International Journal of Innovative Research and Practices Vol.

2, Issue 2, February 2014

ISSN 2321-2926

Staging Prediction in Cervical Cancer Patients – A Machine

Learning Approach

D. Sowjanya Latha1, P.V. Lakshmi2 and Sameen Fathima3

1. Associate Professor, Department of MCA, AMS School of Informatics, Hyderabad

2. Professor, Head of the Department, Department of Information Technology, GITAM University,
Visakhapatnam
3. Professor, Head of the Department, Department of Computer Science, Osmania University,
Hyderabad

Abstract: Cervical cancer is one of the most prominent diseases among women worldwide. It is a disease in
which cells of the cervix become abnormal and start to grow uncontrollably, forming tumors. Staging
specifies the extent to which the cancer has spread from the cervix and to other parts of the body. This is
captured from the various diagnoses (tests) like colposcopy, biopsy, imaging studies which are performed
after doctor’s physical examination. Proper staging is the most important factor in selecting the right
treatment plan. Finding out the correct diagnosis and planning of the proper treatment modality plays a
crucial role in the field of Oncology. Cervical cancer has a more or less well defined treatment modality.
Treatment modalities available are surgery, radiotherapy and chemotherapy or combination of these based on
the stage of the cancer. The objective of this paper is to compare the cancer data using data mining algorithms
and find out the efficient algorithm that helps in identifying the stage of the cancer. Performance of data
mining classification algorithms has been analyzed by computing their accuracy, sensitivity and specificity;
and the efficient algorithm has been identified. J48 has outperformed the other algorithms. Accuracy achieved
by J48 is 93.03%, specificity and sensitivity are above 0.8 for all the stages of the cancerous data (this
accuracy is better than any reported in the literature). The decision tree generated by J48 has picked the
attributes that are closely associated with the staging of the cancer. Sensitivity is above 0.7 and specificity for
all the stages is above 0.9 using J48. This would be useful for the medical oncologists in identifying the stage
of the cancer and plan their treatment accordingly.
Keywords: Cervical cancer, human papilloma virus, machine aided identification, feature selection, risk
factors for cervical cancer, staging, parametria

1. INTRODUCTION invasive cervical cancer is based on the stage, size

and histology grade of the primary tumor and the
Cervical cancer is the prominent form of cancer size of the lymph nodes. Assessment of disease
worldwide and ranks as the first common cancer stage is essential in determining proper
among women in India the age incidence being management in individual cases. The prognosis of
above 15 years [10]. Worldwide, cervical cancer is cervical cancer is influenced by local disease extent
the second leading cause of death among women. as determined on the basis of tumor volume, depth
Around 5, 00, 000 new cases are identified each and degree of tumor invasion, parametrial invasion,
year and more than 85 percent are in developing pelvic side wall extension, lymph node
countries (WHO/NCCC). This cancer is the only involvement, and presence of distant metastases.
major gynecologic malignancy that is staged
clinically according to International Federation of The current study is aimed at identifying the most
Obstetrics and Gynecology (FIGO) influential risk factors among the other factors in
recommendations. Early stage cervical cancer can order to reduce the number of deaths and creating
be cured at an average rate of 80% with either awareness among women.
radical surgery or radiation. Accurate cervical
cancer staging is crucial for appropriate treatment The contents of the paper are organized as follows;
selection and treatment planning. The prognosis of in section 2 talks about the biology of the disease
and diagnosis is presented. In section 3 we present

[Link] 14
International Journal of Innovative Research and Practices Vol.2, Issue 2, February 2014
ISSN 2321-2926

