You are on page 1of 11

A HYBRID FILTER-WRAPPER ATTRIBUTE

REDUCTION APPROACH FOR FETAL RISK


ANTICIPATION

V.Subha1, D.Murugan2, A.Manivanna Boopathi3


1
Assistant Professor, Department of Computer Science and Engineering, Manonmaniam
Sundaranar University, Tirunelveli, India.
subha_velappan@yahoo.co.in
2
Associate Professor, Department of Computer Science and Engineering,
Manonmaniam Sundaranar University, Tirunelveli, India.
dhanushkodim@yahoo.com
3
Associate Professor, Department of Electrical & Electronics Engineering,
PSN College of Engineering & Technology, Tirunelveli, India.
arumugamboopathy@gmail.com
ABSTRACT
In the present days, use of computers has permeated in archiving and analysing vast
amount of medical data. Data mining techniques are widely used on these data to excerpt the
information and by which an accurate and easy prediction of diseases has become possible. In
this paper, Support Vector Machine (SVM) is used to analyse the Cardiotocogram (CTG) data
for the prediction fetal risk. A hybrid method which combines Information Gain (IG) and
Opposition based firefly algorithm (OBFA), is proposed in this paper to excerpt the most
relevant features/attributes which will improve the classification performance of SVM. The
results prove that the proposed hybrid method performs better than the other existing methods.

KEYWORDS
Cardiotocography, SVM Classifier, Feature Selection, Information Gain, Opposition-based firefly
algorithm

1. INTRODUCTION
One of the popular tests done during the third trimester of pregnancy is
Cardiotocography (CTG). This test helps obstetricians in early detection of fetal distress by
monitoring the fetal heart rate (FHR) and uterus contraction (UC). More investigations may be
performed to detect abnormalities such as acute or chronic hypoxia. Mere manual and visual
analysis of CTG data might often lead to incorrect interpretations. Hence, computer aided
systems are used to classify the CTG data in order to help the obstetricians on deciding whether
there can be a chance for natural birth of baby or caesarean.

Many methodologies are found in literature for analysing the CTG data. A SVM
classifier [1] is used to classify the fetal state in to one of two classes, in which genetic
algorithm is used to find out the most relevant features which will result in improved
classification performance. A combination of least squares SVM and binary decision tree with
particle swarm optimization technique is proposed in [2] for the classification of CTG data with
10 fold cross validation. Further, in [3] the CTG data are classified into classes using an
adaptive neuro fuzzy inference system (ANFIS) which is trained well to do the prediction task.
A random forest classifier combined with feature reduction technique is given in [4] for CTG
data classification and thereby automatically predicting the fetal state. Accuracy of classification
is improved in [5] by using discriminant analysis method, decision tree and artificial neural
network for the prediction of fetal distress. SisPorto 2.0 based classifiers for CTG data using
neural network and simple logistic based methods are proposed in [6]. With feature selection
approaches, a Nave Bayes Classifier is used for classification of CTG data in [7]. In this work,
ReliefF, Information Gain, Mutual Information and correlation-based methods of feature
selection are used with the Nave Bayes Classifier. A modular neural network which can
classify the CTG data in to three different classes is developed and proposed in [8]. It attempts
at the objective which conflict in multi class classifiers. Another neural network based
supervised classifier is presented in [9] which improves the classification performance when
compared to other unsupervised methods of clustering. A Nave Bayes Classifier is used to
classify the CTG data in to three classes in [10]. Measures such as Precision, Recall and F-
Score are evaluated for each class to assess the performance of the classifier and shown that this
classifier performs better than other conventional clustering methods. An artificial bee colony
algorithm is used in [11], to find the most relevant feature in order to achieve improved
classification accuracy. A hybrid feature selection method is developed and presented in [12]
using Support Vector Machines for the classification of medical datasets based on artificial bee
colony algorithm, by eliminating the irrelevant, unimportant and obsolete features. This
proposed method is used to diagnose diabetes, hepatitis, disorders in liver, etc. A fetal state
classifier is presented in [13], which uses SVM to classify and Genetic Algorithm (GA) to find
the most relevant features. This method improves the classification performance. To improve
the classification performance further, GA is replaced by Firefly Algorithm (FA) in [14].
Reducing the attributes by extracting most relevant features using Opposition-Based Firefly
Algorithm (OBFA) is proposed in [15].

