An Anomaly Detection Model Using Principal Component Analysis Technique for Medical Wireless Sensor Networks

Full text

An Anomaly Detection Model Using Principal
Component Analysis Technique for Medical
Wireless Sensor Networks
Abstract—A Wireless Medical Sensor Network (WMSN)
connects autonomous nodes such as ( sensors and actuators)
existing on the body, or under a person's skin. The network
usually extends over the entire human body and nodes are
connected via a wireless communication channel. WMSN sense
human physiological signs and monitor a patient's health status.
The framework and preliminary experiment on developing an
anomaly detection model for ubiquitous patient and healthcare
monitoring in medical wireless sensor networks (MWSNs)were
presented in this study. The architecture is a combination of an
improved data mining method and machine learning algorithms
using modern fusion methods. Being that MWSNs are highly
susceptible to failures due to certain limitations, such as low
energy resources, poor reliability, low computational resources,
and considerable susceptibility to post-deployment security
attacks, the proposed model is an anomaly detection method for
MWSNs for the detection of anomalies in an adaptive manner
with high accuracy while maintaining resource constraints
using two phases. First, an anomaly-based detection model was
created using Principal Component Analysis (PCA) technique
to reduce dimensionality, prevent overfitting, and increase
detection accuracy. Second, the detection accuracy of the
proposed model was evaluated and compare before and after
PCA integration. The experimental study showed that the
proposed model can rapidly identify sensor anomalies with high
accuracy.
Keywords— medical wireless sensor network, anomaly
detection, principal component analysis, detection accuracy,
detection rate, data reduction
I. INTRODUCTION
With the continued increase in the average human
lifespan, the number of old people has continued to increase,
causing increases in healthcare-related costs and low patient-
to-doctor ratio due to the ever-increasing demand for medical
care [1]. Caregivers and healthcare providers have devised the
remote monitoring method to cater to the increasing number
of old persons and this has raised the interest in the use of
WSNs in the healthcare sector [2]. Scientists and researchers
have developed MWSNs as a network of wireless sensors
consisting of several miniaturized sensors that can execute
wireless transmission of data from their zones of deployment
(connected or implanted) [3][4]. Hence, physicians rely on
these devices to remotely monitor the vital signs of their
patients even when they are far from the hospital; the sensed
data is communicated to the related professional via a Control
Processing Unit (CPU), such as smartphone, laptop, or tablet
that has more processing power, larger batteries, and a wider
range of transmission than the individual MWSN nodes. So,
the CPU must have the capacity to process the received signals
in real-time, as must be capable of raising medical alarms for
caregivers when patients' health deteriorates, allowing them to
respond immediately by taking relevant actions [5,6]. The
processed data can also be sent to distant databases (DB) by
the CPU for onward long-term analysis and storage. The
advantages of MWSNs include helping healthcare givers in
monitoring patients irrespective of their location, improving
the efficiency and accuracy of diagnosis, and reducing the
overall health-care related costs while allowing constant
monitoring of patients [5-7]. The rate of disease detection can
be improved by using MWSN and it can reduce the risk and
impacts on people's lives during detection of dangerous
diseases. Many medical WSN systems are commercially
available, including Tmote Sky, MICAz, MICA2, IRIS,
Imote2, TelosB, & Shimmer for the monitoring of vital signs
like heart rate (HR), pulse, oxygen saturation (SpO2),
respiration rate (RR), body temperature (BT),
electrocardiogram (ECG), blood pressure (BP),
electromyogram (EMG), blood glucose levels (BGL), etc.
However, these systems are essentially devices for collecting
and reporting crucial data, and may not guarantee data
security; hence, they must be equipped with intrusion
detection systems to guarantee data security. Furthermore,
data reduction techniques that aid in increasing battery life are
not offered in these systems.
While MWSNs have several advantages, they often
have several drawbacks, such as low reliability, limited
energy resources, low computational power, and
considerable susceptibility to security attacks upon their
deployment. Therefore, an adaptive data reduction solution is
required to suit the resource constraints demands of medical
sensor devices in terms of computational complexity while
incurring lower approximation error of original data. To
improve the strengths of these devices and minimize the
chances of any weakness, it is important to first investigate
the weaknesses in further detail to discover some ways of
mitigation. MWSN sensor nodes are vulnerable to data
dynamic changes that can affect their efficiency. In general,
Nabeel Abdulrazaq Yaseen
Faculty of Basic Education,
University of Missan, Iraq
nabilalrashy78@gmail.com
Abbas Abd-Alhussein Hadad
Faculty of Basic Education,
University of Missan, Iraq
abbas@uomisan.edu.iq
Mustafa Sabah Taha
Missan Oil Training Institute,HSE
Center, Ministry of Oil, Iraq
mustafa@moti.oil.gov.iq
2021 International Conference on Data Science and Its Applications (ICoDSA)
978-1-6654-4303-6/21/$31.00 ©2021 IEEE 66
2021 International Conference on Data Science and Its Applications (ICoDSA) | 978-1-6654-4303-6/21/$31.00 ©2021 IEEE | DOI: 10.1109/ICODSA53588.2021.9617547
Authorized licensed use limited to: UNIVERSITY TEKNOLOGI MALAYSIA. Downloaded on February 05,2022 at 13:28:28 UTC from IEEE Xplore. Restrictions apply.