an overview of the studies that were made on early relatives who have suffered from cervical or related
diagnosis and staging of the cancer. Section 4 talks cancer), smoking, having a poor diet and or having
about the methods used. Experiments and results a weakened immune system (due to surgery,
are discussed in section 5 followed by conclusion Human Immunodeficiency Virus (HIV) infection,
in section 6. using immunosuppressive drugs). Use of the
contraceptive pill (over a long period of time) can
2. CERVICAL CANCER also potentially increase the risk of cervical cancer.
However, these factors are very minor compared
Cancer of the cervix occurs when the cells of the with HPV infection, and are more likely to exert
cervix change in a way that leads to abnormal indirect effects; for instance, poor diet can lead to
growth and invasion of other tissues or organs of the body not being able to fight off infection so
the body. The normal cells of the cervix first well, so if a person is infected with a high-risk
gradually develop pre-cancerous changes that turn HPV and has a poor diet, this could mean that the
into cancer. Since the most common form of body is less well equipped to fight off the virus
cervical cancer starts with pre-cancerous changes, allowing the virus to infect cells and potentially
there are two ways to stop this disease from cause changes that may eventually lead to cancer
developing. The first way is to find and treat pre- (these changes are referred to as precancerous
cancers before they become true cancers, and the changes). The carcinogens in cigarettes can cause
second is to prevent the pre-cancers. A well-proven damage to the cervical cells, possibly leading to
way to prevent cervix cancer is to have testing cervical cancer. Studies have shown that smoking
(screening) to find pre-cancers before they can turn can accelerate the cervical damage caused by HPV
into invasive cancer. The Pap test (or Pap smear) [40].
and the human papilloma virus (HPV) test are used
for this. If a pre-cancer is found it can be treated, Symptoms that may occur in the early stages can
stopping cervical cancer before it really starts. The include: abnormal vaginal bleeding between
second way is avoiding exposure to HPV. In periods, after intercourse, or after menopause,
women, HPV infections occur mainly in younger continuous vaginal discharge, which may be pale,
women and are less common in women older than watery, pink, brown, bloody, or foul-smelling,
30. Certain types of sexual behavior increase a periods become heavier and last longer than usual.
woman's risk of getting genital HPV infection, such Cervical cancer may spread to the bladder,
as: having sex at an early age, having many sexual intestines, lungs, and liver. Patients with cervical
partners, having a partner who has had many sex cancer do not usually have problems until the
partners, having sex with uncircumcised males[37]. cancer is advanced and has spread. Symptoms of
HPV enters the body, usually through a break in the advanced cervical cancer may include: back pain,
skin, and then infects the cells in the layers of the bone pain or fractures, fatigue, leaking of urine or
skin. The virus then replicates or multiplies in the feces from the vagina, leg pain, loss of appetite,
body. The time between first contracting HPV and pelvic pain, single swollen leg, weight loss[38].
the appearance of lesions can be weeks to months
or even years. HPV is transmitted by skin-to-skin Early cervical cancer does not present any
contact. HPV infections that cause skin warts (e.g., symptoms so regular screening is vital to ensure it
plantar or common warts) can be acquired through is detected as early as possible. Screening prevents
a cut, but the risk of transmission is low. Walking 84 out of every 100 cervical cancers that would
barefoot in public areas such as the gym or pool otherwise develop. Cervical screening involves a
can be a risk for infection with the types of HPV sample of cells being taken from the cervix and
that cause plantar warts. HPV infections that cause being examined under a microscope for signs of
genital warts are very contagious and are usually abnormality. If any abnormal cells are detected
contracted through sexual activity with an infected then the sample will be graded according to the
person. A mother with a genital HPV infection may severity of the abnormalities.
also transmit the virus to the infant during labor.
The risk factors for HPV infection include age, Precancerous changes of the cervix and cervical
number of sexual partners, immune system. There cancer cannot be seen with the naked eye. Special
are other risks linked to cervical cancer apart from tests and tools are needed to spot such conditions
HPV. These include genetic factors (having close [12]. Papanicolaou or Pap test is a most common

Staging Prediction in Cervical Cancer Patients – A Machine Learning Approach 15

International Journal of Innovative Research and Practices Vol.2, Issue 2, February 2014
ISSN 2321-2926