Albeit there are many methods used for judging the feature used to classify the data set,
the popular method is using information gain [16-19]. Relevant features should be known to
perform the feature selection task. A feature is relevant if its removal from the feature set harms
the prediction power of the feature set. In simple words, a feature is a relevant one if it contains
valuable information about the data set and it cannot be replaced by any other features. The
relevance is measured in terms of Information Gain (IG).

In this paper, a hybrid method incorporating information gain (IG) with opposition
based firefly algorithm (OBFA) together with Support Vector Machine (SVM) classifier is
proposed for classification of CTG data. In the proposed methodology, initially the CTG data
are classified using the full feature set. Further, optimal feature set is produced using the
proposed hybrid method which combines IG and OBFA along with SVM classification.
Optimal feature sets are produced using IG, GA, FA and OBFA separately to compare the
performance of the proposed methodology. The results show that the proposed hybrid method
based optimal feature set considerably improves the accuracy of classification.

This paper is organized in 8 sections. The section 2 describes the CTG data set. The
SVM classifier is given in section 3 followed by the information gain attribute reduction
technique explained in section 4. Section 5 and 6 describe Firefly Algorithm (FA) and
opposition based firefly algorithm (OBFA) optimization techniques respectively. The proposed
methodology to find the optimal and reduced data set is presented in section 7. Finally the
results and discussion are consolidated in section 8 with conclusions in section 9.

2. CTG DATA SET


CTG data contains the measurement information of Fetal Heart Rate (FHR) and uterine
contraction (UC) obtained through cardiotocograms. In the present work, the CTG dataset of
UCI Machine Learning Repository [20] is taken for the experiments. Eight different classes
with 21 attributes and 1 class attribute with a total of 2126 fetal cardiotocograms are available in
this dataset. This dataset has been classified by three expert obstetricians with regard to a
morphologic pattern and to a fetal state consisting measurements of fetal heart rate and uterine
contractions and assigned classification labels to them based on the three fetal heart rate class
codes namely Normal, Suspect and Pathologic. The attributes of CTG data set are shown in
Figure 1.
Figure 1. Attributes of CTG data set

3. SUPPORT VECTOR MACHINE


Support vector machine (SVM) has been widely used for solving classification
problems [18]. SVM separates the classes with an optimal hyperplane that increases the margin
between the classes. The data points closest to this hyperplane are called support vectors. The
nonlinear data used in this work is subjected to nonlinear kernel functions to transform the data
into a new feature space where a hyperplane separates the data. Radial Basis Function (RBF)
has been used widely by researchers for its better generalisation capability and hence RBF
kernel has been adopted in this work. A One-Against-All SVM classifier is used here to classify
the data into three classes.

4. INFORMATION GAIN
Information gain (IG) is one of the key concepts of information theory [16]. A most
common evaluation metric used for finding information gain is entropy, which gives the
measure of the information content of an attribute in the given dataset. More the value of
entropy means higher the information content. The information gain explores the importance of
an attribute in a dataset for classification.

Information Gain attribute evaluation [19] is a filter based supervised feature selection
technique which is based on the measure of dependence between the feature and the class label.
It is one of the most powerful feature selection techniques and it is easy to compute and simple
to interpret. Information Gain (IG) of a feature A and the class label X is calculated as,

IG(A, X) H(A) H(A X) (1)

Where, H(A) is the entropy of A and H(A X) is the entropy of A after observing X.

