Lung Cancer Prognosis System using Data Mining Techniques

Yomna Omar, Abdullah Tasleem, Michel Pasquier and Assim Sagahyroon

American University of Sharjah, Department of Computer Science & Engineering

PO Box 26666, Sharjah, U.A.E.

Keywords: Intelligent Healthcare Tool, Lung Cancer Prognosis, Real-World Data Mining, WEKA.

Abstract: This paper describes a Lung Cancer Prognosis System (LCPS) that aims at providing oncologists with an

accurate estimate of the health status of their patients. The proposed system is born from two observations:

First, lots of efforts are still required in healthcare to improve productivity, accuracy, etc. by providing ad hoc

computer-based solutions; second, while increasing popular, AI and data mining tools cannot be used without

significant training and expertise. LCPS thus aims at providing the former by integrating the latter into a user-

friendly tool, supplementing the knowledge of the expert oncologist with information about their patients, and

leading to improved patient care and treatments. LCPS can accept a variety of lung cancer datasets and

employs several data mining algorithms to uncover relationships between observed health signs and probable

outcomes, and provides oncologists with various statistical results including predictions about their patients’

medical future. Furthermore, LCPS makes it easy to manage patients’ records, allows them view their profiles

and any information as deemed suitable by their doctor, including prognosis and other comments. Lastly,

while the current application is currently limited to lung cancer treatment, it can be considered a prototype

that can be adapted to other diseases.

1 INTRODUCTION

According to scientific experts, cancer remains one of

the leading causes of death in adults and the “number

one disease killer of children”, and half of them are

preventable (American Cancer Society, 2017). Lung

cancer has the highest rate of deaths, despite the fact

that its prevention and treatment in its early stages is

simple and hassle-free. The problem is that early

symptoms are often mistaken for cough or flu, thus

making early-stage lung cancer difficult to diagnose.

Since the initial symptoms of lung cancers can take

years to develop, most patients discover their illness

at an advanced stage where cancer cells are

widespread, making it much harder to cure (American

Cancer Society, 2016). Lung cancer is also one of the

top five fatal diseases in the UAE (Khaleej Times,

2016) and, according to Dr. Ali Al Dameh, Oncology

Consultant from Tawam John Hopkins Hospital, “the

disease is in dire need of being addressed with a solid

plan of action” (WAM, 2015). That goal includes the

two main types of lung cancer i.e., non-small cell lung

cancer and small cell lung cancer.

In order to decrease preventable cases of cancer-

related deaths and improve the overall health system,

a more versatile and powerful technology is needed.

Precisely AI and data mining techniques can process

health data to better forecast people’s future health

conditions and potential risks. But advanced AI tools

cannot be used without significant training and

expertise, which medical practitioners do not possess.

Therefore, we developed a prototype of a Lung

Cancer Prognosis System (LCPS) that aims at making

those techniques accessible in a user-friendly manner

in order to augment the knowledge of the expert

oncologist and improve patient care and treatments.

The proposed system is trained beforehand on a

hospital dataset characteristic of the target population

and can subsequently provide better information and

prognosis specific to each lung cancer patient based

on their vital health signs and test results. LCPS

employs several statistical and data mining tools and

presents their results in a visual, intuitive manner, for

the benefit of both the physicians and their patients.

2 EXISTING SOLUTIONS

GenieMD combines IBM’s Watson with research

from Harvard Medical School to provide users with

accurate self-triage. It records medical history and

can connect to sensors for real-time vitals tracking

Omar, Y., Tasleem, A., Pasquier, M. and Sagahyroon, A.

Lung Cancer Prognosis System using Data Mining Techniques.

DOI: 10.5220/0006553703610368

In Proceedings of the 11th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2018) - Volume 5: HEALTHINF, pages 361-368

ISBN: 978-989-758-281-3

361

(HMW, 2013), but cannot upload other datasets.

WebMD allows users to search through a medical

database to read articles, learn about diagnosis, and

contact medical doctors (WebMD, 2017). However,

this service does not process actual patient data.

Sehhaty is an app that allows Dubai residents to view

their records at the Dubai Health Authority, including

prescriptions, appointments, test results, medical

advice, and so on (Khaleej Times, 2014). Similarly,

My Medical allows users to create a record of their

medical history, enter vital sign readings from health

trackers, and send the data to a doctor. While tailored

for the UAE, these apps do not provide any kind of

diagnosis or prognosis tools.

Lungscreen uses a short lifestyle questionnaire to

assess whether the person has lung cancer symptoms

or not (Mesko, 2015), and gives a prediction for the

next 10-20 years. There is no clear way to assess

prediction accuracy, and it does not use actual patient

data. The Lung Cancer Foundation App is a mobile

application that helps with the self-diagnosis of lung

cancer and provides educational videos and

information (OFWW, 2011). Finally, Cancer Spotter

aims to encourage earlier diagnosis by getting people

to learn and understand the symptoms of cancer, and

to seek help from a doctor as soon as possible (Park,

2001). While it is simple and promotes cancer

awareness, it relies on self-assessment and has no

diagnosis or prognosis.