the MWSN has problems such as sensor calibration, faulty
components, dislocation, and battery exhaustion [8,9].
This paper aims to develop a highly efficient dimensionality
reduction method that can be used to find accurate AD model
for MWSN to detect anomalies with high accuracy while
maintaining resource constraints. To prove the research
hypothesis, the following objectives are presented. To prove
the research hypothesis, the following objectives are
presented:
a) To propose and develop an anomaly-based detection
model by employing Principal Component Analysis
(PCA) technique for dimensionality reduction (this
technique is able to reduce the dimensionality of the
medical data and increases detection accuracy).
b) To evaluate the detection accuracy of the suggested
method by implementing the model and compare the
results before and after integrating PCA.
The rest of this article is arranged as follows: Section 2
presents the proposed method, while Section 3 discusses the
results of the experiments. Section 4 presents the conclusion
of the study.
II. RESEARCH METHOD
The proposed study begins with a review of the
literature. The goal of investigating the state-of-the-arts is to
look for approaches that have been implemented in handling
the problem of AD in MWSN in the existing works. It has
been determined that PCA is a highly efficient dimensionality
reduction method that can be used to find accurate AD
solutions. This research involves three phases, each of which
contributes to the next. Figure 1 depicts a high-level detail of
the entire framework. In Phase 1, a literature review is
undertaken to determine the research problem. The outcome
of this phase is the research gaps that will be addressed in the
subsequent phases. Based on the identified problem and gaps,
Phase 2 carries out the first objective of this study which is
dimensionality reduction in healthcare data using the PCA
technique. The reduced vital data is the result of this phase.
Phase 2's reduced medical data will be employed as an input
for Phase 3. In Phase 3, the reduced medical data will be
applied to train the MWSN anomaly detection model, and
then, the evaluated accuracy will be compared to the existing
methods. The output of this phase is a comparison of the
proposed model's accuracy with PCA and without PCA
technique.
There are two stages to the implementation of the
dimensionality reduction proposed in the first objective; these
are offline training and online implementation. Data was
collected via measurements throughout the training period.
Several pre-processing operations were performed on the
collected medical data, including standardization, noise
removal, normalization, and imputation. The PCA technique
was used in the reduction procedure to get the reduction
parameters, such as eigenvectors and eigenvalues. To ensure
that the dimension of the incoming data from the node is
reduced, these parameters are conserved in the sensor before
the implementation phase. They are also used to calculate the
measurements in the cluster head (CH).
During the online implementation phase, PCA relies on
the stored eigenvectors and eigenvalues determined during
the training phase to performs real-time dimensionality
reduction on the medical data. Then, the low dimensional
medical data is forwarded to the CH where it is restored using
an approximation in its original form based on the same
parameters used during the reduction step. If the observed
changes in medical parameters are beyond a set threshold, the
medical data is recalculated. The readings of the reduced
medical data that replaced the original data during the
detection stage are the outcome of this phase. Consequently,
the communication overhead, memory usage, and
computational complexity of the suggested AD model were
dramatically decreased during the real-time process. This
improves the worthiness of the suggested AD models as it is
trained in limited medical data. Section 3 is the
implementation specifics of the PCA technique and its
incorporation into the suggested model for AD.
For the second objective, PCA was incorporated into the
AD model for MWSN with two stages - offline training, and
online detection. During offline training, the training medical
data were collected from all sensors that constitute the
MWSN. These medical data underwent the pre-processing
step for standardization, normalization, and noise removal.
After that, the PCA technique proposed in the first objective
was used to carry out the dimensionality reduction. logistic
regression algorithm was trained using the reduced data to
construct the classifier that distinguishes between the normal
behaviour and the malicious one based on the normal
reference inferred from the medical data. This normal
reference was then stored at each node for later use during
online detection The implementation of this model is
discussed in detail in section 3.