The values of above listed entropies can be calculated as;

H ( A) i P(ai ) log2 ( P(ai )) (2)

H(A X) i P( x i ) i P(a i x j ) log2 (P(a i x j )) (3)

Initially, the values of Information gain are independently evaluated for all features and
then the features with the top k values are taken in to consideration as the relevant features.
The maximum value of information gain will be 1.

5. FIREFLY ALGORITHM
Firefly algorithm is one of the popular nature-inspired, meta-heuristic and stochastic
methods proposed by Xin-She Yang [23]. FA is to solve diverse engineering optimization and
classification problems by modifying and/or hybridizing it [24]. Fireflies belong to insect
family, having a property of using flashes to communicate amongst them. The light is produced
from firefly as a result of biochemical reaction and this phenomenon is called as
bioluminescence. Each male firefly produces a distinct flashing pattern to attract the female
fireflies. The flash pattern produced by a male firefly contains the information on identity and
gender [25].

Three idealised rules are assumed in Firefly algorithm,


- A firey will attract other rey irrespective of its gender.
- The attractiveness of a rey is proportional to its light intensity, hence the less brighter
firefly will be attracted towards the brighter one.
- Light intensity of a rey is related to the fitness function.

The pseudo code of firefly algorithm is given in Figure 2.

Figure 2. Pseudo code of the standard rey algorithm

The parameters of FA are considered as given in Table 1.

Table 1. Firefly algorithm parameters


d Dimension of function to be optimized
n Number of fireflies
bmax maximum number of generations
light absorption coefcient
attractiveness
Ti light intensity
q distance
randomization parameter
rand ( ) random number generator
The physical distance q between any two reies i and j located at positions Ni and Nj
can be evaluated as follows.
ni ,k n j ,k 2
d
qi j Distance Ni , N j (4)
k 1
The light intensity (T) will increase as the square of the distance (q2) decreases
as approximately given below.
2
T(q) T0 e q (5)

Where T0 is the light intensity at source. The attractiveness () of firefly can be


given in mathematical form as;
(q) = 0 e q
2
(6)
Here, 0 is the attractiveness at the distance of 0.
The movement of a rey i attracted by another more attractive rey j can be
mathematically written as,
n i 1 n i 0 e
q i j 2
n j n i rand( ) 0.5 (7)

The global best solution in a standard firefly algorithm will be the firefly that
possess highest light intensity or attractiveness. The attractiveness may be lost since the
firefly moves in a random fashion in the next iteration. This may lead to premature
convergence [23].

6. OPPOSITION-BASED FIREFLY ALGORITHM


The standard firefly algorithm is modified using opposition-based learning (OBL) to
overcome the aforementioned issue. The candidate solution and its corresponding opposite
solution are taken in to consideration in OBL.

Let c [a, b] be a real number, the opposite number of c is denoted as c` and is dened
as,

c` a b c (8)
Extending this for higher dimensions for R (c1, c2,..., cn) being a n-dimensional vector,
where ci [ai , bi ] and i = 1, 2,..., n, the opposite vector of R is dened by,

R` = (c`1, c`2,..., c`n), where c`i= ai + bi ci (9)

The OBL is utilized at the stage of initialisation of population and creating next
generations. The opposite position of each of m fireflies being initialised are evaluated as given
in equation (8). Based on the evaluated fitness values, the m fittest individuals are found out
from the 2m individuals. The pseudo code of OBFA is given in Figure 3.

During each iteration, the OBFA replaces e number of fireflies whose fitness values are
worst by their opposite fireflies. The value of e, as calculated by equation (10), should be kept
as large during the starting of the iterations to achieve a good global search. However, the value
of e has to be reduced to ensure local exploitation. At the start, variable e should possess a larger
value to provide an effective global search. As the iteration increases, the value of e should be
reduced to provide a local exploitation. Therefore, the value of e is given as follows,

m b
e Round 1 (10)
3 bMax
Figure 3. Pseudo code of OBFA
7. HYBRID IG-OBFA BASED FEATURE SUBSET SELECTION USING SVM
Certain features are ignored based on their irrelevance using wrapper method or filter
method which use the performance of a classifier or feature evaluation techniques respectively.
In the proposed methodology, CTG data with all features is classified using SVM classifier.
Then, the hybrid IG-OBFA is combined with SVM to find the most relevant and optimal set of
features.

