (BodyNets) Beijing China 2011
[19] E Weldon Jr ldquoAn improved Selective-Repeat ARQ strategyrdquo
IEEE Transactions on Communications vol 30 pp 480-486
1982
[20] T Fulford-Jones D Malan M Welsh and S Moulton
ldquoCodeBlue An ad hoc sensor network infrastructure for
emergency medical carerdquo in Proc International Workshop on
Body Sensor Networks 2004
[21] S Jiang Y Cao S Iyengar P Kuryloski R Jafari Y Xue et al
ldquoCareNet An integrated wireless sensor networking environment
for remote healthcarerdquo presented at the Proceedings of the ICST
3rd International Conference on Body Area Networks Tempe
Arizona 2008
[22] M Younis K Akkaya M Eltoweissy and A Wadaa ldquoOn
handling QoS traffic in wireless sensor networksrdquo in Proc 37th
Annual Hawaii International Conference on System Sciences Big
Island HI USA 2004
[23] P Baronti P Pillai V W C Chook S Chessa A Gotta and Y
F Hu ldquoWireless sensor networks A survey on the state of the art
and the 80215 4 and ZigBee standardsrdquo Computer
Communications vol 30 pp 1655-1695 2007
[24] Bluetooth Low Energy Core Specification Version 40 [Online]
AvailablehttpwwwbluetoothcomEnglishTechnologyWorks
PagesBluetooth_low_energy_technologyaspx
[25] K K Sup M A Ameen K Daehan L Cheolhyo and L
Hyungsoo ldquoA study on proposed IEEE 80215 WBAN MAC
protocolsrdquo in Proc 9th International Symposium on
Communications and Information Technology 2009 pp 834-840
[26] M van der Schaar and D S Turaga ldquoCross-Layer packetization
and retransmission strategies for delay-sensitive wireless
multimedia transmissionrdquo IEEE Transactions on Multimedia vol
9 pp 185-197 2007
[27] S Zhang S C Liew and P P Lam ldquoHot topic Physical-layer
network codingrdquo presented at the Proceedings of the 12th annual
International Conference on Mobile Computing and Networking
Los Angeles CA USA 2006
[28] J Ellul B Lo and G Z Yang ldquoThe BSNOS platform A body
sensor networks targeted operating system and toolsetrdquo in Proc
Journal of Communications Vol 10 No 11 November 2015
858copy2015 Journal of Communications
Fifth International Conference on Sensor Technologies and
Applications 2011 pp 381-386
[29] M Li Y Cao and B Prabhakaran ldquoLoss aware sample
packetization strategy for improvement of body sensor data
analysisrdquo in Proc International Conference on Computing
Networking and Communications Anaheim California 2015
[30] L Bao and S S Intille ldquoActivity recognition from user-annotated
acceleration datardquo in Pervasive Computing Springer 2004 pp 1-
17
[31] J R Kwapisz G M Weiss and S A Moore ldquoActivity
recognition using cell phone accelerometersrdquo ACM SigKDD
Explorations Newsletter vol 12 pp 74-82 2011
[32] TinyOS [Online] Available wwwtinyosnet
Ming Li is currently an Associate Professor
and Chair in the Department of Computer
Science California State University Fresno
Prior to that he was an Assistant Professor
from August 2006 to 2012 He received his
MS and PhD degrees in Computer Science
from The University of Texas at Dallas in
2001 and 2006 respectively His research
interests include QoS strategies for IEEE
80211 wireless LANs mobile ad-hoc networks and heterogeneous
wired and wireless networks multimedia streaming over wireless
networks body area networks and robot swarm communications
Yu Cao has been an Assistant Professor at the
Department of Computer Science The
University of Massachusetts Lowell since Aug
2013 From 2010 to 2013 he was a faculty at
The University of Tennessee at Chattanooga
(UTC) From 2007 to 2010 he was a faculty at
California State University Fresno Prior to
that he was a Visiting Fellow of Biomedical
Engineering at Mayo Clinic Rochester
Minnesota He received his MS and PhD degrees in Computer
Science from Iowa State University in 2005 and 2007 respectively He
received the BEng degree from Harbin Engineering University in 1997
the MEng degree from Huazhong University of Science and
Technology in 2000 all in Computer Science His research interests
span a variety of aspects of knowledge discover from complex data
which include the area of biomedical informatics and intelligent system
B Prabhakaran is a Professor of Computer
Science in the University of Texas at Dallas
Prof B Prabhakaran works in the area of
multimedia systems He has published several
research papers in prestigious conferences and
journals in this area Dr Prabhakaran received
the prestigious NSF CAREER Award FY
2003 for his proposal on Animation Databases
Dr Prabhakaran is General Co-Chair of ACM
Multimedia 2011 He is also Technical Program Co-Chair of IEEE
WoWMoM 2012 (World of Wireless Mobile and Multimedia
Networks) He served as the TPC Co-Chair of IEEE ISM 2010
(International Symposium on Multimedia) Dr Prabhakaran is a
Member of the Executive Council of the ACM Special Interest Group
on Multimedia (SIGMM) and is the Co-Chair of IEEE Technical
Committe on Multimedia Computing (TCMC) Special Interest Group
on Video Analytics (SIGVA) Dr Prabhakaran is the Editor-in-Chief of
the ACM SIGMM (Special Interest Group on Multimedia) web
magazine He is Member of the Editorial board of Multimedia Systems
Journal (Springer) and Multimedia Tools and Applications journal
(Springer) He has served as guest-editor (special issue on Multimedia
Authoring and Presentation) for ACM Multimedia Systems journal
Page 3
data aggregator In addition for most sample recovery
methods estimation accuracy of a missing sample is
significantly higher when the immediate previous and
next samples are received
For formal representation we define data recovery
ratio as the percentage of lost samples that can be
estimated using received samples with satisfactory
approximation from the data analysis requirement In
addition we define critical samples as samples that
cannot be recovered with satisfactory accuracy even if
immediate neighboring samples are received Obviously
it is ineffective to recover critical samples with received
ones for data analysis Instead selective repeat can be
adopted
B Sample Loss vs Packet Loss
Existing efforts on QoS support help address packet
loss However from data analysis aspect it is sufficient
as long as enough samples are received or recovered
Therefore understanding the relationship between packet
loss and data recovery ratio is critical For this purpose it
is important to consider the largely overlooked difference
between packet loss and sample loss Given specific
packet loss ratio it is highly desirable that sample loss
pattern yields optimal data recovery ratio
Since samples are usually assembled in packets it is
common that multiple samples are lost at the same time
In FIFO queuing based transport protocols such as UDP
samples are packetized through sequential order by
default ie samples i i+1 i+2 hellip i+L
the sample packet with L being the packet size However
if a packet is dropped then all the L continuous samples
are lost In general recovering L continuous samples
incurs significant error especially when L is not small
As illustrated in Fig 1 it will be better off to distribute
adjacent samples in different packets such that if one
packet is lost each missing sample can still be estimated
by received adjacent samples with higher accuracy It
should be noted that how samples are assembled can be
largely manipulated without affecting network operation
correctness and efficiency We name this process sample
packetization [26] The issue is therefore how to identify
an appropriate pattern for sample packetization such that
the effect of loss on data analysis is minimized
Fig 1 Comparison of two basic packetization approaches
IV LOSS RESILIENT SAMPLE PACKETIZATION (LASP)
Understanding the potential of sample packetization on
improving data recovery ratio we can formulate the
problem statement as follows given a set of N collected
samples channel loss condition and a specific estimation
technique how to packetizesamples such that the overall
chance of missing continuous samples is minimized
Obviously with minimum continuous sample loss the
data sample recovery accuracy can be significantly
improved Then the solution for the aforementioned
problem is to find a pattern of sample packetization such
that the number of continuous sample pairs between any
two packets is minimal
A Possible Solutions and Challenges
To better understand the idea letrsquos take 16 samples
with index from 0 to 15 Fig 2 shows several possible
packetization patterns with corresponding numbers of
continuous sample pairs For simplicity it is assumed that
four samples are assembled in each packet naming P0 P1
P2 and P3 N1 represents the number of continuous
samples in the same packet N2 and N3 represent the
average and maximum numbers of contiguous sample
pairs between any two packets It is clear that approach (a)
is worst since it incurs N1 of 3 which indicates difficulty
of recovering samples even if one packet is lost Approach
(c) works slightly better than (b) since its N2 and N3 are
smaller indicating better performance when multiple
packet loss occurs
Basically given N samples finding the optimal
packetization pattern is computationally expense
especially when N is not small For low energy consuming
body sensors it is not acceptable to enumerate all
possibility for comparison Approach (c) in Fig 2 is
slightly better than (b) but finding such pattern at real-
time incurs overhead On the other hand one may think
that random pattern works well and may be able to
generate small number of continuous sample pairs
However even if it works well the reassembly of samples
according to their original order is quite difficult In order
to restore the order significant effort must be made to
deliver the pattern therefore incurring high network traffic
and leading to a waste of network bandwidth Therefore
we decide to take a heuristic solution based on approach
(b) described in the next subsection
Fig 2 Comparison of packetization approaches of 16 samples
B Proposed Approach
We proposed to design a heuristic packetization pattern
based on the approach (b) in Fig 2 Fig 3 illustrates the
For a total of 8 samples two packetization approaches are
considered
Sequential Packet 1 (0 1 2 3) Packet 2 (4 5 6 7) Alternate Packet 1 (0 2 4 6) Packet 2 (1 3 5 7)
Assuming the first packet gets lost the ldquoalternaterdquo approach yields better data analysis performance since it is easier
to recover recover samples 2 4 6 with good accuracy However
ldquosequentialrdquo approach will have difficulty since the lost samples
are too far away from received ones
Journal of Communications Vol 10 No 11 November 2015
853copy2015 Journal of Communications
-1 are put into
procedure of the proposed approach which is described
as follows
Step 1 for each collected sample group packetization
is performed using alternate pattern Assuming N samples
and M packets for each sample i (0ltiltN) the ID of
corresponding packet that it should be inserted into is
iM
Step 2 for the same group all critical samples are
identified A sample i is considered critical if |estimatedi-
originali| gt α originali Here the estimated value depends
on the specific estimation method and the values of the
neighboring samples α is a constant and is fixed as 025
in this paper
Step 3 all packets are sent to the aggregator At the
same time a special packet containing the IDs of all
critical samples is also sent
Step 4 upon the receipt of packets data aggregator re-
assembles received samples based on the same pattern
Therefore the kth sample in jth packet has the sample ID
of j+kN
Step 5 aggregator identifies lost critical samples by
comparing the IDs of the lost samples and the IDs of all
critical samples
Step 6 aggregator requests the sensor to retransmit
critical samples and ensure its successful delivery
Step 7 aggregator performs data recovery for other
lost samples before performing analysis
Fig 3 Comparison of packetization approaches of 16 samples
It should be noted that the proposed strategy is an
application layer approach and therefore is different from
the physical layer network coding [27] which aims at
coordination of multiple node transmission in order to
achieve higher network throughput especially in wireless
networks However the system performance will
potentially benefit from physical layer network coding if
implemented
C Data Quality Model
In this section we present a variance based data
quality model for the analysis of data recovery ratio We
define estimation variance (EV) as the relative difference
between the estimated sample values and original values
which can calculate it as
(1)
where di is the original value of the sample i and esti is the
estimated value of sample i from received samples
Further the total estimation variance is calculated as
(2)
(3)
where K is the total number of estimated samples in any
evaluation group Clearly smaller EV and AEV indicates
better data recovery which is critical for the success of
data analysis
Now letrsquos assume that a group of M independent
packets are transmitted For each packet j its total
estimation variance of all samples in the packet EV(j)
depends on the specific characteristics of the data samples
in the packet Thus for a given packet loss ratio of l the
expected total estimation variance is calculated as
If EV(j) is the same value (EV) for all packets which is
not usual in real scenarios due to the statistic dropping of
packets we then have that
Clearly the total EV depends on total number of
packets loss ratio and data characteristics Further