form of diagnosis for detecting cervical cancer in staging of cervical cancer with reference to
its early stage. Every woman, above 30 yrs or demographic factors and clinical data cancerous
above 18yrs of age who are sexually active, is and non-cancerous patients. Study also made with
recommended to undergo this test every year. If the reference to the cervix images, pap smear images,
test shows abnormality then further examination and genomics and so on. In this section various
has to be done. If Pap smear results reveal cervical studies have been quoted referring the research
abnormalities, colposcopy is then scheduled. Pieces done at various levels.
of tissue are surgically removed (biopsied) during
this procedure of colposcopy and sent to a A classifier system has been designed for cervical
laboratory for examination. Cone biopsy may also cancer diagnosis using bio chemical parameters of
be done. A small cone shaped sample of the tissue cancer patients using Support Vector Machine
is removed from the cervix. It is examined under a (SVM) and Classification and Regression Trees
microscope for signs of cancer[43]. Treatment of (CART) [21]. Unsupervised modeling techniques
cervical cancer depends on: the stage of the cancer, have been used for feature clustering and
the size and shape of the tumor, the woman's age classification of cervix images to automatically
and general health, desire to have children in the analyze the uterine cervix images by detecting the
future. Early cervical cancer can be cured by detection of cervix boundary and the opening of the
removing or destroying the precancerous or cervix [5]. Decision support system for cervical
cancerous tissue. There are various surgical ways to cancer management and staging was designed
do this without removing the uterus or damaging using soft computing tools like neural networks,
the cervix, so that a woman can still have children genetic algorithm and rough set theory to build an
in the future. Types of surgery for early cervical efficient decision making system for pattern
cancer include: loop electrosurgical excision classification and rule generation [31]. A new
procedure (LEEP) -- uses electricity to remove relaxation ranking algorithm was developed to
abnormal tissue, Cryotherapy -- freezes abnormal supplement the DNA (Deoxyribonucleic acid)
cells, Laser therapy -- uses light to burn abnormal methylation markers in cervical cancer so that the
tissue. A hysterectomy (removal of the uterus but number of validation steps used as part of the
not the ovaries) is not often performed for cervical experimentation would be reduced for detecting the
cancer that has not spread. It may be done in cancer in cervical scrapings [19]. Demographic
women who have repeated LEEP procedures. data, environmental and genetic factors have been
Treatment for more advanced cervical cancer may clubbed for analyzing the risk factors for cervical
include: radical hysterectomy which removes the cancer. A model was developed using induction
uterus and much of the surrounding tissues, technique in finding out the association among the
including lymph nodes and the upper part of the risk factors and hence generate rules for the
vagina, pelvic exenteration an extreme type of management of the disease [29]. A flexible
surgery in which all of the organs of the pelvis, decision based model has been developed using k-
including the bladder and rectum, are removed. means clustering technique for the physician to
Radiation may be used to treat cancer that has know the exact conditions for undertaking biopsy
spread beyond the pelvis, or cancer that has test using the demographic data [32]. Multispectral
returned. Chemotherapy uses drugs to kill cancer. pap smear image classification for cervical cancer
Pre-cancerous conditions are completely curable detection using a novel SVM-based feature
when followed up and treated properly [12]. Pre- screening method [8]. Identification of risk factors
cancerous conditions may take time to change to using fuzzy rough sets for detecting cancer at an
cancer. Staging helps in identifying these changes. early stage applied over demographic data [13].
The objective of this paper is to identify the factor/s Computerized clinical decision support system for
in identifying the stage of the cervical cancer so screening cervical cancer by interpreting the free-
that proper treatment can be given to the patient at text pap reports using Natural Language Processing
the right time. was developed [9]. HPV risk types have been
classified using support vector machine (SVM)
3. LITERATURE REVIEW classifier with gap-spectrum kernel based on k-
spectrum method [14]. Diagnostic performance of
Extensive studies have been made for the early Magnetic resonance Imaging (MRI) was evaluated
diagnosis, prediction of symptoms, prevention,

D. Sowjanya Latha, P.V. Lakshmi and Sameen Fathima 16

International Journal of Innovative Research and Practices Vol.2, Issue 2, February 2014
ISSN 2321-2926