In sum, most available tools provide lung-cancer

related information and some allow self-assessment,

but very few offer any form of medically validated,

automated diagnosis or prognosis, based on actual

lung cancer data and the patient’s vital signs, let alone

tailored to a specific country or population.

3 PROPOSED SOLUTION: LCPS

Based on the above assessment, we have developed a

software system to provide users with a consistent

and accurate health diagnosis at low cost, therefore

facilitating the prognosis of fatal lung cancer. We

plan to make use of large databases of lung cancer

patients from hospitals all around the world and apply

AI and data mining tools to derive from these datasets

diagnosis and prognosis models that can be

subsequently applied to new patients. Another goal is

to integrate all computational techniques and

automate the process as much as possible to offer a

tool that is user-friendly for both doctors and their

patients. Initially, our system was designed to use a

patient database from the Tawam Hospital in UAE,

then it was extended to handle other datasets. In this

paper, the use cases and results presented employ

lung cancer datasets publicly available on the Web.

LCPS is a web-based app that uses JSF for

functionality and HTML/CSS for webpages and UI.

This first version integrates three machine learning

algorithms from WEKA. It provides oncologists and

patients with different access rights. Patients can view

their data, health predictions, and comments from the

oncologist. Doctors manage patients, can run various

tests on the dataset and new patients data. The LCPS

prototype described in this paper provides doctors

with a number of public datasets for demonstration

and training purposes: one is used to predict patient

life expectancy, another to predict whether or not the

patient has a cancer gene. LCPS allows doctors to

load and edit datasets and patient records, to view

statistical indicators, and to apply various algorithms

on patient data. It automatically analyses a dataset to

determine the the most suitable machine learning

algorithm and presents the results to the physician

through many tables and plots.

4 TECHNICAL APPROACH

LCPS is a 3-tiered web application, with a back-end

SQL database, an application layer, and a front-end

user interface. LCPS will make use of several data

mining and machine learning algorithms that have

been trained with patient data to reach conclusions

about the patient’s health within a certain level of

accuracy. LCPS can also use other datasets to allow

flexibility in predicting several lung cancer markers.

By creating several XHTML pages that are styled

by Cascading Style Sheets (CSS), the web app’s main

functions lie in JSF code with baking beans to

implement UI elements, display options, and access

the database. Pages that contain graphs use the

Google Charts API via embedded JavaScript code,

more specifically jQuery. The app’s database access

mainly occurs in baking beans using SQL statements

to insert, update, and delete data. WEKA offers

multiple ways to integrate their machine learning

algorithms using Java. For our purposes, we initially

focused on three algorithms: J48 decision tree, Naïve

Bayes (NB), and K-Nearest Neighbor (KNN). Further

algorithms will be added later on.

5 SYSTEM ARCHITECTURE

LCPS architecture consists of three layers for

presentation, application, and data. This helps

HEALTHINF 2018 - 11th International Conference on Health Informatics

362

developers accommodate future expansions such as

different subcomponents in the data layer (Oracle

database, MySQL, Firebase) or even in the

presentation layer (Java GUI, C# GUI, Python GUI).

Importantly, it minimizes dependencies between

components, enhances security (as the user has no

direct access to the database) and promotes scalability

which is a key aspect in a cloud deployed app. Lastly,

it allows the three tiers to be coded and tested in

parallel, and therefore reduces development and

maintenance costs over the long term.

Figure 1: LCPS system multi-layered architecture.

The presentation layer consists of either command

line or GUI, which are used to get user inputs and

provide information and feedback. The presentation

layer consists of a web based interface programmed

using HTML and CSS. The application layer consists

of the decision logic i.e., the preprocessing and

machine learning algorithms from WEKA that will

mine the dataset selected by the oncologist and

evaluate new patient data accordingly. The

application uses the SQL database on Apache Derby

to store its data. The tables and schema can be

managed via the interface, or programmatically. The

node architecture can be displayed through the use of

a deployment diagram, which captures the system

hardware’s topology. Each layer is deployed within a

different hardware component and the presentation

layer is mainly deployed via the web interface.

Since one distinguishing feature of LCPS is to

provide automated data analysis and prognosis, the

choice of data mining component has been subject to

extensive study (Witten, 2016). Moreover, the need

to offer a user-friendly tool that hides technical

complexity and presents results in an intuitive, visual

manner, means favoring models based on rules or

decision trees for instance, and providing adequate

visualizations via tables and plots. As such many

algorithms are available, as well as several data

mining platforms, that we experimented with.

6 DATA MINING TOOLS

We started with KEEL (Knowledge Extraction based

on Evolutionary Learning), an open source data

mining framework that is increasingly being used in

healthcare. KEEL includes many algorithms for

regression, classification, clustering, and more. It

allows creating experiments using multiple datasets

and algorithms, independently scripted from the user

interface. However, we didn’t find the UI user-

friendly enough, and KEEL lacks good visualization

techniques (Rangra, 2014).