average expected estimation variance is now
(6)
That means if packets are independently recovered
then loss ratio does not affect the level of data recovery
However when two or more adjacent packets are lost it is
difficult to have independent data recovery therefore
higher loss ration does decrease the data recovery ratio
TABLE I EV VS DATA CHARACTERISTICS
Strategy Data Characteristics
sqrt linear square cubic quartic
Sequential 0075 0 039 090 140
Alternate 0003 0 002 006 012
To illustrate the effect of the packetization on data with
different characteristics we take a data segment that
follows different patterns and calculate the total EV for
the segment with two different packetization schemes
Each data segment contains 8 samples which are
packetized into two different packets with either the
ldquosequentialrdquo or ldquoalternaterdquo approach under the condition
that the second packet is lost Table I shows the result It
can be seen that as the curve deviates more from the linear
pattern higher variance is yielded for both ldquosequentialrdquo
and ldquoalternaterdquo approach However ldquoalternaterdquo approach
shows a much slower increase compared with ldquosequentialrdquo
V PERFORMANCE EVALUATION
We performed extensive simulations to study the
performance of the proposed packetization strategy For
simplicity we label the proposed strategy as LASP and
sequential packetization as ldquoExistingrdquo The following four
strategies were compared
Existing samples are packetized in order of creation
ie sequential
Collect
Samples
Perform
PacketizationSample
Assembling
Identify Critical
Samples
Request Lost
Critical Samples
Resend Critical
Samples
Data
Recovery
Sensor Side Aggregator
iiii ddestEV
i iEVEV
KEVAEV i i
M
m
m
j
mMm jEVllm
MEV
0 1
)()1(
EVlMmllm
MEVEV
M
m
mMm
0
)1(
EVlMEVlMAEV
Journal of Communications Vol 10 No 11 November 2015
854copy2015 Journal of Communications
(4)
(5)
Finally the Average Estimation Variance (AEV) is
calculated as
Existing+cs existing approach with lost critical
samples retransmitted
LASP proposed approach described in section 42
LASP-cs proposed approach but lost critical samples
are not retransmitted
We use the Body Sensor Network (BSN) Development
Kit v3 [28] developed by Prof Guang-Zhong Yang of
Imperial College to collect the accelerometer data from
users The BSN development kit v3 includes a USB
programming board a sensor board a battery board a
prototype board and a pair of the new BSN v3 nodes The
sensor board has a 3D accelerometer and a temperature
sensor In our experiments we recruit three subjects and
attach the sensor board to the arm of the subject Each
subject is asked to perform three types of Activity of Daily
Living (ADL) jogging sitting and walking The duration
of each activity is around 3 minutes The 3D
accelerometer data was collected when the subjects
performed the ADL activity We chose to use the
sampling rate as 10 Hz We also implemented linear
interpolation to calculate a sample directly between any
two neighboring ones the interpolation effectively yields
a 33 Hz sampling rate with the smoothness of the 16 Hz
rate Using the 10 Hz sampling rate with linear
interpolation yielded a rather smooth graph without a lot
of noise and yet it was fast enough that the peaks and
troughs of ADL were all present
Our evaluation consists of two parts In part 1 received
samples with different approaches are compared in term
of several key metrics In part 2 recovered data are
analyzed by various representative analysis algorithms for
comparison of the accuracy
Fig 4 Comparison of Average Estimation Variance (AEV)
A Sample Quality Analysis
For each received samples we evaluate how
satisfactory the data is recovered by comparing the
estimated values with original ones Fig 4 shows the
comparison of average estimation variance of lost samples
for x values of jogging data It can be seen that LASP has
only about 6-9 normalized variance on average while
existing approach yields an average of 15-17 Also
retransmission of critical samples does not help much for
existing approach mainly due to the high estimation
inaccuracy when a contiguous sample train gets lost Fig 5
compares number of lost samples with normalized
variance higher than 25 Itrsquos clear that LASP is much
better than existing approach even if critical samples are
not retransmitted This is due to the fact that LASP creates
a much nicer sample loss pattern as described earlier and
therefore minimizes the variance for lost samples
Fig 5 Number of samples with low recovery ratio
Fig 6 Comparison of data recovery
To provide a visual representation of how well LASP
improves data recovery 100 samples are extracted from a
sequence of 1043 samples for jogging Itrsquos clear that
LASP while having some variance creates a good
approximation of the original values However existing
approach easily mis- estimate many samples
B Data Analysis
In order to determine whether the proposed
packetization strategy could contribute to improved
accuracy of data analysis we design our experiments as
follow First we employ the sliding window techniques
with dynamic window length For each sliding window
we extract high level features (eg average absolute
difference time between peaks and etc) instead of using
the raw sensor value Second we choose two widely used
data mining and machine learning classification
algorithms Naive Bayes Classifier and Support Vector
Machines (SVM) Classifier Next we train two models
using the aforementioned classifiers We define these two
models as ldquoNBCrdquo (which is trained using Naive Bayes
Classifier) and ldquoSVMrdquo (which is trained using Support
Vector Machines) Once the two models were trained we
will apply each model to the data sets generated by the
four strategies (introduced in the beginning of Section 5)
For each testing six accuracy measurements are used
True Positive Rate (TP Rate) False Positive Rate (FPR)
Precision Recall F-Measurement and Area Under an
ROC Curve (ROC Area)
We will first introduce our feature extraction
techniques for generating semantic features from sliding
windows In our previous work [29] we mainly employ
raw time-series data In the experiments we extend our
previous work by employing six types of features which
are listed as follows The first and second type of feature
is Average and Standard Deviation respectively The
third type of feature is called Average Resultant
Acceleration [30] [31] which is defined as the average
of the square roots of the summation of the sensor
Journal of Communications Vol 10 No 11 November 2015
855copy2015 Journal of Communications
Journal of Communications Vol 10 No 11 November 2015
856copy2015 Journal of Communications
readings within each sliding window The fourth type of
feature is named as Average Absolute Difference [30]
[31] which is defined as the average absolute difference
between the value of each of the raw sensor reading in the
sliding window and the mean value over of all the sensor
reading The fifth category of features is called Time
Between Peaks [30] which is defined as time differences
(calculated in milliseconds) between peaks in the
sinusoidal waveform from the raw sensor readings within
the sliding window
1) Comparison with various models
Table II shows the detailed results using the NBC
model As a representative scenario the results for a loss
rate of 20 are studied As we can see from this table the
proposed enhancement in this paper does contribute to the
improvement of the accuracy For example if we compare
the ldquoF-Measurementrdquo (row six of Table II) values without
lost critical samples retransmitted and with lost critical
samples retransmitted (+cs) we will find that the values
have been improved more than 5
TABLE II ACCURACY COMPARISON AMONG DIFFERENT STRATEGIES
USING NBC MODEL
Strategy Name
Existing Existing+cs LASP-cs LASP
TP Rate 0791 0803 0824 0837
FP Rate 0115 0112 0090 0089
Precision 0809 0814 0831 0847
Recall 0793 0904 0815 0817
F-Measure 0785 0801 0821 0824
ROC Area 0924 0927 0946 0949
TABLE III ACCURACY COMPARISON AMONG DIFFERENT STRATEGIES
USING SVM MODEL
Strategy Name
Existing Existing+cs LASP-cs LASP
TP Rate 0633 0645 0693 0699
FP Rate 0174 0169 0147 0145
Precision 0836 0839 0849 0851
Recall 0635 0645 0692 0699
F-Measure 0631 0630 0689 0695
ROC Area 0728 0737 0783 0782
We can draw similar conclusions from Table III if we
employ the second classification model (SVM
classification model) Again we only show the results
when the loss rate is 20 and the results under other loss
rate are similar As shown in Table III the average
performance improvements are around 10 We also
performs two sample t tests Results show that the p-value
was less than 5 which means the performance
improvements are statistically significant
2) Accuracy under various loss rates
We then study whether or not the proposed LASP
strategy is robust under different loss rates To do so we
evaluated the classification accuracy under different loss
rate within LASP Specifically we evaluated the accuracy
(precision and recall) under three loss rates 10 20
and 30 using the first classification model (NBC
model) Fig 7 shows the precision and recall of our NBC
classification algorithms As shown in this figure while
the classification performance does suffering when the
loss rate increases the reduction of classification
performance is minor For example when the loss rate is
increased for 20 (from 10 to 30) the precision and
recall are reduced by less than 3 These results are
strong indication that our proposed LASP strategy are
very robust for sample loss
Fig 7 Accuracy of LASP under different loss rate with NBC
classification model
3) Comparison with various sample lengths
Finally we conducted a third experimentin order to
determine whether the proposed approach can reduce the
detection time For example given two groups of time
series T1 and T2 we apply the same classification model
to both of them If the classification accuracy for T1 is
better than the result for T2 and the difference between the
accuracy is statistically significant we can claim that
strategy for T1 is more effective (in terms of data analytics)
than the strategy for T2 In our experiments we choose
three lengths 15 seconds (L1) 30 seconds (L2) and 45
seconds (L3) For each length we generate two data
streams the first data stream is generated by sequential
packetization (ldquoExistingrdquo) and the second data stream is
generated by the proposed strategy (LASP) As shown in
Table IV each column represents the results of data
stream with different length For example the second row
(ldquoExisting-L1rdquo) indicates the data stream whose length is
15 seconds (L1) using the sequential packetization
(ldquoExistingrdquo) strategy Each row represents different
accuracy measurements which are introduced in the
beginning of section VB From Table IV we can see that
given the same length of the data stream the accuracy
measurement for data stream from LASP is substantially
higher than the accuracy measurement for data stream
from ldquoExistingrdquo strategy This verifies that if we employ
the proposed approach we could achieve faster detection
speed In another world LASP has a good potential in
applications where real-time classification is desired
TABLE IV ACCURACY COMPARISON AMONG DIFFERENT LENGTHS
GENERATED BY DIFFERENT STRATEGIES
Strategy Existing LASP
L1 L2 L3 L1 L2 L3
TP Rate 0582 0609 0611 0655 0677 0681
FP Rate 0228 0206 0201 0188 0164 0167
Precision 0789 0811 0815 0804 0828 0827
Recall 0570 06 0611 0645 0679 0688
F-Measure 0558 0581 0587 0645 0673 0677
ROC Area 0679 0699 0701 0729 0755 0759
VI SUMMARY
Existing strategies for transmission of time series body
sensor data assumes sequential sample packetization
Journal of Communications Vol 10 No 11 November 2015
857copy2015 Journal of Communications
which leads to difficulty of data recovery in the high loss
ratio WBANs In this paper we investigated the issue of
sample packetization pattern and proposed a heuristic
approach to minimize the effect of channel loss on success
of data analysis The proposed approach is very efficient
for sensor data processing incurs no extra overhead and
does not require modification of the transport protocol
In the future we will implement the packetization
scheme in real sensor network platform such as TinyOS
[32] and experiment with more biomedical sensor data to
validate its effectiveness
ACKNOWLEDGMENT
The authors wish to thank ICNC 2015 reviewers for
their valuable comments that significantly improved the
quality of the paper This work was partially supported by
the US National Science Foundation under Grant No
1229213
REFERENCES
[1] G Z Yang Body Sensor Networks New York USA Springer
Science+Business Media LLC 2006
[2] C Cordeiro and M Patel ldquoBody area networking standardization
present and future directionsrdquo presented at the Proceedings of the
ICST 2nd International Conference on Body Area Networks
Florence Italy 2007
[3] J Ko C Lu M B Srivastava J A Stankovic A Terzis and M
Welsh ldquoWireless sensor networks for healthcarerdquo Proceedings of
the IEEE vol 98 pp 1947-1960 2010
[4] O Amft H Junker P Lukowicz G Troster and C Schuster
ldquoSensing muscle activities with body-worn sensorsrdquo in Proc
International Workshop on Wearable and Implantable Body
Sensor Networks 2006
[5] B French A Smailagic D Siewiorek V Ambur and D
Tyamagundlu ldquoClassifying wheelchair propulsion patterns with a