in the pretreatment evaluation of invasive cervical The performance of each learning algorithm on
cancer especially for parametrial invasion and each fold can be tracked using some predetermined
lymph node involvement . A study was made on performance metric like accuracy. In our study we
the correlation between MRI involvement and applied ten-fold cross validation.
parametrial invasion on histology. It was found that
MRI measured tumor volume does not help as a Systems that construct classifiers are one of the
diagnostic criterion rather parametrial invasion is commonly used tools in data mining. Such systems
an important factor for cancer treatment because of take as input a collection of cases, each belonging
low accuracy; less than 60% [42]. Medical imaging to one of a small number of classes and described
techniques often detect cancer at its early stage by its values for a fixed set of attributes, and output
when it is curable and least costly to be treated a classifier that can accurately predict the class to
upon [24]. which a new case belongs. This classification is
based on the type of classification like rule based,
4. METHODOLOGY tree based, function based, fuzzy rule based and so
on.
Data pre-processing is essential for successful data
mining process [4]. Feature selection is one of the Decision trees are powerful classification
important and frequently used techniques in data algorithms that are becoming increasingly more
preprocessing for data mining [3]. This process popular with the growth of data mining in the field
reduces the number of features, removes irrelevant, of information systems. This technique recursively
redundant or noisy data thereby improving mining separates observations in branches to construct a
performance such as predictive accuracy and result tree for the purpose of improving the prediction
comprehensibility. The removal of irrelevant and accuracy. In doing so, they use mathematical
redundant information often improves the algorithms (e.g., information gain, Gini index, and
performance of the classification algorithms [16]. Chi-squared test) to identify a variable and
The feature selection process as shown in fig 1 corresponding threshold for the variable that splits
consists of subset generation, subset evaluation and the input observation into two or more subgroups.
result validation. This step is repeated at each leaf node until the
complete tree is constructed. The objective of the
Feature selection algorithms broadly are splitting algorithm is to find a variable-threshold
categorized into a) filter model (univariate, pair that maximizes the homogeneity (order) of the
multivariate) b) wrapper model and c) embedded resulting two or more subgroups of samples. In our
model. Filter model evaluates and selects the subset study we have used J48 (implementation of C4.5 in
from the data without involving any data mining Weka). C4.5 [34] is a software extension of the
algorithm. Wrapper model searches for features basic ID3 algorithm designed by Quinlan.
that best suites the predetermined mining
algorithm; it is computationally expensive than the C4.5 is a supervised learning algorithm. The
filter model. Embedded model is a combination of algorithm analyzes the training set and builds a
the above two models like C4.5 algorithm. classifier that must be able to correctly classify
both training and test examples. The classifier used
In our study multivariate filter based model CFS by C4.5 is a decision tree.
(Correlation based feature selection) and embedded
model C4.5 have been applied on the cancer data. Algorithm: C4.5
The purpose of applying these models is they
model the dependencies among the features, which 1. Check for the base case
is the basic advantage of feature selection. Cross
validation is method for estimating the true error of 2. Find the attribute with highest information
a model. Cross validation is used to evaluate or gain (A-best)
compare learning algorithms as follows: in each
iteration one or more learning algorithms use k -1 3. Partition S into S1, S2, .. according to the
folds of data to learn one or more models, and values of A-best
subsequently the learned models are asked to make
4. Repeat the steps for S1, S2, ..
predictions about the data in the validation fold.

Staging Prediction in Cervical Cancer Patients – A Machine Learning Approach 17

International Journal of Innovative Research and Practices Vol.2, Issue 2, February 2014
ISSN 2321-2926

The base cases are as follows: Bayesian classifiers assign the most likely class to a
given example described by its feature vector.
• All the examples from the training set Learning such classifiers can be greatly simplified
belong to the same class by assuming that features are independent given
class, that is, P(X|C) = П P(Xi | C), where X = (X1,
• The training set is empty X2, … Xn ) is a feature vector and C is a
class. Despite this unrealistic assumption, the
• The attribute list is empty resulting classifier known as Naive Bayes is
remarkably successful in practice, often competing
Rule-based expert systems are often applied to with much more sophisticated techniques. Naive
classification problems in various application Bayes has proven effective in many practical
fields, like fault detection, biology, and medicine. applications, including text classification, medical
Fuzzy logic can improve such classification and diagnosis, and systems performance management.
decision support systems by using fuzzy sets to
define overlapping class definitions. The 5. EXPERIMENTS AND RESULTS
application of fuzzy if-then rules also improves the
interpretability of the results and provides more The dataset consists of 203 cervical cancer patient
insight into the classifier structure and decision cases (Courtesy: Mitra et al., 2000). It consists of
making process. In our study we have applied NN 21 boolean features that indicate the signs and
and Fuzzy RoughNN [28] as these are applied in symptoms observed upon physical examination
the literature. containing the 4 stages of the cancer. The 21
Boolean input features refer to Vulva: healthy
Support vector machines are supervised learning (Vu(h)), Vulva: lesioned (Vu(l)) , Vagina: healthy
models with associated learning algorithms that (Va(h)), Vagina:spread to upper part (Va(u)),
analyze data and recognize patterns, used for Vagina: spread to middle part (Va(m)), Vagina:
classification and regression analysis. The basic spread to lower part (Va(l)), Cervix:healthy
SVM takes a set of input data and predicts, for each (Cx(h)), Cervix: eroded (Cx(e)), Cervix: small
given input, which of two possible classes forms ulcer (Cx(su)), Cervix: ulcerative growth (Cx(u)),
the output, making it a non-probabilistic binary Cervix: proliferative growth (Cx(p)), Cervix:
linear classifier. Given a set of training examples, ulcero-proliferative growth (Cx(l)), Paracervix:
each marked as belonging to one of two categories, free (PCx(f)), Paracervix: infiltrated (PCx(i)),
an SVM training algorithm builds a model that Urinary bladder base: soft (BB(s)), Urinary
assigns new examples into one category or the bladder base: hard (BB(h)), Rectrovaginal septum:
other. In addition to performing linear free (RVS(f)), Rectrovaginal septum: infiltrated
classification, SVMs can efficiently perform non- (RVS(i)), Parametrium: free (Para(f)),
linear classification using what is called the kernel Parametrium: spread, but not upto (Para(nu)) and
trick, implicitly mapping their inputs into high- Parametrium: spread upto (Para(u)), respectively.
dimensional feature spaces. Staging of cervical cancer is given by FIGO.