Procedures Action
1-
Study the Literature Determine the related problems
2-
Design and implement the
PCA
 Design the proposed
dimensionality reduction
technique.
3-
Train the MWSN anomaly
detection model
 Training the model.
 Validation.
 Testing.
 Coding.
4-
Experiments and
evaluation
 Result.
 Discussion.
Fig. 1. The high- level detail of the entire proposed framework
For the online detection phase, the new vital data
observed by any node in the MWSN are collected. Similar to
the training phase, the new observation undergoes pre-
processing activities like standardization, normalization,
imputation, and noise removal. PCA technique was also used
for dimensionality reduction. After that, the observation was
compared against the normal reference stored at the offline
training in the respective node to identify whether it is
normal. It is worth noting that both training and detection
phases took place locally at the node nodes level. The
outcome of this objective is the efficient online anomaly
detection model that trained the logistic regression algorithm
using reduced data from the proposed first objective. The
2021 International Conference on Data Science and Its Applications (ICoDSA)
67
Authorized licensed use limited to: UNIVERSITY TEKNOLOGI MALAYSIA. Downloaded on February 05,2022 at 13:28:28 UTC from IEEE Xplore. Restrictions apply.

design and implementation of this model were elaborated in
section 3.
III. RESULT DISCUSSION AND EVALUATIONS
This section discussed the proposed model's design and
implementation, as well as the facts acquired as a result of its
execution. The proposed model was described in Section 3.1.
The suggested PCA-based anomaly detection model's design
was discussed in Section 3.2, while Section 3.3 discussed the
evaluation metrics. Finally, the experimental results and
comparison were detailed in Section 3.4
3.1. PCA Technique-Based Anomaly Detection Model
This study implemented a similar transformation
procedure as found in most previous studies. The value of
each feature was normalized to a range of 0 and 1. Note that
the transformation and standardization of all medical data
occur at the CH in the centralized scenario. But in the
distributed scenario, a summary of the collected data by each
node is sent to the CH and this information contains the
number, linear sum, & linear sum of squares of the local data
vectors. Furthermore, each node transmits the minimum and
maximum values of each feature to the CH for the
computation of the global max, global min, global mean, and
global variance after receiving this data. The sensor nodes
receive these global values and use them for their local data
pre-processing. Having a reduced number of features in a
dataset improves the performance of the anomaly detector via
speeding up the detection process and improving its accuracy.
The hierarchical or cluster-based network structure was
used at the present study, as illustrated in Figure 2. The
network was divided into clusters in this design, with every
cluster having an strategy node called the CH; this CH has
extra processing power and energy and is responsible for data
processing and propagation from the other sensors to the base
station. The use of these clusters and the assignment of
specific tasks to the CHs significantly improves the
scalability, energy efficiency, and network lifetime of the
system. After applying PCA locally, other sensor nodes
(SNs) participate in the dimensionality reduction process, as
well as provide only a summary of their reduced data to CH.
The CH, which operates as a sensor node, performs an
approximation of the original data before sending it to the
sink. Each cluster's nodes are assumed to be static,
homogeneous, and time-synchronized.
Fig. 2. The hierarchical or cluster-based sensor network structure
3.2. The design of the suggested PCA-based AD model
In this phase, the PCA technique was applied in each node
to extract the principal components that represent the medical
data in the feature space. This phase was carried out offline
after collecting sufficient data from the sensors. The features
extracted by PCA were then fed to a machine learning
classifier and train the detection model using the logistic
regression algorithm that can distinguish between normal and
anomalous medical data. The same procedure was applied
online as the technique is lightweight in terms of computation
and requires no intensive operations. The eigenvectors and
eigenvalues are the products of the PCA process and were
used as the features for the training of the ML classifier
model. The pseudocode of the initialization step of the model
is shown in Algorithm 1 while Figure 3 depicted the proposed
model's design.
Algorithm 1. The pseudocode of the initialization step of the model
3.3. Evaluation Metrics
The performance of the proposed PCA-based
dimensionality reduction technique was evaluated in terms of
the approximation error, and approximation accuracy; these
were the popular performance evaluation metrics used by
most of the previous studies to determine the effectiveness of
novel PCA-based dimensionality reduction techniques
[11],[12]. Furthermore, the effectiveness of the suggested AD
system for MWSN were evaluated based by several metrics,
such as the detection rate (DR), false-positive rates (FPR),
detection accuracy (DA), as well as false-negative rates
(FNR); these metrics were adopted from previous studies
[13]. The efficiency of the new method was also evaluated
based on the communication overhead, memory usage, and
computational complexity as obtainable in the related [14].