wrist mounted accelerometerrdquo presented at the Proceedings of the
ICST 3rd International Conference on Body Area Networks
Tempe Arizona 2008
[6] L Jones N Deo and B Lockyer ldquoWireless physiological sensor
system for ambulatory userdquo in Proc International Workshop on
Wearable and Implantable Body Sensor Networks 2006 pp 4
pp-149
[7] K V Laerhoven H W Gellersen and Y G Malliaris ldquoLong
term activity monitoring with a wearable sensor noderdquo in Proc
International Workshop on Wearable and Implantable Body
Sensor Networks 2006
[8] U Maurer A Smailagic D P Siewiorek and M Deisher
ldquoActivity recognition and monitoring using multiple sensors on
different body positionsrdquo in Proc International Workshop on
Wearable and Implantable Body Sensor Networks 2006
[9] M Quwaider and S Biswas ldquoBody posture identification using
hidden Markov model with a wearable sensor networkrdquo presented
at the Proceedings of the ICST 3rd International Conference on
Body Area Networks Tempe Arizona 2008
[10] R Chavarriaga H Sagha A Calatroni S T Digumarti G Tr et
al ldquoThe opportunity challenge A benchmark database for on-
body sensor-based activity recognitionrdquo Pattern Recogn Lett vol
34 pp 2033-2042 2013
[11] M Zhang and A A Sawchuk ldquoUSC-HAD A daily activity
dataset for ubiquitous activity recognition using wearable sensorsrdquo
presented at the Proceedings of the 2012 ACM Conference on
Ubiquitous Computing Pittsburgh Pennsylvania 2012
[12] A Bulling U Blanke and B Schiele ldquoA tutorial on human
activity recognition using body-worn inertial sensorsrdquo ACM
Comput Surv vol 46 pp 1-33 2014
[13] G Zhou J Lu C Wan M Yarvis and J Stankovic ldquoBodyqos
Adaptive and radio-agnostic qos for body sensor networksrdquo in
Proc IEEE INFOCOM Phoenix AZ 2008 pp 565-573
[14] G Zhou C Y Wan M D Yarvis and J A Stankovic
ldquoAggregator-centric QoS for body sensor networksrdquo presented at
the Proceedings of the 6th International Conference on
Information Processing in Sensor Networks Cambridge
Massachusetts USA 2007
[15] B Otal L Alonso and C Verikoukis ldquoNovel QoS scheduling
and energy-saving MAC protocol for body sensor networks
optimizationrdquo presented at the Proceedings of the ICST 3rd
International Conference on Body Area Networks Tempe
Arizona 2008
[16] J J Garcia and T Falck ldquoQuality of service for IEEE 802154-
based wireless body sensor networksrdquo in Proc 3rd International
Conference on Pervasive Computing Technologies for Healthcare
2009 pp 1-6
[17] F Gengfa and E Dutkiewicz ldquoBodyMAC Energy efficient
TDMA-based MAC protocol for wireless body area networksrdquo in
Proc 9th International Symposium on Communications and
Information Technology 2009 pp 1455-1459
[18] N Read M Li Y Cao S H Liu T Wilson and B Prabhakaran
ldquoLoss resilient strategy in body sensor networksrdquo presented at the
Proc of ACMIEEE International Conference on Body Area
Networks (BodyNets) Beijing China 2011
[19] E Weldon Jr ldquoAn improved Selective-Repeat ARQ strategyrdquo
IEEE Transactions on Communications vol 30 pp 480-486
1982
[20] T Fulford-Jones D Malan M Welsh and S Moulton
ldquoCodeBlue An ad hoc sensor network infrastructure for
emergency medical carerdquo in Proc International Workshop on
Body Sensor Networks 2004
[21] S Jiang Y Cao S Iyengar P Kuryloski R Jafari Y Xue et al
ldquoCareNet An integrated wireless sensor networking environment
for remote healthcarerdquo presented at the Proceedings of the ICST
3rd International Conference on Body Area Networks Tempe
Arizona 2008
[22] M Younis K Akkaya M Eltoweissy and A Wadaa ldquoOn
handling QoS traffic in wireless sensor networksrdquo in Proc 37th
Annual Hawaii International Conference on System Sciences Big
Island HI USA 2004
[23] P Baronti P Pillai V W C Chook S Chessa A Gotta and Y
F Hu ldquoWireless sensor networks A survey on the state of the art
and the 80215 4 and ZigBee standardsrdquo Computer
Communications vol 30 pp 1655-1695 2007
[24] Bluetooth Low Energy Core Specification Version 40 [Online]
AvailablehttpwwwbluetoothcomEnglishTechnologyWorks
PagesBluetooth_low_energy_technologyaspx
[25] K K Sup M A Ameen K Daehan L Cheolhyo and L
Hyungsoo ldquoA study on proposed IEEE 80215 WBAN MAC
protocolsrdquo in Proc 9th International Symposium on
Communications and Information Technology 2009 pp 834-840
[26] M van der Schaar and D S Turaga ldquoCross-Layer packetization
and retransmission strategies for delay-sensitive wireless
multimedia transmissionrdquo IEEE Transactions on Multimedia vol
9 pp 185-197 2007
[27] S Zhang S C Liew and P P Lam ldquoHot topic Physical-layer
network codingrdquo presented at the Proceedings of the 12th annual
International Conference on Mobile Computing and Networking
Los Angeles CA USA 2006
[28] J Ellul B Lo and G Z Yang ldquoThe BSNOS platform A body
sensor networks targeted operating system and toolsetrdquo in Proc
Journal of Communications Vol 10 No 11 November 2015
858copy2015 Journal of Communications
Fifth International Conference on Sensor Technologies and
Applications 2011 pp 381-386
[29] M Li Y Cao and B Prabhakaran ldquoLoss aware sample
packetization strategy for improvement of body sensor data
analysisrdquo in Proc International Conference on Computing
Networking and Communications Anaheim California 2015
[30] L Bao and S S Intille ldquoActivity recognition from user-annotated
acceleration datardquo in Pervasive Computing Springer 2004 pp 1-
17
[31] J R Kwapisz G M Weiss and S A Moore ldquoActivity
recognition using cell phone accelerometersrdquo ACM SigKDD
Explorations Newsletter vol 12 pp 74-82 2011
[32] TinyOS [Online] Available wwwtinyosnet
Ming Li is currently an Associate Professor
and Chair in the Department of Computer
Science California State University Fresno
Prior to that he was an Assistant Professor
from August 2006 to 2012 He received his
MS and PhD degrees in Computer Science
from The University of Texas at Dallas in
2001 and 2006 respectively His research
interests include QoS strategies for IEEE
80211 wireless LANs mobile ad-hoc networks and heterogeneous
wired and wireless networks multimedia streaming over wireless
networks body area networks and robot swarm communications
Yu Cao has been an Assistant Professor at the
Department of Computer Science The
University of Massachusetts Lowell since Aug
2013 From 2010 to 2013 he was a faculty at
The University of Tennessee at Chattanooga
(UTC) From 2007 to 2010 he was a faculty at
California State University Fresno Prior to
that he was a Visiting Fellow of Biomedical
Engineering at Mayo Clinic Rochester
Minnesota He received his MS and PhD degrees in Computer
Science from Iowa State University in 2005 and 2007 respectively He
received the BEng degree from Harbin Engineering University in 1997
the MEng degree from Huazhong University of Science and
Technology in 2000 all in Computer Science His research interests
span a variety of aspects of knowledge discover from complex data
which include the area of biomedical informatics and intelligent system
B Prabhakaran is a Professor of Computer
Science in the University of Texas at Dallas
Prof B Prabhakaran works in the area of
multimedia systems He has published several
research papers in prestigious conferences and
journals in this area Dr Prabhakaran received
the prestigious NSF CAREER Award FY
2003 for his proposal on Animation Databases
Dr Prabhakaran is General Co-Chair of ACM
Multimedia 2011 He is also Technical Program Co-Chair of IEEE
WoWMoM 2012 (World of Wireless Mobile and Multimedia
Networks) He served as the TPC Co-Chair of IEEE ISM 2010
(International Symposium on Multimedia) Dr Prabhakaran is a
Member of the Executive Council of the ACM Special Interest Group
on Multimedia (SIGMM) and is the Co-Chair of IEEE Technical
Committe on Multimedia Computing (TCMC) Special Interest Group
on Video Analytics (SIGVA) Dr Prabhakaran is the Editor-in-Chief of
the ACM SIGMM (Special Interest Group on Multimedia) web
magazine He is Member of the Editorial board of Multimedia Systems
Journal (Springer) and Multimedia Tools and Applications journal
(Springer) He has served as guest-editor (special issue on Multimedia
Authoring and Presentation) for ACM Multimedia Systems journal
Page 4
procedure of the proposed approach which is described
as follows
Step 1 for each collected sample group packetization
is performed using alternate pattern Assuming N samples
and M packets for each sample i (0ltiltN) the ID of
corresponding packet that it should be inserted into is
iM
Step 2 for the same group all critical samples are
identified A sample i is considered critical if |estimatedi-
originali| gt α originali Here the estimated value depends
on the specific estimation method and the values of the
neighboring samples α is a constant and is fixed as 025
in this paper
Step 3 all packets are sent to the aggregator At the
same time a special packet containing the IDs of all
critical samples is also sent
Step 4 upon the receipt of packets data aggregator re-
assembles received samples based on the same pattern
Therefore the kth sample in jth packet has the sample ID
of j+kN
Step 5 aggregator identifies lost critical samples by
comparing the IDs of the lost samples and the IDs of all
critical samples
Step 6 aggregator requests the sensor to retransmit
critical samples and ensure its successful delivery
Step 7 aggregator performs data recovery for other
lost samples before performing analysis
Fig 3 Comparison of packetization approaches of 16 samples
It should be noted that the proposed strategy is an
application layer approach and therefore is different from
the physical layer network coding [27] which aims at
coordination of multiple node transmission in order to
achieve higher network throughput especially in wireless
networks However the system performance will
potentially benefit from physical layer network coding if
implemented
C Data Quality Model
In this section we present a variance based data
quality model for the analysis of data recovery ratio We
define estimation variance (EV) as the relative difference
between the estimated sample values and original values
which can calculate it as
(1)
where di is the original value of the sample i and esti is the
estimated value of sample i from received samples
Further the total estimation variance is calculated as
(2)
(3)
where K is the total number of estimated samples in any
evaluation group Clearly smaller EV and AEV indicates
better data recovery which is critical for the success of
data analysis
Now letrsquos assume that a group of M independent
packets are transmitted For each packet j its total
estimation variance of all samples in the packet EV(j)
depends on the specific characteristics of the data samples
in the packet Thus for a given packet loss ratio of l the
expected total estimation variance is calculated as
If EV(j) is the same value (EV) for all packets which is
not usual in real scenarios due to the statistic dropping of
packets we then have that
Clearly the total EV depends on total number of
packets loss ratio and data characteristics Further
average expected estimation variance is now
(6)
That means if packets are independently recovered
then loss ratio does not affect the level of data recovery
However when two or more adjacent packets are lost it is
difficult to have independent data recovery therefore
higher loss ration does decrease the data recovery ratio
TABLE I EV VS DATA CHARACTERISTICS
Strategy Data Characteristics
sqrt linear square cubic quartic
Sequential 0075 0 039 090 140
Alternate 0003 0 002 006 012
To illustrate the effect of the packetization on data with
different characteristics we take a data segment that
follows different patterns and calculate the total EV for
the segment with two different packetization schemes
Each data segment contains 8 samples which are
packetized into two different packets with either the
ldquosequentialrdquo or ldquoalternaterdquo approach under the condition
that the second packet is lost Table I shows the result It
can be seen that as the curve deviates more from the linear
pattern higher variance is yielded for both ldquosequentialrdquo
and ldquoalternaterdquo approach However ldquoalternaterdquo approach
shows a much slower increase compared with ldquosequentialrdquo
V PERFORMANCE EVALUATION
We performed extensive simulations to study the
performance of the proposed packetization strategy For
simplicity we label the proposed strategy as LASP and
sequential packetization as ldquoExistingrdquo The following four
strategies were compared
Existing samples are packetized in order of creation
ie sequential
Collect
Samples
Perform
PacketizationSample
Assembling
Identify Critical
Samples
Request Lost
Critical Samples
Resend Critical
Samples
Data
Recovery
Sensor Side Aggregator
iiii ddestEV
i iEVEV
KEVAEV i i
M
m
m
j
mMm jEVllm
MEV
0 1
)()1(
EVlMmllm
MEVEV
M
m
mMm
0
)1(
EVlMEVlMAEV
Journal of Communications Vol 10 No 11 November 2015
854copy2015 Journal of Communications
(4)
(5)
Finally the Average Estimation Variance (AEV) is
calculated as
Existing+cs existing approach with lost critical
samples retransmitted
LASP proposed approach described in section 42
LASP-cs proposed approach but lost critical samples
are not retransmitted
We use the Body Sensor Network (BSN) Development
Kit v3 [28] developed by Prof Guang-Zhong Yang of
Imperial College to collect the accelerometer data from
users The BSN development kit v3 includes a USB