A multilayer perceptron (MLP) is a feed forward The purpose of this study is to identify the
artificial neural network model that maps sets of attributes and extract rules for easily identifying the
input data onto a set of appropriate output. An MLP stage based on the signs and symptoms to identify
consists of multiple layers of nodes in a directed the right treatment.
graph, with each layer fully connected to the next
one. Except for the input nodes, each node is a Accuracy of C4.5 is better as compared to the
neuron (or processing element) with a nonlinear accuracy achieved in [31]. Accuracies achieved by
activation function. MLP utilizes a supervised Mitra are 81.5 and 80.2 on training and test data
learning technique called back propagation for respectively whereas the accuracies obtained in this
training the network. MLP is a modification of the paper are 93.06 and 90.19 for training and test data
standard linear perception and can distinguish data respectively after applying Correlation based
that is not linearly separable. Feature Selection (Cfs). Among the algorithms
compared C4.5 (J48 is its implementation in Weka)
has outperformed as seen in table 1.

D. Sowjanya Latha, P.V. Lakshmi and Sameen Fathima 18

International Journal of Innovative Research and Practices Vol.2, Issue 2, February 2014
ISSN 2321-2926

Precision is a measure of the accuracy provided generated give a picture of the most influential
that a specific class has been predicted. Recall or attributes for identifying the stage of the cancer.
Sensitivity is a measure of the ability of a
prediction model to select instances of a certain 6. CONCLUSION
class from a dataset. It corresponds to the true
positive rate. Specificity is a measure commonly The comparative study of multiple classifiers
used in two class problems where one is more identifying the stage of cervical cancer using a
interested in a particular class. It corresponds to the dataset of size 221 records provided us with an
true-negative rate. insight into the predictive ability of different data
mining methods. Accuracy achieved by J48
Sensitivity is computed using the following algorithm is better than any given in the literature.
formula (TP/(TP+FN)) where TP is True Positive, Sensitivity and specificity analysis on these
FN is False Negative. Specificity is computed algorithms provided us with the prioritized
using the following formula (TN/(FP+TN)) where importance of the prognostic factors that lead to the
TN is True Negative, FP is False Positive, TN is staging of the cancer. This analysis was not
True Negative. performed in any given in the literature. Data
analysis was done using 10-fold cross validation.
In this paper sensitivity and specificity values have We can conclude saying that by applying data
been computed for all these algorithms as shown in mining algorithms the invaluable efforts of the
table 2. J48 has values greater than 0.9 for both medical professionals can be enhanced to save
sensitivity and specificity. The results show that more human lives by giving proper treatment at the
J48 is able to diagnose all the stages of the cancer. right time.