The process of reducing the attributes is done in two stages. Firstly, the IG for the full
feature set is evaluated and ranked as features with large values of IG on the top and least value
at the bottom. Then, the top 15 features with maximum IG are selected as reduced feature set.
Secondly, the obtained reduced feature set is set as initial population to OBFA and the optimal
feature set is produced by OBFA.

The presence and absence of a particular feature set in the data set in represented by 1
and 0 respectively. The full CTG data set is divided in to 75:25 ratio (1594:532 instances) and
the larger part is used for training and the smaller part is used for testing the classifier.

The fitness function (F) can be written as the function of average accuracy rate (Ec) of classifier
and the number of features being absent (Tf) in the data set, as given in equation (11).

F w1 E c w2T f (11)

The values assigned for the parameters of Firefly algorithm are listed in Table 2.

Table 2. Values of Firefly algorithm parameters

Number of fireflies 30
Number of generations 100
Randomisation parameter () 0.5
Attractiveness () 0.2
Light absorption coefficient () 1
Weights (w1 & w2) 1, 0.1

8. RESULTS & DISCUSSION


The original and obtained optimal data sets are used in experiments to assess the
performance of the proposed method. Various performance measures used to assess the
performance of proposed method are given in Table 3.

Table 3. Expressions for performance measures

TP TN TN
Accuracy = Negative Predictive Value: NPV
TP TN FP FN TN FN

TP Geometric mean:
Sensitivity =
TP FN Gmean speificity sensitivity

Specificity =
TN F-measure = 2 precision sensitivity

TN FP precision sensitivity

Positive Predictive Value:


TP Area under ROC = Sensitivity Specificity
(Precision): PPV 2
TP FP

TP - True Positives, TN - True Negatives, FP - False Positives and FN - False Negatives

The evaluated SVM accuracies found with full feature set and reduced feature sets
obtained using IG, FA, OBFA and hybrid IG-OBFA are consolidated in Table 4. It is found that
the average accuracy is 88.75% with full feature set (without feature selection) and the same is
achieved as 89.47 with optimal feature set produced by IG, 91.92% with optimal feature set
produced by FA and as 92.85 with optimal feature set produced by OBFA. The maximum
accuracy 96.24% is achieved with optimal feature set produced by the proposed hybrid IG-
OBFA.

Table 4: SVM accuracies

Data set Accuracy


Without FS Full feature set 88.75
IG 89.47
FA 91.92
With FS
OBFA 92.85
Hybrid IG-OBFA 96.24

The values of other performance measures given in Table 3 for various methods are
tabulated in Table 5.
Table 5: Performance metrics of SVM with and without feature selection

With FS
Performance Without Filter
Wrapper method Hybrid
Metrics (%) FS method
method
IG GA FA OBFA
Specificity 90.22 91.14 92.50 93.78 93.72 96.26
Sensitivity 77.79 81.07 80.71 84.83 83.81 91.92
PPV 78.29 78.48 83.06 83.14 85.45 93.33
NPV 90.70 91.29 93.77 93.26 95.02 97.44
G-mean 83.77 85.96 85.92 89.19 88.62 94.06
F-measure 78.08 79.75 81.87 83.94 84.62 92.61
Area under ROC 84.00 86.11 86.61 89.30 88.76 94.09

Figures 4 and 5 are the graphical representations of the values given in Table 4 and 5
respectively.