programming board a sensor board a battery board a
prototype board and a pair of the new BSN v3 nodes The
sensor board has a 3D accelerometer and a temperature
sensor In our experiments we recruit three subjects and
attach the sensor board to the arm of the subject Each
subject is asked to perform three types of Activity of Daily
Living (ADL) jogging sitting and walking The duration
of each activity is around 3 minutes The 3D
accelerometer data was collected when the subjects
performed the ADL activity We chose to use the
sampling rate as 10 Hz We also implemented linear
interpolation to calculate a sample directly between any
two neighboring ones the interpolation effectively yields
a 33 Hz sampling rate with the smoothness of the 16 Hz
rate Using the 10 Hz sampling rate with linear
interpolation yielded a rather smooth graph without a lot
of noise and yet it was fast enough that the peaks and
troughs of ADL were all present
Our evaluation consists of two parts In part 1 received
samples with different approaches are compared in term
of several key metrics In part 2 recovered data are
analyzed by various representative analysis algorithms for
comparison of the accuracy
Fig 4 Comparison of Average Estimation Variance (AEV)
A Sample Quality Analysis
For each received samples we evaluate how
satisfactory the data is recovered by comparing the
estimated values with original ones Fig 4 shows the
comparison of average estimation variance of lost samples
for x values of jogging data It can be seen that LASP has
only about 6-9 normalized variance on average while
existing approach yields an average of 15-17 Also
retransmission of critical samples does not help much for
existing approach mainly due to the high estimation
inaccuracy when a contiguous sample train gets lost Fig 5
compares number of lost samples with normalized
variance higher than 25 Itrsquos clear that LASP is much
better than existing approach even if critical samples are
not retransmitted This is due to the fact that LASP creates
a much nicer sample loss pattern as described earlier and
therefore minimizes the variance for lost samples
Fig 5 Number of samples with low recovery ratio
Fig 6 Comparison of data recovery
To provide a visual representation of how well LASP
improves data recovery 100 samples are extracted from a
sequence of 1043 samples for jogging Itrsquos clear that
LASP while having some variance creates a good
approximation of the original values However existing
approach easily mis- estimate many samples
B Data Analysis
In order to determine whether the proposed
packetization strategy could contribute to improved
accuracy of data analysis we design our experiments as
follow First we employ the sliding window techniques
with dynamic window length For each sliding window
we extract high level features (eg average absolute
difference time between peaks and etc) instead of using
the raw sensor value Second we choose two widely used
data mining and machine learning classification
algorithms Naive Bayes Classifier and Support Vector
Machines (SVM) Classifier Next we train two models
using the aforementioned classifiers We define these two
models as ldquoNBCrdquo (which is trained using Naive Bayes
Classifier) and ldquoSVMrdquo (which is trained using Support
Vector Machines) Once the two models were trained we
will apply each model to the data sets generated by the
four strategies (introduced in the beginning of Section 5)
For each testing six accuracy measurements are used
True Positive Rate (TP Rate) False Positive Rate (FPR)
Precision Recall F-Measurement and Area Under an
ROC Curve (ROC Area)
We will first introduce our feature extraction
techniques for generating semantic features from sliding
windows In our previous work [29] we mainly employ
raw time-series data In the experiments we extend our
previous work by employing six types of features which
are listed as follows The first and second type of feature
is Average and Standard Deviation respectively The
third type of feature is called Average Resultant
Acceleration [30] [31] which is defined as the average
of the square roots of the summation of the sensor
Journal of Communications Vol 10 No 11 November 2015
855copy2015 Journal of Communications
Journal of Communications Vol 10 No 11 November 2015
856copy2015 Journal of Communications
readings within each sliding window The fourth type of
feature is named as Average Absolute Difference [30]
[31] which is defined as the average absolute difference
between the value of each of the raw sensor reading in the
sliding window and the mean value over of all the sensor
reading The fifth category of features is called Time
Between Peaks [30] which is defined as time differences
(calculated in milliseconds) between peaks in the
sinusoidal waveform from the raw sensor readings within
the sliding window
1) Comparison with various models
Table II shows the detailed results using the NBC
model As a representative scenario the results for a loss
rate of 20 are studied As we can see from this table the
proposed enhancement in this paper does contribute to the
improvement of the accuracy For example if we compare
the ldquoF-Measurementrdquo (row six of Table II) values without
lost critical samples retransmitted and with lost critical
samples retransmitted (+cs) we will find that the values
have been improved more than 5
TABLE II ACCURACY COMPARISON AMONG DIFFERENT STRATEGIES
USING NBC MODEL
Strategy Name
Existing Existing+cs LASP-cs LASP
TP Rate 0791 0803 0824 0837
FP Rate 0115 0112 0090 0089
Precision 0809 0814 0831 0847
Recall 0793 0904 0815 0817
F-Measure 0785 0801 0821 0824
ROC Area 0924 0927 0946 0949
TABLE III ACCURACY COMPARISON AMONG DIFFERENT STRATEGIES
USING SVM MODEL
Strategy Name
Existing Existing+cs LASP-cs LASP
TP Rate 0633 0645 0693 0699
FP Rate 0174 0169 0147 0145
Precision 0836 0839 0849 0851
Recall 0635 0645 0692 0699
F-Measure 0631 0630 0689 0695
ROC Area 0728 0737 0783 0782
We can draw similar conclusions from Table III if we
employ the second classification model (SVM
classification model) Again we only show the results
when the loss rate is 20 and the results under other loss
rate are similar As shown in Table III the average
performance improvements are around 10 We also
performs two sample t tests Results show that the p-value
was less than 5 which means the performance
improvements are statistically significant
2) Accuracy under various loss rates
We then study whether or not the proposed LASP
strategy is robust under different loss rates To do so we
evaluated the classification accuracy under different loss
rate within LASP Specifically we evaluated the accuracy
(precision and recall) under three loss rates 10 20
and 30 using the first classification model (NBC
model) Fig 7 shows the precision and recall of our NBC
classification algorithms As shown in this figure while
the classification performance does suffering when the
loss rate increases the reduction of classification
performance is minor For example when the loss rate is
increased for 20 (from 10 to 30) the precision and
recall are reduced by less than 3 These results are
strong indication that our proposed LASP strategy are
very robust for sample loss
Fig 7 Accuracy of LASP under different loss rate with NBC
classification model
3) Comparison with various sample lengths
Finally we conducted a third experimentin order to
determine whether the proposed approach can reduce the
detection time For example given two groups of time
series T1 and T2 we apply the same classification model
to both of them If the classification accuracy for T1 is
better than the result for T2 and the difference between the
accuracy is statistically significant we can claim that
strategy for T1 is more effective (in terms of data analytics)
than the strategy for T2 In our experiments we choose
three lengths 15 seconds (L1) 30 seconds (L2) and 45
seconds (L3) For each length we generate two data
streams the first data stream is generated by sequential
packetization (ldquoExistingrdquo) and the second data stream is
generated by the proposed strategy (LASP) As shown in
Table IV each column represents the results of data
stream with different length For example the second row
(ldquoExisting-L1rdquo) indicates the data stream whose length is
15 seconds (L1) using the sequential packetization
(ldquoExistingrdquo) strategy Each row represents different
accuracy measurements which are introduced in the
beginning of section VB From Table IV we can see that
given the same length of the data stream the accuracy
measurement for data stream from LASP is substantially
higher than the accuracy measurement for data stream
from ldquoExistingrdquo strategy This verifies that if we employ
the proposed approach we could achieve faster detection
speed In another world LASP has a good potential in
applications where real-time classification is desired
TABLE IV ACCURACY COMPARISON AMONG DIFFERENT LENGTHS
GENERATED BY DIFFERENT STRATEGIES
Strategy Existing LASP
L1 L2 L3 L1 L2 L3
TP Rate 0582 0609 0611 0655 0677 0681
FP Rate 0228 0206 0201 0188 0164 0167
Precision 0789 0811 0815 0804 0828 0827
Recall 0570 06 0611 0645 0679 0688
F-Measure 0558 0581 0587 0645 0673 0677
ROC Area 0679 0699 0701 0729 0755 0759
VI SUMMARY
Existing strategies for transmission of time series body
sensor data assumes sequential sample packetization
Journal of Communications Vol 10 No 11 November 2015
857copy2015 Journal of Communications
which leads to difficulty of data recovery in the high loss
ratio WBANs In this paper we investigated the issue of
sample packetization pattern and proposed a heuristic
approach to minimize the effect of channel loss on success
of data analysis The proposed approach is very efficient
for sensor data processing incurs no extra overhead and
does not require modification of the transport protocol
In the future we will implement the packetization
scheme in real sensor network platform such as TinyOS
[32] and experiment with more biomedical sensor data to
validate its effectiveness
ACKNOWLEDGMENT
The authors wish to thank ICNC 2015 reviewers for
their valuable comments that significantly improved the
quality of the paper This work was partially supported by
the US National Science Foundation under Grant No
1229213
REFERENCES
[1] G Z Yang Body Sensor Networks New York USA Springer
Science+Business Media LLC 2006
[2] C Cordeiro and M Patel ldquoBody area networking standardization
present and future directionsrdquo presented at the Proceedings of the
ICST 2nd International Conference on Body Area Networks
Florence Italy 2007
[3] J Ko C Lu M B Srivastava J A Stankovic A Terzis and M
Welsh ldquoWireless sensor networks for healthcarerdquo Proceedings of
the IEEE vol 98 pp 1947-1960 2010
[4] O Amft H Junker P Lukowicz G Troster and C Schuster
ldquoSensing muscle activities with body-worn sensorsrdquo in Proc
International Workshop on Wearable and Implantable Body
Sensor Networks 2006
[5] B French A Smailagic D Siewiorek V Ambur and D
Tyamagundlu ldquoClassifying wheelchair propulsion patterns with a
wrist mounted accelerometerrdquo presented at the Proceedings of the
ICST 3rd International Conference on Body Area Networks
Tempe Arizona 2008
[6] L Jones N Deo and B Lockyer ldquoWireless physiological sensor
system for ambulatory userdquo in Proc International Workshop on
Wearable and Implantable Body Sensor Networks 2006 pp 4
pp-149
[7] K V Laerhoven H W Gellersen and Y G Malliaris ldquoLong
term activity monitoring with a wearable sensor noderdquo in Proc
International Workshop on Wearable and Implantable Body
Sensor Networks 2006
[8] U Maurer A Smailagic D P Siewiorek and M Deisher
ldquoActivity recognition and monitoring using multiple sensors on
different body positionsrdquo in Proc International Workshop on
Wearable and Implantable Body Sensor Networks 2006
[9] M Quwaider and S Biswas ldquoBody posture identification using
hidden Markov model with a wearable sensor networkrdquo presented
at the Proceedings of the ICST 3rd International Conference on
Body Area Networks Tempe Arizona 2008
[10] R Chavarriaga H Sagha A Calatroni S T Digumarti G Tr et
al ldquoThe opportunity challenge A benchmark database for on-
body sensor-based activity recognitionrdquo Pattern Recogn Lett vol
34 pp 2033-2042 2013
[11] M Zhang and A A Sawchuk ldquoUSC-HAD A daily activity
dataset for ubiquitous activity recognition using wearable sensorsrdquo
presented at the Proceedings of the 2012 ACM Conference on
Ubiquitous Computing Pittsburgh Pennsylvania 2012
[12] A Bulling U Blanke and B Schiele ldquoA tutorial on human
activity recognition using body-worn inertial sensorsrdquo ACM
Comput Surv vol 46 pp 1-33 2014
[13] G Zhou J Lu C Wan M Yarvis and J Stankovic ldquoBodyqos
Adaptive and radio-agnostic qos for body sensor networksrdquo in
Proc IEEE INFOCOM Phoenix AZ 2008 pp 565-573
[14] G Zhou C Y Wan M D Yarvis and J A Stankovic
ldquoAggregator-centric QoS for body sensor networksrdquo presented at
the Proceedings of the 6th International Conference on
Information Processing in Sensor Networks Cambridge
Massachusetts USA 2007
[15] B Otal L Alonso and C Verikoukis ldquoNovel QoS scheduling
and energy-saving MAC protocol for body sensor networks
optimizationrdquo presented at the Proceedings of the ICST 3rd
International Conference on Body Area Networks Tempe
Arizona 2008
[16] J J Garcia and T Falck ldquoQuality of service for IEEE 802154-
based wireless body sensor networksrdquo in Proc 3rd International