The results shown in table 3 project the better ACKNOWLEDGMENTS

performance of C4.5 algorithm with feature
selection in the working paper compared to the We would like to thank Dr Kodati Vijaya Lakshmi,
other similar studies. C4.5 algorithm is efficient, Vasavi Hospital, Dr Sunil and Dr Kameswari,
the rules generated are easier to understand. A Fernandez Hospital, Dr Benjamin and Dr Usha,
major disadvantage that has been identified in this MNJC Hospital, Dr Sudarshan and Dr Sushma,
paper is that the rules generated by C4.5 are clear Indo-American Cancer Hospital who have helped
and precise but are not sufficient enough for in gathering information about the disease and to
predicting the stages of the cancer. A further study interact with a few patients. Special thanks to Dr
needs to be made to enhance the performance of the Sushmita Mitra and Prof Pabitra Mitra, Indian
algorithm. This would be the scope of this paper. Statistical Institute, Calcutta for sharing the data.

The rules generated by J48 algorithm are as REFERENCES:

follows:
Ambika Satija, Cervical Cancer in India, South
Asia Centre for Chronic Disease.
If {Para(u)=Y and BB(s)}=N then stage = IV
American College of Obstetricians and
If {Para(u) = Y and BB(s)}=Y then stage = III Gynecologists. ACOG Practice Bulletin No. 99:
Management of abnormal cervical cytology and
If {Para(u)= N and (Para(f)= N) or histology. Obstet Gynecol. 2008;112:1419-1444.

H. Liu and H. Motoda, Feature Extraction,

(Para(f) = Y and Va(u) = Y)} then stage = II
Construction and Selection: A Data Mining
Perspective, eds. Boston: Kluwer Academic, 1998,
If {Para(u) = Y and Va(u) = N} then stage = I second printing, 2001.

Huan Liu, Lei Yu, Toward Integrating Feature

The decision tree generated by J48 algorithm is Selection Algorithms for Classification and
shown in figure 2. The decision tree and the rules

Staging Prediction in Cervical Cancer Patients – A Machine Learning Approach 19

International Journal of Innovative Research and Practices Vol.2, Issue 2, February 2014
ISSN 2321-2926

Clustering, IEEE Transactions on Knowledge and countries, Asian Pacific Journal of Cancer
Data Engineering, Vol 17, No 4, April 2005 Prevention, 7(1) (Jan-March 2006): 36-40.

Hayit Greenspan etal., Automatic Detection of Lei Yu, Huan Liu, Feature Selection for High-
Anatomical Landmarks in Uterine Cervix Images, Dimensional Data: A Fast Correlation-Based
IEEE Transactions on Medical Imaging, Vol 28, Filter Solution, Proceedings of twentieth
No. 3, March 200 International Conference on Machine Learning
(ICML), 2003
J Han, M Kamber, Data Mining: Concepts and
Techniques, Elsevier, Second Edition, 2006 Margaret H Dunham, S Sridhar, Data Mining:
Introductory and Advanced Topics, Pearson
Jyotsna A Saonere, Awareness screening Education, 2006
programme reduces the risk of cervical cancer in
women, African Journal of Pharmacy and Martin-Hirsch PPL, Paraskevaidis E, Bryant A,
Pharmacology, Vol. 4(6), pp 314-323, June 2010 Dickinson HO, Keep SL. Surgery for cervical
intraepithelial neoplasia. Cochrane Database of
Jiayong Zhang, Yanxi Liu, Cervical Cancer Systematic Reviews 2010, Issue 6. Art. No.:
detection using SVM based Feature Screening, CD001318.
Carnegie Mellon University, NIH program
Mate Ongenaert et al., Discovery of DNA
Kavishvar, Kathy, Michael et. al., Clinical methylation markers in cervical cancer using
Decision support with automated text processing relaxation ranking, BMC Medical Genomics,
for cervical cancer screening, Biomedical Journal, November 2008
2012. Available online at
[Link]/content/early/recent. Mohd Otham, Thomas Yau, Comparision of
different Classification Techniques using WEKA
Kaarthigeyan K, Cervical Cancer in India and for Breast Cancer, Biomed 06, IFMBE
HPV Vacination,Indian J Med Paediatr Oncol Proceedings 15, pp 520-523, 2007
2012, 33:7-12
Nester Jeyakumar M et al., Improved Classifier
Katz VL, Lentz GM, Lobo RA, Gershenson DM,, performance through genetic algorithm for cervical
Noller KL. Intraepithelial neoplasia of the lower cancer prediction, Journal of Research in
genital tract (cervix, vulva): Etiology, screening, Bioinformatics, Apr 2012.
diagnostic techniques, management. In: eds.
Comprehensive Gynecology. 5th ed. Philadelphia, Pavani Sowjanya et al., Prevalence and
Pa: Mosby Elsevier; 2007:chap 28. distribution of high-risk human papilloma virus
(HPV) types in invasive squamous cell carcinoma
Krishnakumar Duraisamy et al., Methods of of the cervix and in normal women in Andhra
Detecting Cervical Cancer, Advances in Biological Pradesh, India, BMC Infectious Diseases 2005, 5:
Research 5(4), 2011, pp 226-232, ISSN 1992-0067 116