Figure 4. SVM accuracies (with & without FS)

Figure 5. Performance metrics of SVM with and without feature selection


9. CONCLUSION
In this paper, a hybrid attribute reduction method incorporating Information Gain and
Opposition based Firefly algorithm is proposed to find optimal feature set for CTG data
classification. The CTG dataset has been taken from UCI Machine Learning Repository to
assess the performance of the proposed method and comparison with that of existing methods.
The comparison is done in terms of the performance measures such as Accuracy, Sensitivity,
Specificity, Positive Predictive Value, Negative Predictive Value, Geometric Mean, F-measure
and Area under ROC. From the obtained results and comparison, it is very obvious that the
proposed method of classification performs much better than the existing methods. Hence, this
proposed method will eventually be much useful for making a better and accurate diagnosis on
fetal state from the CTG data.

REFERENCES
[1] Hasan Ocak, A medical decision support system based on support vector machines and the
genetic algorithm for the evaluation of fetal well-being, Journal of Medical Systems, vol.
37, no.2, Article ID: 9913, 2013.
[2] Ersen Ylmaz and alar Klker, Determination of fetal state from using LS-SVM with
particle swarm optimization and binary decision tree, Computational and Mathematical
Methods in Medicine, 2013.
[3] Hasan Ocak and Huseyin Metin Ertunc, Prediction of Fetal State From The
Cardiotocogram Recordings using Adaptive Neuro-Fuzzy Inference Systems, Neural
Comput & Applic., vol.23, no.6., pp. 15831589, 2013.
[4] Peterek Tomas, Jana Krohova, Pavel Dohnalek and Petr Gajdos, Classification of
Cardiotocography Records by Random Forest, IEEE 36th International conference on
telecommunications and signal processing, pp.620-623, July 2013.
[5] Mei-Ling Huang and Yung-Yan Hsu, Fetal distress prediction using discriminant analysis,
decision tree, and artificial neural network, J. Biomedical Science and Engineering, vol. 5,
no. 9, pp.526-533, 2012.
[6] Hakan Sahin and Abdulhamit Subasi, Classification of Fetal State from the Cardiotocogram
Recordings using ANN and Simple Logistic, In: 3rd International Symposium on
Sustainable Development, Sarajevo, pp.499-505, 2012.
[7] Mohamed El Bachir Menai, Fatimah J. Mohder and Fayha Al-mutairi, Influence of Feature
Selection on Nave Bayes Classifier for Recognizing Patterns in Cardiotocograms, Journal
of Medical and Bioengineering, vol. 2, no.1, pp.66-70, 2013.
[8] Shivajirao Jadhav, Sanjay Nalbalwar and Ashok Ghatol, Modular Neural Network Model
Based Foetal State Classification, IEEE International Conference on Bioinformatics and
Biomedicine Workshops, pp.915-917, November 2011.
[9] Sundar.C., M. Chitradevi and G. Geetharamani, An Overview of Research Challenges for
Classification of Cardiotocogram Data, Journal of Computer Science, vol. 9, no. 2, pp.
198-206, 2013.
[10] Sundar.C, M.Chitradevi and G.Geetharamani, An Analysis on the Performance of Nave
Bayes Probabilistic Model Based Classifier for Cardiotocogram Data Classification,
International Journal on Computational Sciences & Applications, vol. 3, np. 1, pp.17-26,
2013.
[11] Mauricio Schiezaro and Helio Pedrini, Data feature selection based on Artificial Bee
Colony algorithm, EURASIP Journal on Image and Video Processing, vol. 47, pp.1-8,
2013.
[12] Mustafa Serter Uzer, Nihat Yilmaz and Onur Inan, Feature Selection Method based on
Artificial Bee Colony Algorithm and Support Vector Machines for Medical Datasets
Classification, The Scientific World Journal, Article ID 419187, 2013.
[13] V.Subha, D.Murugan, S.Prabha and A.Manivanna Boopathi, Genetic Algorithm based
feature subset selection for fetal state classificaion, Journal of communications technology,
Electronics and Computer Science, Issue 2, pp.13-17, 2015.
[14] V.Subha and D.Murugan, Fetal State Determination using Support Vector Machine and
Firefly Optimization, International Journal of Knowledge Based Computer Systems, vol. 2,
no. 2, pp. 7-12, 2014.
[15] V.Subha & Dr.D.Murugan, Opposition-based Firefly Algorithm Optimized Feature Subset
Selection Approach for Fetal Risk Anticipation, Machine Learning and Applications: An
International Journal, (MLAIJ) Vol.3, No.2, pp. 55-64, June 2016.
[16] Bangsheng Sui, Information Gain Feature Selection Based On Feature Interactions, M.S.
thesis, University of Houston, Dec 2013.
[17] B.Azhagusundari and Antony Selvadoss Thanamani, Feature Selection based on
Information Gain, International Journal of Innovative Technology and Exploring
Engineering (IJITEE) ISSN: 2278-3075, Vol.2, No.2, pp.18-21, January 2013.
[18] T. Mitchell, Machine Learning, McGraw-Hill, New York, 1997.
[19] R.Porkodi, "Comparison of filter based feature selection algorithms an overview",
International journal of Innovative Research in Technology & Science, Vol.2, No.2, pp.108-
113, 2014.
[20] K.Bache and M.Lichman, UCI Machine Learning Repository
[http://archive.ics.uci.edu/ml]. Irvine, CA: University of California, School of Information
and Computer Science, 2013.
[21] Shuhao Yu, Shenglong Zhu, Yan Ma, Demei Mao, Enhancing firefly algorithm using
generalized opposition-based learning, Computing, pp.741-754, 2015.
[22] Tizhoosh HR, Opposition-based learning : a new scheme for machine intelligence,
International Conference on Computational Intelligence for Modelling, Control and
Automation Jointly with International Conference on Intelligence Agents, Web
Te chnologies and Internet Commerce, Vol 1, pp. 695-701, 2006.
[23] X.S. Yang, Nature-inspired metaheuristic algorithms, 2nd edition, Luniver Press, UK,
2010.
[24] A.Manivanna Boopathi and A.Abudhahir, Firefly Algorithm tuned Fuzzy Set-point
Weighted PID Controller for Antilock Braking Systems, Journal of Engineering Research,
Vol. 3, No. 2, pp.79-94, 2015.
[25] I.Fister, I.Fister Jr, X.S.Yang and J.Brest, A comprehensive review of firefly algorithms,
Swarm and Evolutionary Computation, Vol. 13, pp. 34-46, 2013. DOI: 10.1016/
j.swevo.2013.06.001.