Conference on Pervasive Computing Technologies for Healthcare
2009 pp 1-6
[17] F Gengfa and E Dutkiewicz ldquoBodyMAC Energy efficient
TDMA-based MAC protocol for wireless body area networksrdquo in
Proc 9th International Symposium on Communications and
Information Technology 2009 pp 1455-1459
[18] N Read M Li Y Cao S H Liu T Wilson and B Prabhakaran
ldquoLoss resilient strategy in body sensor networksrdquo presented at the
Proc of ACMIEEE International Conference on Body Area
Networks (BodyNets) Beijing China 2011
[19] E Weldon Jr ldquoAn improved Selective-Repeat ARQ strategyrdquo
IEEE Transactions on Communications vol 30 pp 480-486
1982
[20] T Fulford-Jones D Malan M Welsh and S Moulton
ldquoCodeBlue An ad hoc sensor network infrastructure for
emergency medical carerdquo in Proc International Workshop on
Body Sensor Networks 2004
[21] S Jiang Y Cao S Iyengar P Kuryloski R Jafari Y Xue et al
ldquoCareNet An integrated wireless sensor networking environment
for remote healthcarerdquo presented at the Proceedings of the ICST
3rd International Conference on Body Area Networks Tempe
Arizona 2008
[22] M Younis K Akkaya M Eltoweissy and A Wadaa ldquoOn
handling QoS traffic in wireless sensor networksrdquo in Proc 37th
Annual Hawaii International Conference on System Sciences Big
Island HI USA 2004
[23] P Baronti P Pillai V W C Chook S Chessa A Gotta and Y
F Hu ldquoWireless sensor networks A survey on the state of the art
and the 80215 4 and ZigBee standardsrdquo Computer
Communications vol 30 pp 1655-1695 2007
[24] Bluetooth Low Energy Core Specification Version 40 [Online]
AvailablehttpwwwbluetoothcomEnglishTechnologyWorks
PagesBluetooth_low_energy_technologyaspx
[25] K K Sup M A Ameen K Daehan L Cheolhyo and L
Hyungsoo ldquoA study on proposed IEEE 80215 WBAN MAC
protocolsrdquo in Proc 9th International Symposium on
Communications and Information Technology 2009 pp 834-840
[26] M van der Schaar and D S Turaga ldquoCross-Layer packetization
and retransmission strategies for delay-sensitive wireless
multimedia transmissionrdquo IEEE Transactions on Multimedia vol
9 pp 185-197 2007
[27] S Zhang S C Liew and P P Lam ldquoHot topic Physical-layer
network codingrdquo presented at the Proceedings of the 12th annual
International Conference on Mobile Computing and Networking
Los Angeles CA USA 2006
[28] J Ellul B Lo and G Z Yang ldquoThe BSNOS platform A body
sensor networks targeted operating system and toolsetrdquo in Proc
Journal of Communications Vol 10 No 11 November 2015
858copy2015 Journal of Communications
Fifth International Conference on Sensor Technologies and
Applications 2011 pp 381-386
[29] M Li Y Cao and B Prabhakaran ldquoLoss aware sample
packetization strategy for improvement of body sensor data
analysisrdquo in Proc International Conference on Computing
Networking and Communications Anaheim California 2015
[30] L Bao and S S Intille ldquoActivity recognition from user-annotated
acceleration datardquo in Pervasive Computing Springer 2004 pp 1-
17
[31] J R Kwapisz G M Weiss and S A Moore ldquoActivity
recognition using cell phone accelerometersrdquo ACM SigKDD
Explorations Newsletter vol 12 pp 74-82 2011
[32] TinyOS [Online] Available wwwtinyosnet
Ming Li is currently an Associate Professor
and Chair in the Department of Computer
Science California State University Fresno
Prior to that he was an Assistant Professor
from August 2006 to 2012 He received his
MS and PhD degrees in Computer Science
from The University of Texas at Dallas in
2001 and 2006 respectively His research
interests include QoS strategies for IEEE
80211 wireless LANs mobile ad-hoc networks and heterogeneous
wired and wireless networks multimedia streaming over wireless
networks body area networks and robot swarm communications
Yu Cao has been an Assistant Professor at the
Department of Computer Science The
University of Massachusetts Lowell since Aug
2013 From 2010 to 2013 he was a faculty at
The University of Tennessee at Chattanooga
(UTC) From 2007 to 2010 he was a faculty at
California State University Fresno Prior to
that he was a Visiting Fellow of Biomedical
Engineering at Mayo Clinic Rochester
Minnesota He received his MS and PhD degrees in Computer
Science from Iowa State University in 2005 and 2007 respectively He
received the BEng degree from Harbin Engineering University in 1997
the MEng degree from Huazhong University of Science and
Technology in 2000 all in Computer Science His research interests
span a variety of aspects of knowledge discover from complex data
which include the area of biomedical informatics and intelligent system
B Prabhakaran is a Professor of Computer
Science in the University of Texas at Dallas
Prof B Prabhakaran works in the area of
multimedia systems He has published several
research papers in prestigious conferences and
journals in this area Dr Prabhakaran received
the prestigious NSF CAREER Award FY
2003 for his proposal on Animation Databases
Dr Prabhakaran is General Co-Chair of ACM
Multimedia 2011 He is also Technical Program Co-Chair of IEEE
WoWMoM 2012 (World of Wireless Mobile and Multimedia
Networks) He served as the TPC Co-Chair of IEEE ISM 2010
(International Symposium on Multimedia) Dr Prabhakaran is a
Member of the Executive Council of the ACM Special Interest Group
on Multimedia (SIGMM) and is the Co-Chair of IEEE Technical
Committe on Multimedia Computing (TCMC) Special Interest Group
on Video Analytics (SIGVA) Dr Prabhakaran is the Editor-in-Chief of
the ACM SIGMM (Special Interest Group on Multimedia) web
magazine He is Member of the Editorial board of Multimedia Systems
Journal (Springer) and Multimedia Tools and Applications journal
(Springer) He has served as guest-editor (special issue on Multimedia
Authoring and Presentation) for ACM Multimedia Systems journal
Page 5
Existing+cs existing approach with lost critical
samples retransmitted
LASP proposed approach described in section 42
LASP-cs proposed approach but lost critical samples
are not retransmitted
We use the Body Sensor Network (BSN) Development
Kit v3 [28] developed by Prof Guang-Zhong Yang of
Imperial College to collect the accelerometer data from
users The BSN development kit v3 includes a USB
programming board a sensor board a battery board a
prototype board and a pair of the new BSN v3 nodes The
sensor board has a 3D accelerometer and a temperature
sensor In our experiments we recruit three subjects and
attach the sensor board to the arm of the subject Each
subject is asked to perform three types of Activity of Daily
Living (ADL) jogging sitting and walking The duration
of each activity is around 3 minutes The 3D
accelerometer data was collected when the subjects
performed the ADL activity We chose to use the
sampling rate as 10 Hz We also implemented linear
interpolation to calculate a sample directly between any
two neighboring ones the interpolation effectively yields
a 33 Hz sampling rate with the smoothness of the 16 Hz
rate Using the 10 Hz sampling rate with linear
interpolation yielded a rather smooth graph without a lot
of noise and yet it was fast enough that the peaks and
troughs of ADL were all present
Our evaluation consists of two parts In part 1 received
samples with different approaches are compared in term
of several key metrics In part 2 recovered data are
analyzed by various representative analysis algorithms for
comparison of the accuracy
Fig 4 Comparison of Average Estimation Variance (AEV)
A Sample Quality Analysis
For each received samples we evaluate how
satisfactory the data is recovered by comparing the
estimated values with original ones Fig 4 shows the
comparison of average estimation variance of lost samples
for x values of jogging data It can be seen that LASP has
only about 6-9 normalized variance on average while
existing approach yields an average of 15-17 Also
retransmission of critical samples does not help much for
existing approach mainly due to the high estimation
inaccuracy when a contiguous sample train gets lost Fig 5
compares number of lost samples with normalized
variance higher than 25 Itrsquos clear that LASP is much
better than existing approach even if critical samples are
not retransmitted This is due to the fact that LASP creates
a much nicer sample loss pattern as described earlier and
therefore minimizes the variance for lost samples
Fig 5 Number of samples with low recovery ratio
Fig 6 Comparison of data recovery
To provide a visual representation of how well LASP
improves data recovery 100 samples are extracted from a
sequence of 1043 samples for jogging Itrsquos clear that
LASP while having some variance creates a good
approximation of the original values However existing
approach easily mis- estimate many samples
B Data Analysis
In order to determine whether the proposed
packetization strategy could contribute to improved
accuracy of data analysis we design our experiments as
follow First we employ the sliding window techniques
with dynamic window length For each sliding window
we extract high level features (eg average absolute
difference time between peaks and etc) instead of using
the raw sensor value Second we choose two widely used
data mining and machine learning classification
algorithms Naive Bayes Classifier and Support Vector
Machines (SVM) Classifier Next we train two models
using the aforementioned classifiers We define these two
models as ldquoNBCrdquo (which is trained using Naive Bayes
Classifier) and ldquoSVMrdquo (which is trained using Support
Vector Machines) Once the two models were trained we
will apply each model to the data sets generated by the
four strategies (introduced in the beginning of Section 5)
For each testing six accuracy measurements are used
True Positive Rate (TP Rate) False Positive Rate (FPR)
Precision Recall F-Measurement and Area Under an
ROC Curve (ROC Area)
We will first introduce our feature extraction
techniques for generating semantic features from sliding
windows In our previous work [29] we mainly employ
raw time-series data In the experiments we extend our
previous work by employing six types of features which
are listed as follows The first and second type of feature
is Average and Standard Deviation respectively The
third type of feature is called Average Resultant
Acceleration [30] [31] which is defined as the average
of the square roots of the summation of the sensor
Journal of Communications Vol 10 No 11 November 2015
855copy2015 Journal of Communications
Journal of Communications Vol 10 No 11 November 2015
856copy2015 Journal of Communications
readings within each sliding window The fourth type of
feature is named as Average Absolute Difference [30]
[31] which is defined as the average absolute difference
between the value of each of the raw sensor reading in the
sliding window and the mean value over of all the sensor
reading The fifth category of features is called Time
Between Peaks [30] which is defined as time differences
(calculated in milliseconds) between peaks in the
sinusoidal waveform from the raw sensor readings within
the sliding window
1) Comparison with various models
Table II shows the detailed results using the NBC
model As a representative scenario the results for a loss
rate of 20 are studied As we can see from this table the
proposed enhancement in this paper does contribute to the
improvement of the accuracy For example if we compare
the ldquoF-Measurementrdquo (row six of Table II) values without
lost critical samples retransmitted and with lost critical
samples retransmitted (+cs) we will find that the values
have been improved more than 5
TABLE II ACCURACY COMPARISON AMONG DIFFERENT STRATEGIES
USING NBC MODEL
Strategy Name
Existing Existing+cs LASP-cs LASP
TP Rate 0791 0803 0824 0837
FP Rate 0115 0112 0090 0089
Precision 0809 0814 0831 0847
Recall 0793 0904 0815 0817
F-Measure 0785 0801 0821 0824
ROC Area 0924 0927 0946 0949
TABLE III ACCURACY COMPARISON AMONG DIFFERENT STRATEGIES
USING SVM MODEL
Strategy Name
Existing Existing+cs LASP-cs LASP
TP Rate 0633 0645 0693 0699
FP Rate 0174 0169 0147 0145
Precision 0836 0839 0849 0851
Recall 0635 0645 0692 0699
F-Measure 0631 0630 0689 0695
ROC Area 0728 0737 0783 0782
We can draw similar conclusions from Table III if we
employ the second classification model (SVM
classification model) Again we only show the results
when the loss rate is 20 and the results under other loss
rate are similar As shown in Table III the average
performance improvements are around 10 We also
performs two sample t tests Results show that the p-value
was less than 5 which means the performance
improvements are statistically significant
2) Accuracy under various loss rates
We then study whether or not the proposed LASP
strategy is robust under different loss rates To do so we
evaluated the classification accuracy under different loss
rate within LASP Specifically we evaluated the accuracy
(precision and recall) under three loss rates 10 20
and 30 using the first classification model (NBC
model) Fig 7 shows the precision and recall of our NBC
classification algorithms As shown in this figure while
the classification performance does suffering when the
loss rate increases the reduction of classification
performance is minor For example when the loss rate is
increased for 20 (from 10 to 30) the precision and
recall are reduced by less than 3 These results are
strong indication that our proposed LASP strategy are
very robust for sample loss