Kuzhali, Rajendran, et al, Feature Selection Pieman, Wahid, Azuraliza, A Comparative study
Algorithm Using Fuzzy Rough Sets for Predicting for various methods of classification, 2012
Cervical Cancer Risks, Modern Applied Science, International conference on Information and
Vol 4 No 8, August 2010 Computer Networks, IPCSIT vol. 27(2012)

Kiam and Zhang, HPV Risk Type Classification Polidais LLC, Medical Imaging in Cancer care:
from Protein Sequences using SVM, LCNS 3907, Charting the progress Executive Summary, New
pp 57-66, 2006 England Journal of Medicine, 342.1, 2000, pg 42-
49
Kurkure AP, and Yeole BB, Social inequalities in
cancer with special reference to South Asian Rakesh Dikshit, et. al, Cancer mortality in India: a
nationally representative survey, [Document

D. Sowjanya Latha, P.V. Lakshmi and Sameen Fathima 20

International Journal of Innovative Research and Practices Vol.2, Issue 2, February 2014
ISSN 2321-2926

online], Available from [Link] [Link]

March 2012, DOI:10.1016/S0140-6736(12)60358- [Link]
4
Cervical Cancer Risk Factors [Document-online]
R Prasad, Cancer Killed 5.56 lakh in India in 2010, Available from
The Hindu, Chennai, March 28, 2012. [Link]
iledGuide/cervical-cancer-risk-factors
Ross Quinlan J., Machine Learning, 1st ed. Morgan
Kaufmann Publishers Inc., 1993 A Jena, [Link]., Parametrial invasion in carcinoma
of cervix: role of MRI measured tumour volume,
Bustanur Rosidi, [Link]., Calssification of Cervical The British Journal of Radiology, 78 (2005), 1075
Cells Based on Labeled Colour Intensity – 1077
Distribution, International Journal of Biology and
Biomedical Engineering, 2011 [Link]
004359
Seung Hee Hoo, Sun Ha Jee, Jong Eun, Jong Sup,
Analysis on risk factors for cervical cancer using Authors
Induction technique, Expert Systems with
applications, Elsevier, 2004 Ms D Sowjanya Latha is an Associate Professor at
AMS School of Informatics, Hyderabad, India. She
Shravya Reddy Konda, A Comparative evaluation is MCA with 14 years of experience. She has 3
of Symbolic Learning Methods and Neural publications to her credit. She is currently pursuing
Learning Methods, University of Maryland her PhD (CSE) from GITAM University. Her areas
of research are Data Mining, Bioinformatics,
Sushmita Mitra, Pabitra Mitra, Staging of Cervical Network Security.
Cancer with Soft Computing, IEEE Transactions on
Biomedical Engineering, Vol 47, No 7, July 2000 Dr PV Lakshmi is Professor & Head, Dept of IT,
GITAM University, Visakhapatnam, India. She is
Thangavel, Jagannathan, Easmi, Data Mining PhD (CSE) and MTech (CSE) from Andhra
Approach to Cervical Cancer Patients Analysis University, Visakhapatnam, India. She has around
using Clustering Technique, Asian Journal of 14 papers published in International Journals, 2
Information Technology, 5(4), 2006 papers in National journals, 8 in International
conferences, 2 in National conferences. She has
Thales Senh Korting, C4.5 algorithm and been awarded as the Best Academician in 2010
multivariate decision trees, Image Processing from CSE Dept, GITAM University. She has
Division, National Institute for Space Research, around 16 years of teaching experience. Her areas
Brazil of research are Bioinformatics, Cryptography,
Network Security.
Veronica S Moertinin, Towards the use of C4.5
algorithm for classifying banking dataset, Integral Dr Sameen Fatima is Professor & Head, Dept of
Vol 8 No 2, October 2003 CSE, Osmania University, Hyderabad, India. She is
PhD (CSE) from Osmania University, Hyderabad,
Zhang Liu, Tong Zhao, Yanxi Liu, SVM Based India, MS (CSE) from University of
Feature Screening applied to Hierarchical Massacheusetts, USA, MPhil (Computer Methods),
Cervical Cancer Detection, Carnegie Mellon University of Hyderabad, She has around 14
University, NIH program publications to her credit. Her areas of research are
Information Retrieval System, Artificial
WEKA (Data Mining Software). Available at Intelligence, Machine Learning.
[Link] 2006.