Authors

V. Subha received B.E. (Electronics and Communication Engineering)


from Bharathiyar University, Coimbatore. She received her M.E.
(Computer Science and Engineering) from Manonmaniam Sundaranar
University, Tirunelveli. She is presently working as Assistant Professor
and pursuing research towards Ph.D. degree in Manonmaniam
Sundaranar University. Her areas of interest are Data Mining and
Pattern recognition.

Dr.D.Murugan received B.E. (Electronics and Communication


Engineering) and M.E. (Computer Science and Engineering) from
Madurai Kamaraj University respectively. He received his Ph.D. from
Manonmaniam Sundaranar University, Tirunelveli. He is presently
working as Associate Professor in the Department of Computer Science
and Engineering, Manonmaniam Sundaranar University. Presently he is
a Principal Investigator of the two-funded research projects granted by
UGC and DST. His areas of interest are Image processing, Software
engineering, Data mining and Pattern recognition.
Dr.A.Manivanna Boopathi received his B.E. (Instrumentation and
Control Engineering) from Madurai Kamaraj University. He received
his Master in Control System Engineering degree from Jadavpur
University, Kolkata and Ph.D. from Anna university, Chennai. He is
presently working as Associate Professor at PSN College of Engineering
& Technology (Autonomous), Tirunelveli. His areas of interest are
Nonlinear control, Adaptive control and Pattern recognition.

You might also like