Fig 7 Accuracy of LASP under different loss rate with NBC
classification model
3) Comparison with various sample lengths
Finally we conducted a third experimentin order to
determine whether the proposed approach can reduce the
detection time For example given two groups of time
series T1 and T2 we apply the same classification model
to both of them If the classification accuracy for T1 is
better than the result for T2 and the difference between the
accuracy is statistically significant we can claim that
strategy for T1 is more effective (in terms of data analytics)
than the strategy for T2 In our experiments we choose
three lengths 15 seconds (L1) 30 seconds (L2) and 45
seconds (L3) For each length we generate two data
streams the first data stream is generated by sequential
packetization (ldquoExistingrdquo) and the second data stream is
generated by the proposed strategy (LASP) As shown in
Table IV each column represents the results of data
stream with different length For example the second row
(ldquoExisting-L1rdquo) indicates the data stream whose length is
15 seconds (L1) using the sequential packetization
(ldquoExistingrdquo) strategy Each row represents different
accuracy measurements which are introduced in the
beginning of section VB From Table IV we can see that
given the same length of the data stream the accuracy
measurement for data stream from LASP is substantially
higher than the accuracy measurement for data stream
from ldquoExistingrdquo strategy This verifies that if we employ
the proposed approach we could achieve faster detection
speed In another world LASP has a good potential in
applications where real-time classification is desired
TABLE IV ACCURACY COMPARISON AMONG DIFFERENT LENGTHS
GENERATED BY DIFFERENT STRATEGIES
Strategy Existing LASP
L1 L2 L3 L1 L2 L3
TP Rate 0582 0609 0611 0655 0677 0681
FP Rate 0228 0206 0201 0188 0164 0167
Precision 0789 0811 0815 0804 0828 0827
Recall 0570 06 0611 0645 0679 0688
F-Measure 0558 0581 0587 0645 0673 0677
ROC Area 0679 0699 0701 0729 0755 0759
VI SUMMARY
Existing strategies for transmission of time series body
sensor data assumes sequential sample packetization
Journal of Communications Vol 10 No 11 November 2015
857copy2015 Journal of Communications
which leads to difficulty of data recovery in the high loss
ratio WBANs In this paper we investigated the issue of
sample packetization pattern and proposed a heuristic
approach to minimize the effect of channel loss on success
of data analysis The proposed approach is very efficient
for sensor data processing incurs no extra overhead and
does not require modification of the transport protocol
In the future we will implement the packetization
scheme in real sensor network platform such as TinyOS
[32] and experiment with more biomedical sensor data to
validate its effectiveness
ACKNOWLEDGMENT
The authors wish to thank ICNC 2015 reviewers for
their valuable comments that significantly improved the
quality of the paper This work was partially supported by
the US National Science Foundation under Grant No
1229213
REFERENCES
[1] G Z Yang Body Sensor Networks New York USA Springer
Science+Business Media LLC 2006
[2] C Cordeiro and M Patel ldquoBody area networking standardization
present and future directionsrdquo presented at the Proceedings of the
ICST 2nd International Conference on Body Area Networks
Florence Italy 2007
[3] J Ko C Lu M B Srivastava J A Stankovic A Terzis and M
Welsh ldquoWireless sensor networks for healthcarerdquo Proceedings of
the IEEE vol 98 pp 1947-1960 2010
[4] O Amft H Junker P Lukowicz G Troster and C Schuster
ldquoSensing muscle activities with body-worn sensorsrdquo in Proc
International Workshop on Wearable and Implantable Body
Sensor Networks 2006
[5] B French A Smailagic D Siewiorek V Ambur and D
Tyamagundlu ldquoClassifying wheelchair propulsion patterns with a
wrist mounted accelerometerrdquo presented at the Proceedings of the
ICST 3rd International Conference on Body Area Networks
Tempe Arizona 2008
[6] L Jones N Deo and B Lockyer ldquoWireless physiological sensor
system for ambulatory userdquo in Proc International Workshop on
Wearable and Implantable Body Sensor Networks 2006 pp 4
pp-149
[7] K V Laerhoven H W Gellersen and Y G Malliaris ldquoLong
term activity monitoring with a wearable sensor noderdquo in Proc
International Workshop on Wearable and Implantable Body
Sensor Networks 2006
[8] U Maurer A Smailagic D P Siewiorek and M Deisher
ldquoActivity recognition and monitoring using multiple sensors on
different body positionsrdquo in Proc International Workshop on
Wearable and Implantable Body Sensor Networks 2006
[9] M Quwaider and S Biswas ldquoBody posture identification using
hidden Markov model with a wearable sensor networkrdquo presented
at the Proceedings of the ICST 3rd International Conference on
Body Area Networks Tempe Arizona 2008
[10] R Chavarriaga H Sagha A Calatroni S T Digumarti G Tr et
al ldquoThe opportunity challenge A benchmark database for on-
body sensor-based activity recognitionrdquo Pattern Recogn Lett vol
34 pp 2033-2042 2013
[11] M Zhang and A A Sawchuk ldquoUSC-HAD A daily activity
dataset for ubiquitous activity recognition using wearable sensorsrdquo
presented at the Proceedings of the 2012 ACM Conference on
Ubiquitous Computing Pittsburgh Pennsylvania 2012
[12] A Bulling U Blanke and B Schiele ldquoA tutorial on human
activity recognition using body-worn inertial sensorsrdquo ACM
Comput Surv vol 46 pp 1-33 2014
[13] G Zhou J Lu C Wan M Yarvis and J Stankovic ldquoBodyqos
Adaptive and radio-agnostic qos for body sensor networksrdquo in
Proc IEEE INFOCOM Phoenix AZ 2008 pp 565-573
[14] G Zhou C Y Wan M D Yarvis and J A Stankovic
ldquoAggregator-centric QoS for body sensor networksrdquo presented at
the Proceedings of the 6th International Conference on
Information Processing in Sensor Networks Cambridge
Massachusetts USA 2007
[15] B Otal L Alonso and C Verikoukis ldquoNovel QoS scheduling
and energy-saving MAC protocol for body sensor networks
optimizationrdquo presented at the Proceedings of the ICST 3rd
International Conference on Body Area Networks Tempe
Arizona 2008
[16] J J Garcia and T Falck ldquoQuality of service for IEEE 802154-
based wireless body sensor networksrdquo in Proc 3rd International
Conference on Pervasive Computing Technologies for Healthcare
2009 pp 1-6
[17] F Gengfa and E Dutkiewicz ldquoBodyMAC Energy efficient
TDMA-based MAC protocol for wireless body area networksrdquo in
Proc 9th International Symposium on Communications and
Information Technology 2009 pp 1455-1459
[18] N Read M Li Y Cao S H Liu T Wilson and B Prabhakaran
ldquoLoss resilient strategy in body sensor networksrdquo presented at the
Proc of ACMIEEE International Conference on Body Area
Networks (BodyNets) Beijing China 2011
[19] E Weldon Jr ldquoAn improved Selective-Repeat ARQ strategyrdquo
IEEE Transactions on Communications vol 30 pp 480-486
1982
[20] T Fulford-Jones D Malan M Welsh and S Moulton
ldquoCodeBlue An ad hoc sensor network infrastructure for
emergency medical carerdquo in Proc International Workshop on
Body Sensor Networks 2004
[21] S Jiang Y Cao S Iyengar P Kuryloski R Jafari Y Xue et al
ldquoCareNet An integrated wireless sensor networking environment
for remote healthcarerdquo presented at the Proceedings of the ICST
3rd International Conference on Body Area Networks Tempe
Arizona 2008
[22] M Younis K Akkaya M Eltoweissy and A Wadaa ldquoOn
handling QoS traffic in wireless sensor networksrdquo in Proc 37th
Annual Hawaii International Conference on System Sciences Big
Island HI USA 2004
[23] P Baronti P Pillai V W C Chook S Chessa A Gotta and Y
F Hu ldquoWireless sensor networks A survey on the state of the art
and the 80215 4 and ZigBee standardsrdquo Computer
Communications vol 30 pp 1655-1695 2007
[24] Bluetooth Low Energy Core Specification Version 40 [Online]
AvailablehttpwwwbluetoothcomEnglishTechnologyWorks
PagesBluetooth_low_energy_technologyaspx
[25] K K Sup M A Ameen K Daehan L Cheolhyo and L
Hyungsoo ldquoA study on proposed IEEE 80215 WBAN MAC
protocolsrdquo in Proc 9th International Symposium on
Communications and Information Technology 2009 pp 834-840
[26] M van der Schaar and D S Turaga ldquoCross-Layer packetization
and retransmission strategies for delay-sensitive wireless
multimedia transmissionrdquo IEEE Transactions on Multimedia vol
9 pp 185-197 2007
[27] S Zhang S C Liew and P P Lam ldquoHot topic Physical-layer
network codingrdquo presented at the Proceedings of the 12th annual
International Conference on Mobile Computing and Networking
Los Angeles CA USA 2006
[28] J Ellul B Lo and G Z Yang ldquoThe BSNOS platform A body
sensor networks targeted operating system and toolsetrdquo in Proc
Journal of Communications Vol 10 No 11 November 2015
858copy2015 Journal of Communications
Fifth International Conference on Sensor Technologies and
Applications 2011 pp 381-386
[29] M Li Y Cao and B Prabhakaran ldquoLoss aware sample
packetization strategy for improvement of body sensor data
analysisrdquo in Proc International Conference on Computing
Networking and Communications Anaheim California 2015
[30] L Bao and S S Intille ldquoActivity recognition from user-annotated
acceleration datardquo in Pervasive Computing Springer 2004 pp 1-
17
[31] J R Kwapisz G M Weiss and S A Moore ldquoActivity
recognition using cell phone accelerometersrdquo ACM SigKDD
Explorations Newsletter vol 12 pp 74-82 2011
[32] TinyOS [Online] Available wwwtinyosnet
Ming Li is currently an Associate Professor
and Chair in the Department of Computer
Science California State University Fresno
Prior to that he was an Assistant Professor
from August 2006 to 2012 He received his
MS and PhD degrees in Computer Science
from The University of Texas at Dallas in
2001 and 2006 respectively His research
interests include QoS strategies for IEEE
80211 wireless LANs mobile ad-hoc networks and heterogeneous
wired and wireless networks multimedia streaming over wireless
networks body area networks and robot swarm communications
Yu Cao has been an Assistant Professor at the
Department of Computer Science The
University of Massachusetts Lowell since Aug
2013 From 2010 to 2013 he was a faculty at
The University of Tennessee at Chattanooga
(UTC) From 2007 to 2010 he was a faculty at
California State University Fresno Prior to
that he was a Visiting Fellow of Biomedical
Engineering at Mayo Clinic Rochester
Minnesota He received his MS and PhD degrees in Computer
Science from Iowa State University in 2005 and 2007 respectively He
received the BEng degree from Harbin Engineering University in 1997
the MEng degree from Huazhong University of Science and
Technology in 2000 all in Computer Science His research interests
span a variety of aspects of knowledge discover from complex data
which include the area of biomedical informatics and intelligent system
B Prabhakaran is a Professor of Computer
Science in the University of Texas at Dallas
Prof B Prabhakaran works in the area of
multimedia systems He has published several
research papers in prestigious conferences and
journals in this area Dr Prabhakaran received
the prestigious NSF CAREER Award FY
2003 for his proposal on Animation Databases
Dr Prabhakaran is General Co-Chair of ACM
Multimedia 2011 He is also Technical Program Co-Chair of IEEE
WoWMoM 2012 (World of Wireless Mobile and Multimedia
Networks) He served as the TPC Co-Chair of IEEE ISM 2010
(International Symposium on Multimedia) Dr Prabhakaran is a
Member of the Executive Council of the ACM Special Interest Group
on Multimedia (SIGMM) and is the Co-Chair of IEEE Technical
Committe on Multimedia Computing (TCMC) Special Interest Group
on Video Analytics (SIGVA) Dr Prabhakaran is the Editor-in-Chief of
the ACM SIGMM (Special Interest Group on Multimedia) web
magazine He is Member of the Editorial board of Multimedia Systems
Journal (Springer) and Multimedia Tools and Applications journal
(Springer) He has served as guest-editor (special issue on Multimedia
Authoring and Presentation) for ACM Multimedia Systems journal
Page 6
Journal of Communications Vol 10 No 11 November 2015
856copy2015 Journal of Communications
readings within each sliding window The fourth type of
feature is named as Average Absolute Difference [30]
[31] which is defined as the average absolute difference
between the value of each of the raw sensor reading in the
sliding window and the mean value over of all the sensor
reading The fifth category of features is called Time
Between Peaks [30] which is defined as time differences
(calculated in milliseconds) between peaks in the
sinusoidal waveform from the raw sensor readings within
the sliding window
1) Comparison with various models
Table II shows the detailed results using the NBC
model As a representative scenario the results for a loss
rate of 20 are studied As we can see from this table the
proposed enhancement in this paper does contribute to the
improvement of the accuracy For example if we compare
the ldquoF-Measurementrdquo (row six of Table II) values without
lost critical samples retransmitted and with lost critical
samples retransmitted (+cs) we will find that the values
have been improved more than 5
TABLE II ACCURACY COMPARISON AMONG DIFFERENT STRATEGIES
USING NBC MODEL
Strategy Name
Existing Existing+cs LASP-cs LASP
TP Rate 0791 0803 0824 0837
FP Rate 0115 0112 0090 0089
Precision 0809 0814 0831 0847
Recall 0793 0904 0815 0817
F-Measure 0785 0801 0821 0824