The DA, DR, precision, and F-measure were calculated using
Equations 1, 2, 3, & 4, respectively based on the confusion
matrix and related evaluation metrics for evaluating anomaly
detection models.
Input: Data collected from sensors
Output: The classification of the data as normal or
anomalous
##Training Phase (offline)
1: Collect data from sensors.
2: Do data normalization and Standardization
3: Extract the raw features from data
4: Apply PCA on the raw data and features
5: select the best n eigenvector as PCA features.
6: feed the selected PCA features into logistic regression
classifier.
##Testing phase (online):
7: read the new measurement (data) from the sensor
8: apply PCA to newly obtained data
9: Feed the data with PCA features into the classifier
10: The classifier will determine whether the
measurement is normal or anomalous.
2021 International Conference on Data Science and Its Applications (ICoDSA)
68
Authorized licensed use limited to: UNIVERSITY TEKNOLOGI MALAYSIA. Downloaded on February 05,2022 at 13:28:28 UTC from IEEE Xplore. Restrictions apply.

Fig. 3. The proposed model's design
4. RESULT ANALYSIS AND COMPARISON
Figures 4, 5, and 6 showed the distribution of the medical
data gathered from the sensor nodes. The medical data were
not normally distributed as seen in Figures 4, 6, and 7 due to
the randomness of medical data as body temperature, heart
rate, and blood pressure were all shaped by a multitude of
variables. These variables are uncontrollable; furthermore,
these three sensors captured medical data that are relatively
associated. As a result, variability in one of these critical facts
causes randomness in the others. The medical data collected
by the body temperature sensor, on the other hand, was
distributed normally due to the measurement's predictability
and the restricted range of values for body temperature.
Fig. 4. The distribution of medical data collected by the Oxygenation ratio
sensor.
Fig. 5. The distribution of medical data gathered by the body temperature node
Fig. 6. The distribution of medical data gathered by the blood pressure node
The performance of the suggested AD method with and
without the PCA technique was shown in Figure 8. As we can
see, the suggested PCA-based system performed better with
PCA as the values of the precision and recall of the proposed
F1 model were greater than the values for the model without
PCA. This improved performance is due to the data
dimensionality capability of PCA as it selected the most
2021 International Conference on Data Science and Its Applications (ICoDSA)
69
Authorized licensed use limited to: UNIVERSITY TEKNOLOGI MALAYSIA. Downloaded on February 05,2022 at 13:28:28 UTC from IEEE Xplore. Restrictions apply.

relevant features for use by the detection model. The impact
of this data dimensionality reduction capability on the model
performance is that the problem of overfitting which
deteriorates DA is prevented. Furthermore, the performance
of the model in terms of accuracy was slightly higher with
PCA than without PCA because of the issue of high false
alarms associated with the traditional AD techniques. The
extent of improvement achieved by the new model over the
conventional models was determined by checking for the t-
test value at the α value of 0.05. Observably, the model
achieved a p-value of 0.03 which was <0.05 [15], suggesting
the significance of the suggested PCA-based AD model.
Fig. 7. The distribution of medical data gathered by the heart rate node
Fig. 8. The comparison of the performance of the proposed model with and
without the PCA-based AD technique.
5. CONCLUSION
This study designed and implemented a PCA-based AD
model for MWSNs. The model was built in two main phases
which are the training and testing phases. In the training
phase, data were collected from different sensors and pre-
processed. During the pre-processing step, several
procedures were implemented on the data, such as data
normalization and standardization. PCA was implemented on
the data to reduce the dimensionality and prevent overfitting.
The features extracted by PCA were fed to a machine learning
classifier for the training of the detection model using the
logistic regression algorithm. The developed model with and
without the PCA-based dimensionality reduction step was
evaluated and compared in terms of performance using
various performance metrics.
ACKNOWLEDGMENT
We would like to express our very great appreciation to
Dr. Hiyam N. Khakid for her valuable and constructive
suggestions during the planning and development of this
research work. Her willingness to give her time so generously
has been very much appreciated.
We would also like to thank the staff of the following
institutions for enabling us to visit their Labs to do our
operations:
 University of Misan
 Missan Oil Training Institute
REFERENCES
[1] M. A. M. El-Bendary, H. Kasban, A. Haggag, and M.
A. R. El-Tokhy, “Investigating of nodes and personal
authentications utilizing multimodal biometrics for
medical application of WBANs security,” Multimed.
Tools Appl., vol. 79, no. 33, pp. 24507–24535, 2020.