Cervical cancer incidence [Document-online]

Available from

Staging Prediction in Cervical Cancer Patients – A Machine Learning Approach 21

International Journal of Innovative Research and Practices Vol.2, Issue 2, February 2014
ISSN 2321-2926

Original Subset
Subset Generation Subset Evaluation
Set

Goodness
of subset

No Stopping Yes
Result Validation
Criterion

Figure 1: Unified process of feature selection

Table 1: Performance analysis using accuracy on data mining algorithms

Accuracy
Train Test Set Full Set Full Set (10-fold cross 10-fold cross validation
Set (10-fold validation without with Wrapper and Subset
cross CFS) Evaluator
validation
with CFS)
J48 93.069 90.196 87.192 87.192 87.192
SVM 86.418 92.156 80.295 81.773 85.221
FuzzyRoughNN 92.079 100 61.0837 66.5025 67.487
NN 89.108 98.039 80.295 84.236 85.714
NaiveBayes 93.069 92.156 85.221 84.729 87.684
MLP 95.049 94.117 82.758 78.325 87.192

Table 2(A): Comparison of the Performance of mining algorithms

Sensitivity Specificity
J48 SV Fuzzy NN Naïv MLP J48 SVM Fuzzy NN Naïv MLP
M Roug eBay Roug eBay
hNN es hNN es
Stage I 0.85 0.14 1 0.5 0.57 0.571 0.96 0.98 0.893 0.97 0.97 0.96
7 3 1 8 3 1 3 8
Stage II 0.71 0.78 0.421 0.63 0.73 0.711 0.96 0.9 0.775 0.92 0.94 0.93
0 9 2 7 0 4 0 5
Stage III 0.92 0.94 0.638 0.97 0.95 0.926 0.92 0.77 0.857 0.67 0.83 0.84
6 1 1 6 2 6 1 5 9
Stage IV 0.8 0.2 0.066 0 0.46 0.467 0.96 0.97 0.990 0.99 0.97 0.96
7 3 3 4 8 2

D. Sowjanya Latha, P.V. Lakshmi and Sameen Fathima 22

International Journal of Innovative Research and Practices Vol.2, Issue 2, February 2014
ISSN 2321-2926

Table 2(B): Comparison of the Performance of mining algorithms

AUC
J48 SVM Fuzzy NN Naïv MLP
Roug eBay
hNN es
Stage I 0.95 0.56 0.95 0.96 0.96 0.95
Stage II 0.86 0.84 0.74 0.86 0.88 0.85
Stage III 0.90 0.85 0.90 0.93 0.92 0.91
Stage IV 0.89 0.58 0.87 0.87 0.81 0.87

Table 3: Comparison with other similar studies

Methods Used Sensitivit Specificit Accurac

y y y
[31] Pabitra, 2001 (Staging RoughSet Theory 81.03
data, clinical) ID3 82.74
C4.5 80.2
[42] Jena, 2005 (MR images 59.26 61.54 60.95
of cervical cancer)
[29] Seung Hee Ho, [Link]., CHAID 64.00 77.83 72.96
2004 (demographic, Logostic Regression 40.80 88.70 71.83
environmental and genetic
factors)
[28] Bustanur Rosidi, [Link]., Labeled Color Intensity 78.6 75.9 77.0
2011 (Cervix cells) Distribution
Working Paper (Staging data, C4.5 (with feature 82 95 87.19
clinical) selection) 51 91 80.29
SVM 53 88 61.08
FuzzyRoughNN 52 89 80.29
NN 68 93 85.22
NaiveBayes 67 93 82.75
MLP

Staging Prediction in Cervical Cancer Patients – A Machine Learning Approach 23