ROC Area 0924 0927 0946 0949
TABLE III ACCURACY COMPARISON AMONG DIFFERENT STRATEGIES
USING SVM MODEL
Strategy Name
Existing Existing+cs LASP-cs LASP
TP Rate 0633 0645 0693 0699
FP Rate 0174 0169 0147 0145
Precision 0836 0839 0849 0851
Recall 0635 0645 0692 0699
F-Measure 0631 0630 0689 0695
ROC Area 0728 0737 0783 0782
We can draw similar conclusions from Table III if we
employ the second classification model (SVM
classification model) Again we only show the results
when the loss rate is 20 and the results under other loss
rate are similar As shown in Table III the average
performance improvements are around 10 We also
performs two sample t tests Results show that the p-value
was less than 5 which means the performance
improvements are statistically significant
2) Accuracy under various loss rates
We then study whether or not the proposed LASP
strategy is robust under different loss rates To do so we
evaluated the classification accuracy under different loss
rate within LASP Specifically we evaluated the accuracy
(precision and recall) under three loss rates 10 20
and 30 using the first classification model (NBC
model) Fig 7 shows the precision and recall of our NBC
classification algorithms As shown in this figure while
the classification performance does suffering when the
loss rate increases the reduction of classification
performance is minor For example when the loss rate is
increased for 20 (from 10 to 30) the precision and
recall are reduced by less than 3 These results are
strong indication that our proposed LASP strategy are
very robust for sample loss
Fig 7 Accuracy of LASP under different loss rate with NBC
classification model
3) Comparison with various sample lengths
Finally we conducted a third experimentin order to
determine whether the proposed approach can reduce the
detection time For example given two groups of time
series T1 and T2 we apply the same classification model
to both of them If the classification accuracy for T1 is
better than the result for T2 and the difference between the
accuracy is statistically significant we can claim that
strategy for T1 is more effective (in terms of data analytics)
than the strategy for T2 In our experiments we choose
three lengths 15 seconds (L1) 30 seconds (L2) and 45
seconds (L3) For each length we generate two data
streams the first data stream is generated by sequential
packetization (ldquoExistingrdquo) and the second data stream is
generated by the proposed strategy (LASP) As shown in
Table IV each column represents the results of data
stream with different length For example the second row
(ldquoExisting-L1rdquo) indicates the data stream whose length is
15 seconds (L1) using the sequential packetization
(ldquoExistingrdquo) strategy Each row represents different
accuracy measurements which are introduced in the
beginning of section VB From Table IV we can see that
given the same length of the data stream the accuracy
measurement for data stream from LASP is substantially
higher than the accuracy measurement for data stream
from ldquoExistingrdquo strategy This verifies that if we employ
the proposed approach we could achieve faster detection
speed In another world LASP has a good potential in
applications where real-time classification is desired
TABLE IV ACCURACY COMPARISON AMONG DIFFERENT LENGTHS
GENERATED BY DIFFERENT STRATEGIES
Strategy Existing LASP
L1 L2 L3 L1 L2 L3
TP Rate 0582 0609 0611 0655 0677 0681
FP Rate 0228 0206 0201 0188 0164 0167
Precision 0789 0811 0815 0804 0828 0827
Recall 0570 06 0611 0645 0679 0688
F-Measure 0558 0581 0587 0645 0673 0677
ROC Area 0679 0699 0701 0729 0755 0759
VI SUMMARY
Existing strategies for transmission of time series body
sensor data assumes sequential sample packetization
Journal of Communications Vol 10 No 11 November 2015
857copy2015 Journal of Communications
which leads to difficulty of data recovery in the high loss
ratio WBANs In this paper we investigated the issue of
sample packetization pattern and proposed a heuristic
approach to minimize the effect of channel loss on success
of data analysis The proposed approach is very efficient
for sensor data processing incurs no extra overhead and
does not require modification of the transport protocol
In the future we will implement the packetization
scheme in real sensor network platform such as TinyOS
[32] and experiment with more biomedical sensor data to
validate its effectiveness
ACKNOWLEDGMENT
The authors wish to thank ICNC 2015 reviewers for
their valuable comments that significantly improved the
quality of the paper This work was partially supported by
the US National Science Foundation under Grant No
1229213
REFERENCES
[1] G Z Yang Body Sensor Networks New York USA Springer
Science+Business Media LLC 2006
[2] C Cordeiro and M Patel ldquoBody area networking standardization
present and future directionsrdquo presented at the Proceedings of the
ICST 2nd International Conference on Body Area Networks
Florence Italy 2007
[3] J Ko C Lu M B Srivastava J A Stankovic A Terzis and M
Welsh ldquoWireless sensor networks for healthcarerdquo Proceedings of
the IEEE vol 98 pp 1947-1960 2010
[4] O Amft H Junker P Lukowicz G Troster and C Schuster
ldquoSensing muscle activities with body-worn sensorsrdquo in Proc
International Workshop on Wearable and Implantable Body
Sensor Networks 2006
[5] B French A Smailagic D Siewiorek V Ambur and D
Tyamagundlu ldquoClassifying wheelchair propulsion patterns with a
wrist mounted accelerometerrdquo presented at the Proceedings of the
ICST 3rd International Conference on Body Area Networks
Tempe Arizona 2008
[6] L Jones N Deo and B Lockyer ldquoWireless physiological sensor
system for ambulatory userdquo in Proc International Workshop on
Wearable and Implantable Body Sensor Networks 2006 pp 4
pp-149
[7] K V Laerhoven H W Gellersen and Y G Malliaris ldquoLong
term activity monitoring with a wearable sensor noderdquo in Proc
International Workshop on Wearable and Implantable Body
Sensor Networks 2006
[8] U Maurer A Smailagic D P Siewiorek and M Deisher
ldquoActivity recognition and monitoring using multiple sensors on
different body positionsrdquo in Proc International Workshop on
Wearable and Implantable Body Sensor Networks 2006
[9] M Quwaider and S Biswas ldquoBody posture identification using
hidden Markov model with a wearable sensor networkrdquo presented
at the Proceedings of the ICST 3rd International Conference on
Body Area Networks Tempe Arizona 2008
[10] R Chavarriaga H Sagha A Calatroni S T Digumarti G Tr et
al ldquoThe opportunity challenge A benchmark database for on-
body sensor-based activity recognitionrdquo Pattern Recogn Lett vol
34 pp 2033-2042 2013
[11] M Zhang and A A Sawchuk ldquoUSC-HAD A daily activity
dataset for ubiquitous activity recognition using wearable sensorsrdquo
presented at the Proceedings of the 2012 ACM Conference on
Ubiquitous Computing Pittsburgh Pennsylvania 2012
[12] A Bulling U Blanke and B Schiele ldquoA tutorial on human
activity recognition using body-worn inertial sensorsrdquo ACM
Comput Surv vol 46 pp 1-33 2014
[13] G Zhou J Lu C Wan M Yarvis and J Stankovic ldquoBodyqos
Adaptive and radio-agnostic qos for body sensor networksrdquo in
Proc IEEE INFOCOM Phoenix AZ 2008 pp 565-573
[14] G Zhou C Y Wan M D Yarvis and J A Stankovic
ldquoAggregator-centric QoS for body sensor networksrdquo presented at
the Proceedings of the 6th International Conference on
Information Processing in Sensor Networks Cambridge
Massachusetts USA 2007
[15] B Otal L Alonso and C Verikoukis ldquoNovel QoS scheduling
and energy-saving MAC protocol for body sensor networks
optimizationrdquo presented at the Proceedings of the ICST 3rd
International Conference on Body Area Networks Tempe
Arizona 2008
[16] J J Garcia and T Falck ldquoQuality of service for IEEE 802154-
based wireless body sensor networksrdquo in Proc 3rd International
Conference on Pervasive Computing Technologies for Healthcare
2009 pp 1-6
[17] F Gengfa and E Dutkiewicz ldquoBodyMAC Energy efficient
TDMA-based MAC protocol for wireless body area networksrdquo in
Proc 9th International Symposium on Communications and
Information Technology 2009 pp 1455-1459
[18] N Read M Li Y Cao S H Liu T Wilson and B Prabhakaran
ldquoLoss resilient strategy in body sensor networksrdquo presented at the
Proc of ACMIEEE International Conference on Body Area
Networks (BodyNets) Beijing China 2011
[19] E Weldon Jr ldquoAn improved Selective-Repeat ARQ strategyrdquo
IEEE Transactions on Communications vol 30 pp 480-486
1982
[20] T Fulford-Jones D Malan M Welsh and S Moulton
ldquoCodeBlue An ad hoc sensor network infrastructure for
emergency medical carerdquo in Proc International Workshop on
Body Sensor Networks 2004
[21] S Jiang Y Cao S Iyengar P Kuryloski R Jafari Y Xue et al
ldquoCareNet An integrated wireless sensor networking environment
for remote healthcarerdquo presented at the Proceedings of the ICST
3rd International Conference on Body Area Networks Tempe
Arizona 2008
[22] M Younis K Akkaya M Eltoweissy and A Wadaa ldquoOn
handling QoS traffic in wireless sensor networksrdquo in Proc 37th
Annual Hawaii International Conference on System Sciences Big
Island HI USA 2004
[23] P Baronti P Pillai V W C Chook S Chessa A Gotta and Y
F Hu ldquoWireless sensor networks A survey on the state of the art
and the 80215 4 and ZigBee standardsrdquo Computer
Communications vol 30 pp 1655-1695 2007
[24] Bluetooth Low Energy Core Specification Version 40 [Online]
AvailablehttpwwwbluetoothcomEnglishTechnologyWorks
PagesBluetooth_low_energy_technologyaspx
[25] K K Sup M A Ameen K Daehan L Cheolhyo and L
Hyungsoo ldquoA study on proposed IEEE 80215 WBAN MAC
protocolsrdquo in Proc 9th International Symposium on
Communications and Information Technology 2009 pp 834-840
[26] M van der Schaar and D S Turaga ldquoCross-Layer packetization
and retransmission strategies for delay-sensitive wireless
multimedia transmissionrdquo IEEE Transactions on Multimedia vol
9 pp 185-197 2007
[27] S Zhang S C Liew and P P Lam ldquoHot topic Physical-layer
network codingrdquo presented at the Proceedings of the 12th annual
International Conference on Mobile Computing and Networking
Los Angeles CA USA 2006
[28] J Ellul B Lo and G Z Yang ldquoThe BSNOS platform A body
sensor networks targeted operating system and toolsetrdquo in Proc
Journal of Communications Vol 10 No 11 November 2015
858copy2015 Journal of Communications
Fifth International Conference on Sensor Technologies and
Applications 2011 pp 381-386
[29] M Li Y Cao and B Prabhakaran ldquoLoss aware sample
packetization strategy for improvement of body sensor data
analysisrdquo in Proc International Conference on Computing
Networking and Communications Anaheim California 2015
[30] L Bao and S S Intille ldquoActivity recognition from user-annotated
acceleration datardquo in Pervasive Computing Springer 2004 pp 1-
17
[31] J R Kwapisz G M Weiss and S A Moore ldquoActivity
recognition using cell phone accelerometersrdquo ACM SigKDD
Explorations Newsletter vol 12 pp 74-82 2011
[32] TinyOS [Online] Available wwwtinyosnet
Ming Li is currently an Associate Professor
and Chair in the Department of Computer
Science California State University Fresno
Prior to that he was an Assistant Professor
from August 2006 to 2012 He received his
MS and PhD degrees in Computer Science
from The University of Texas at Dallas in
2001 and 2006 respectively His research
interests include QoS strategies for IEEE
80211 wireless LANs mobile ad-hoc networks and heterogeneous
wired and wireless networks multimedia streaming over wireless
networks body area networks and robot swarm communications
Yu Cao has been an Assistant Professor at the
Department of Computer Science The
University of Massachusetts Lowell since Aug
2013 From 2010 to 2013 he was a faculty at
The University of Tennessee at Chattanooga
(UTC) From 2007 to 2010 he was a faculty at
California State University Fresno Prior to
that he was a Visiting Fellow of Biomedical
Engineering at Mayo Clinic Rochester
Minnesota He received his MS and PhD degrees in Computer
Science from Iowa State University in 2005 and 2007 respectively He
received the BEng degree from Harbin Engineering University in 1997
the MEng degree from Huazhong University of Science and
Technology in 2000 all in Computer Science His research interests
span a variety of aspects of knowledge discover from complex data
which include the area of biomedical informatics and intelligent system
B Prabhakaran is a Professor of Computer
Science in the University of Texas at Dallas
Prof B Prabhakaran works in the area of
multimedia systems He has published several
research papers in prestigious conferences and
journals in this area Dr Prabhakaran received
the prestigious NSF CAREER Award FY
2003 for his proposal on Animation Databases
Dr Prabhakaran is General Co-Chair of ACM
Multimedia 2011 He is also Technical Program Co-Chair of IEEE
WoWMoM 2012 (World of Wireless Mobile and Multimedia
Networks) He served as the TPC Co-Chair of IEEE ISM 2010
(International Symposium on Multimedia) Dr Prabhakaran is a
Member of the Executive Council of the ACM Special Interest Group
on Multimedia (SIGMM) and is the Co-Chair of IEEE Technical
Committe on Multimedia Computing (TCMC) Special Interest Group
on Video Analytics (SIGVA) Dr Prabhakaran is the Editor-in-Chief of
the ACM SIGMM (Special Interest Group on Multimedia) web
magazine He is Member of the Editorial board of Multimedia Systems