[2] G. Shanthi and M. Sundarambal, “FSO–PSO based
multihop clustering in WSN for efficient medical
building management system,” Cluster Comput., vol.
22, no. 5, pp. 12157–12168, 2019.
[3] S. Izza, M. Benssalah, and K. Drouiche, “An
enhanced scalable and secure RFID authentication
protocol for WBAN within an IoT environment,” J.
Inf. Secur. Appl., vol. 58, p. 102705, 2021.
[4] A. S. H. Altamimi, O. R. K. Al-Dulaimi, A. A.
Mahawish, M. M. Hashim, and M. S. Taha, “Power
minimization of WBSN using adaptive routing
protocol,” Indones. J. Electr. Eng. Comput. Sci., vol.
19, no. 2, pp. 837–846, 2020.
[5] M. Salayma, A. Al-Dubai, I. Romdhani, and Y.
Nasser, “Wireless body area network (WBAN) a
survey on reliability, fault tolerance, and
technologies coexistence,” ACM Comput. Surv., vol.
50, no. 1, pp. 1–38, 2017.
[6] Taha, M. S., Rahim, M. S. M., Hashim, M. M., &
Khalid, H. N. (2020). Information Hiding: A Tools
for Securing Biometric Information. Technology
Reports of Kansai University, 62(04), 1383-1394. [7]
M. S. Hajar, M. O. Al-Kadri, and H. K. Kalutarage,
“A survey on wireless body area networks:
architecture, security challenges and research
opportunities,” Comput. Secur., p. 102211, 2021.
[8] S. J. Hussain, M. Irfan, N. Z. Jhanjhi, K. Hussain, and
M. Humayun, “Performance enhancement in
wireless body area networks with secure
communication,” Wirel. Pers. Commun., vol. 116,
no. 1, pp. 1–22, 2021.
[9] S. GS and R. Balakrishnan, “A Statistical-Based
2021 International Conference on Data Science and Its Applications (ICoDSA)
70
Authorized licensed use limited to: UNIVERSITY TEKNOLOGI MALAYSIA. Downloaded on February 05,2022 at 13:28:28 UTC from IEEE Xplore. Restrictions apply.

Light-Weight Anomaly Detection Framework for
Wireless Body Area Networks,” Comput. J., 2021.
[10] A. El Aalaoui and A. Hajraoui, “Energy efficiency of
organized cluster election method in wireless sensor
networks,” Indones. J. Electr. Eng. Comput. Sci., vol.
18, no. 1, pp. 218–226, 2020.
[11] Alrikabi, Hanan Ali, et al. "Using FFNN classifier
with HOS-WPD method for epileptic seizure
detection." 2019 IEEE 9th International Conference
on System Engineering and Technology (ICSET).
IEEE, 2019.
[12] Y.-A. Le Borgne, S. Raybaud, and G. Bontempi,
“Distributed principal component analysis for
wireless sensor networks,” Sensors, vol. 8, no. 8, pp.
4821–4850, 2008.
[13] M. Xie, J. Hu, S. Han, and H.-H. Chen, “Scalable
hypergrid k-NN-based online anomaly detection in
wireless sensor networks,” IEEE Trans. Parallel
Distrib. Syst., vol. 24, no. 8, pp. 1661–1670, 2012.
[14] M. Moshtaghi et al., “Clustering ellipses for anomaly
detection,” Pattern Recognit., vol. 44, no. 1, pp. 55–
69, 2011.
[15] J. Wang, J.-M. Wei, Z. Yang, and S.-Q. Wang,
“Feature selection by maximizing independent
classification information,” IEEE Trans. Knowl.
Data Eng., vol. 29, no. 4, pp. 828–841, 2017.
AUTHOR BIOGRAPHY
Mustafa Sabah Taha is a senior researcher
at Missan Oil Training Institute. He
earned his Ph.D. degree in Information
Security from University Technology
Malaysia (UTM) in 2020. During his
Ph.D. study, he was accorded several
honorable awards as recognition for his
level of excellence and tenacity, such as
GOT (Graduate-on-Time) award, and the
Best Researcher award from UTM. His research work has
been published in several reputable academic journals, book
chapters, and refereed conference proceedings. His main
research interest is in Image Processing, Information
Security, Wireless Sensor Networks, Wireless Body Area
Network and Internet of Things (IoT). He is a member of
IEEE since 2017.
2021 International Conference on Data Science and Its Applications (ICoDSA)
71
Authorized licensed use limited to: UNIVERSITY TEKNOLOGI MALAYSIA. Downloaded on February 05,2022 at 13:28:28 UTC from IEEE Xplore. Restrictions apply.