Journal (Springer) and Multimedia Tools and Applications journal
(Springer) He has served as guest-editor (special issue on Multimedia
Authoring and Presentation) for ACM Multimedia Systems journal
Page 7
Journal of Communications Vol 10 No 11 November 2015
857copy2015 Journal of Communications
which leads to difficulty of data recovery in the high loss
ratio WBANs In this paper we investigated the issue of
sample packetization pattern and proposed a heuristic
approach to minimize the effect of channel loss on success
of data analysis The proposed approach is very efficient
for sensor data processing incurs no extra overhead and
does not require modification of the transport protocol
In the future we will implement the packetization
scheme in real sensor network platform such as TinyOS
[32] and experiment with more biomedical sensor data to
validate its effectiveness
ACKNOWLEDGMENT
The authors wish to thank ICNC 2015 reviewers for
their valuable comments that significantly improved the
quality of the paper This work was partially supported by
the US National Science Foundation under Grant No
1229213
REFERENCES
[1] G Z Yang Body Sensor Networks New York USA Springer
Science+Business Media LLC 2006
[2] C Cordeiro and M Patel ldquoBody area networking standardization
present and future directionsrdquo presented at the Proceedings of the
ICST 2nd International Conference on Body Area Networks
Florence Italy 2007
[3] J Ko C Lu M B Srivastava J A Stankovic A Terzis and M
Welsh ldquoWireless sensor networks for healthcarerdquo Proceedings of
the IEEE vol 98 pp 1947-1960 2010
[4] O Amft H Junker P Lukowicz G Troster and C Schuster
ldquoSensing muscle activities with body-worn sensorsrdquo in Proc
International Workshop on Wearable and Implantable Body
Sensor Networks 2006
[5] B French A Smailagic D Siewiorek V Ambur and D
Tyamagundlu ldquoClassifying wheelchair propulsion patterns with a
wrist mounted accelerometerrdquo presented at the Proceedings of the
ICST 3rd International Conference on Body Area Networks
Tempe Arizona 2008
[6] L Jones N Deo and B Lockyer ldquoWireless physiological sensor
system for ambulatory userdquo in Proc International Workshop on
Wearable and Implantable Body Sensor Networks 2006 pp 4
pp-149
[7] K V Laerhoven H W Gellersen and Y G Malliaris ldquoLong
term activity monitoring with a wearable sensor noderdquo in Proc
International Workshop on Wearable and Implantable Body
Sensor Networks 2006
[8] U Maurer A Smailagic D P Siewiorek and M Deisher
ldquoActivity recognition and monitoring using multiple sensors on
different body positionsrdquo in Proc International Workshop on
Wearable and Implantable Body Sensor Networks 2006
[9] M Quwaider and S Biswas ldquoBody posture identification using
hidden Markov model with a wearable sensor networkrdquo presented
at the Proceedings of the ICST 3rd International Conference on
Body Area Networks Tempe Arizona 2008
[10] R Chavarriaga H Sagha A Calatroni S T Digumarti G Tr et
al ldquoThe opportunity challenge A benchmark database for on-
body sensor-based activity recognitionrdquo Pattern Recogn Lett vol
34 pp 2033-2042 2013
[11] M Zhang and A A Sawchuk ldquoUSC-HAD A daily activity
dataset for ubiquitous activity recognition using wearable sensorsrdquo
presented at the Proceedings of the 2012 ACM Conference on
Ubiquitous Computing Pittsburgh Pennsylvania 2012
[12] A Bulling U Blanke and B Schiele ldquoA tutorial on human
activity recognition using body-worn inertial sensorsrdquo ACM
Comput Surv vol 46 pp 1-33 2014
[13] G Zhou J Lu C Wan M Yarvis and J Stankovic ldquoBodyqos
Adaptive and radio-agnostic qos for body sensor networksrdquo in
Proc IEEE INFOCOM Phoenix AZ 2008 pp 565-573
[14] G Zhou C Y Wan M D Yarvis and J A Stankovic
ldquoAggregator-centric QoS for body sensor networksrdquo presented at
the Proceedings of the 6th International Conference on
Information Processing in Sensor Networks Cambridge
Massachusetts USA 2007
[15] B Otal L Alonso and C Verikoukis ldquoNovel QoS scheduling
and energy-saving MAC protocol for body sensor networks
optimizationrdquo presented at the Proceedings of the ICST 3rd
International Conference on Body Area Networks Tempe
Arizona 2008
[16] J J Garcia and T Falck ldquoQuality of service for IEEE 802154-
based wireless body sensor networksrdquo in Proc 3rd International
Conference on Pervasive Computing Technologies for Healthcare
2009 pp 1-6
[17] F Gengfa and E Dutkiewicz ldquoBodyMAC Energy efficient
TDMA-based MAC protocol for wireless body area networksrdquo in
Proc 9th International Symposium on Communications and
Information Technology 2009 pp 1455-1459
[18] N Read M Li Y Cao S H Liu T Wilson and B Prabhakaran
ldquoLoss resilient strategy in body sensor networksrdquo presented at the
Proc of ACMIEEE International Conference on Body Area
Networks (BodyNets) Beijing China 2011
[19] E Weldon Jr ldquoAn improved Selective-Repeat ARQ strategyrdquo
IEEE Transactions on Communications vol 30 pp 480-486
1982
[20] T Fulford-Jones D Malan M Welsh and S Moulton
ldquoCodeBlue An ad hoc sensor network infrastructure for
emergency medical carerdquo in Proc International Workshop on
Body Sensor Networks 2004
[21] S Jiang Y Cao S Iyengar P Kuryloski R Jafari Y Xue et al
ldquoCareNet An integrated wireless sensor networking environment
for remote healthcarerdquo presented at the Proceedings of the ICST
3rd International Conference on Body Area Networks Tempe
Arizona 2008
[22] M Younis K Akkaya M Eltoweissy and A Wadaa ldquoOn
handling QoS traffic in wireless sensor networksrdquo in Proc 37th
Annual Hawaii International Conference on System Sciences Big
Island HI USA 2004
[23] P Baronti P Pillai V W C Chook S Chessa A Gotta and Y
F Hu ldquoWireless sensor networks A survey on the state of the art
and the 80215 4 and ZigBee standardsrdquo Computer
Communications vol 30 pp 1655-1695 2007
[24] Bluetooth Low Energy Core Specification Version 40 [Online]
AvailablehttpwwwbluetoothcomEnglishTechnologyWorks
PagesBluetooth_low_energy_technologyaspx
[25] K K Sup M A Ameen K Daehan L Cheolhyo and L
Hyungsoo ldquoA study on proposed IEEE 80215 WBAN MAC
protocolsrdquo in Proc 9th International Symposium on
Communications and Information Technology 2009 pp 834-840
[26] M van der Schaar and D S Turaga ldquoCross-Layer packetization
and retransmission strategies for delay-sensitive wireless
multimedia transmissionrdquo IEEE Transactions on Multimedia vol
9 pp 185-197 2007
[27] S Zhang S C Liew and P P Lam ldquoHot topic Physical-layer
network codingrdquo presented at the Proceedings of the 12th annual
International Conference on Mobile Computing and Networking
Los Angeles CA USA 2006
[28] J Ellul B Lo and G Z Yang ldquoThe BSNOS platform A body
sensor networks targeted operating system and toolsetrdquo in Proc
Journal of Communications Vol 10 No 11 November 2015
858copy2015 Journal of Communications
Fifth International Conference on Sensor Technologies and
Applications 2011 pp 381-386
[29] M Li Y Cao and B Prabhakaran ldquoLoss aware sample
packetization strategy for improvement of body sensor data
analysisrdquo in Proc International Conference on Computing
Networking and Communications Anaheim California 2015
[30] L Bao and S S Intille ldquoActivity recognition from user-annotated
acceleration datardquo in Pervasive Computing Springer 2004 pp 1-
17
[31] J R Kwapisz G M Weiss and S A Moore ldquoActivity
recognition using cell phone accelerometersrdquo ACM SigKDD
Explorations Newsletter vol 12 pp 74-82 2011
[32] TinyOS [Online] Available wwwtinyosnet
Ming Li is currently an Associate Professor
and Chair in the Department of Computer
Science California State University Fresno
Prior to that he was an Assistant Professor
from August 2006 to 2012 He received his
MS and PhD degrees in Computer Science
from The University of Texas at Dallas in
2001 and 2006 respectively His research
interests include QoS strategies for IEEE
80211 wireless LANs mobile ad-hoc networks and heterogeneous
wired and wireless networks multimedia streaming over wireless
networks body area networks and robot swarm communications
Yu Cao has been an Assistant Professor at the
Department of Computer Science The
University of Massachusetts Lowell since Aug
2013 From 2010 to 2013 he was a faculty at
The University of Tennessee at Chattanooga
(UTC) From 2007 to 2010 he was a faculty at
California State University Fresno Prior to
that he was a Visiting Fellow of Biomedical
Engineering at Mayo Clinic Rochester
Minnesota He received his MS and PhD degrees in Computer
Science from Iowa State University in 2005 and 2007 respectively He
received the BEng degree from Harbin Engineering University in 1997
the MEng degree from Huazhong University of Science and
Technology in 2000 all in Computer Science His research interests
span a variety of aspects of knowledge discover from complex data
which include the area of biomedical informatics and intelligent system
B Prabhakaran is a Professor of Computer
Science in the University of Texas at Dallas
Prof B Prabhakaran works in the area of
multimedia systems He has published several
research papers in prestigious conferences and
journals in this area Dr Prabhakaran received
the prestigious NSF CAREER Award FY
2003 for his proposal on Animation Databases
Dr Prabhakaran is General Co-Chair of ACM
Multimedia 2011 He is also Technical Program Co-Chair of IEEE
WoWMoM 2012 (World of Wireless Mobile and Multimedia
Networks) He served as the TPC Co-Chair of IEEE ISM 2010
(International Symposium on Multimedia) Dr Prabhakaran is a
Member of the Executive Council of the ACM Special Interest Group
on Multimedia (SIGMM) and is the Co-Chair of IEEE Technical
Committe on Multimedia Computing (TCMC) Special Interest Group
on Video Analytics (SIGVA) Dr Prabhakaran is the Editor-in-Chief of
the ACM SIGMM (Special Interest Group on Multimedia) web
magazine He is Member of the Editorial board of Multimedia Systems
Journal (Springer) and Multimedia Tools and Applications journal
(Springer) He has served as guest-editor (special issue on Multimedia
Authoring and Presentation) for ACM Multimedia Systems journal
Page 8
Journal of Communications Vol 10 No 11 November 2015
858copy2015 Journal of Communications
Fifth International Conference on Sensor Technologies and
Applications 2011 pp 381-386
[29] M Li Y Cao and B Prabhakaran ldquoLoss aware sample
packetization strategy for improvement of body sensor data
analysisrdquo in Proc International Conference on Computing
Networking and Communications Anaheim California 2015
[30] L Bao and S S Intille ldquoActivity recognition from user-annotated
acceleration datardquo in Pervasive Computing Springer 2004 pp 1-
17
[31] J R Kwapisz G M Weiss and S A Moore ldquoActivity
recognition using cell phone accelerometersrdquo ACM SigKDD
Explorations Newsletter vol 12 pp 74-82 2011
[32] TinyOS [Online] Available wwwtinyosnet
Ming Li is currently an Associate Professor
and Chair in the Department of Computer
Science California State University Fresno
Prior to that he was an Assistant Professor
from August 2006 to 2012 He received his
MS and PhD degrees in Computer Science
from The University of Texas at Dallas in
2001 and 2006 respectively His research
interests include QoS strategies for IEEE
80211 wireless LANs mobile ad-hoc networks and heterogeneous
wired and wireless networks multimedia streaming over wireless
networks body area networks and robot swarm communications
Yu Cao has been an Assistant Professor at the
Department of Computer Science The
University of Massachusetts Lowell since Aug
2013 From 2010 to 2013 he was a faculty at
The University of Tennessee at Chattanooga
(UTC) From 2007 to 2010 he was a faculty at
California State University Fresno Prior to
that he was a Visiting Fellow of Biomedical
Engineering at Mayo Clinic Rochester
Minnesota He received his MS and PhD degrees in Computer
Science from Iowa State University in 2005 and 2007 respectively He
received the BEng degree from Harbin Engineering University in 1997
the MEng degree from Huazhong University of Science and
Technology in 2000 all in Computer Science His research interests
span a variety of aspects of knowledge discover from complex data
which include the area of biomedical informatics and intelligent system
B Prabhakaran is a Professor of Computer
Science in the University of Texas at Dallas
Prof B Prabhakaran works in the area of
multimedia systems He has published several
research papers in prestigious conferences and
journals in this area Dr Prabhakaran received
the prestigious NSF CAREER Award FY
2003 for his proposal on Animation Databases
Dr Prabhakaran is General Co-Chair of ACM
Multimedia 2011 He is also Technical Program Co-Chair of IEEE
WoWMoM 2012 (World of Wireless Mobile and Multimedia
Networks) He served as the TPC Co-Chair of IEEE ISM 2010
(International Symposium on Multimedia) Dr Prabhakaran is a
Member of the Executive Council of the ACM Special Interest Group
on Multimedia (SIGMM) and is the Co-Chair of IEEE Technical
Committe on Multimedia Computing (TCMC) Special Interest Group
on Video Analytics (SIGVA) Dr Prabhakaran is the Editor-in-Chief of
the ACM SIGMM (Special Interest Group on Multimedia) web
magazine He is Member of the Editorial board of Multimedia Systems
Journal (Springer) and Multimedia Tools and Applications journal
(Springer) He has served as guest-editor (special issue on Multimedia
Authoring and Presentation) for ACM Multimedia Systems journal