KNIME (Konstanz Information Miner) is

another open source tool based on data analytics,

reporting and integrating platform, and has been used

extensively in pharmaceutical research. It has a

modular design and offers interactive execution and

data pipelines. It covers all major areas of data

analysis but has limited error measurement methods

and does not support techniques such as wrapper

methods for descriptor selection (Rangra, 2014).

Eventually we selected the open source WEKA

framework, which has all the desired data processing

and machine learning algorithms and is widely used

for research and development. It has a great user

interface and provides excellent visualization tools to

help understand the models (Witten, 2016).

7 SYSTEM FUNCTIONALITY

The user interface was initially modeled using Adobe

Dreamweaver through the use of HTML and CSS.

Once tested and approved by users, the design was

transferred to the development environment, where

core functions were added in. LCPS employs several

web interface mechanisms, as depicted in the figures

hereafter, to enhance usability.

Upon login and redirection to their home page, an

oncologist can view all his/her respective patients,

further analyze their medical profile, add comments,

deciding for each whether it should be visible to the

patient or not. The oncologist can also compare the

patient’s health data against a preloaded dataset or

one of his/her choosing. The doctor can also edit

existing datasets and view previous versions. When

the physician compares his/her patient’s data against

a chosen dataset using selected algorithms, he/she

will be provided with a prediction of the patient’s

health status. The physician can also compare results

from different algorithms in order to gain a better

understanding of the data and the various models.

When a cancer patient logs into LCPS, he/she is

redirected to his/her main page where he/she can view

Lung Cancer Prognosis System using Data Mining Techniques

363

Figure 2: Sample LCPS information page.

all the tests conducted on him/her by his/her

respective oncologist. The patient can also view any

comments left by the physician on his/her health

status, as well as data plots and results as included.

8 DATA MINING ALGORITHMS

WEKA features hundreds of algorithms for data

processing, feature selection, clustering, finding

association rules, classification, etc. (Witten, 2016).

LCPS integrates selected methods for removing

outliers and irrelevant features, and predicting the

patients’ health status using classification rules,

decision trees, instance based learning, probabilistic

approach, and later regression trees. Since choosing

the “best” predictive algorithm for a given dataset

requires a lot of computations, automation is needed.

This in turns requires making assumptions, which is

another reason why the current system supports

databases for lung cancer only.

Including decision trees in LCPS was an obvious

choice, and currently the J48 algorithm is used

(Witten, 2016). The decision tree generated shows the

most relevant input variables for prediction, that are

determined using information gain. As an illustrative

example, we chose a dataset obtained by Harvard

University (Zhu, 2007), that contains 189 instances

with all numeric values except for the class attribute

(ADEN, SQUA, COID, SCLS, NORMAL). It

includes different specimens of lung cancer using

various gene selection tests and classified as tumors.

Accuracy results and Confusion Matrix are shown

and explained to the doctor. The latter table further

describes the performance of the classification model.

In this example, the user is told that 131 out of the 189

instances were correctly classified as

adenocarcinoma (ADEN), 18 as squamous cell lung

Figure 3: Classifier output of WEKA’s J4.8 algorithm.

carcinomas (SQUA), 20 as pulmonary carcinoids

(COID), 4 as small-cell lung carcinomas (SCLC) and

that 16 instances were normal lung patients. Other

values outside the diagonal show exactly how some

instances were incorrectly classified. The data can be

examined by the oncologist, if desired, to try and

analyze the cause of the error, which could be due to

some data entry mistake, some noise in

measurements, or simply some borderline case that

was too difficult for the classifier to figure out.

Furthermore, the Detailed Accuracy by Class shows

that the TP rate of true positives (correctly classified

instances) has a high average of 0.931 while the FP

rate of false positives (falsely classified instances)

had a very low average of 0.058. Such results are

pointed out to the doctor so he/she can appreciate

diagnosis or prognosis accuracy, possibly examine

the data to further his/her understanding. In this

example, J48 achieved a 93% accuracy and new

patients can be evaluated with high confidence. The

decision tree in Fig. 4 is shown to the doctor.

Another popular classifier is Naïve Bayes (NB),

which realizes probabilistic inference based on

assumptions of conditional independence between

predictors (Witten, 2016). One of the reasons we

chose to include it is that, despite its simplicity, it

often outperforms more sophisticated classifiers.

Another is that it can be easily explained to the doctor.

Naïve Bayes is essentially a quick method for

composition of statistical predictive models, that

analyses the subjection between attribute values and

classes to derive a conditional probability. Ranking

HEALTHINF 2018 - 11th International Conference on Health Informatics

364

Figure 4: Generated decision tree for the J4.8 algorithm.

class probabilities then gives an indication of which

result is most likely. Using the same lung cancer

dataset, we can observe a loss in the TP rate and a

gain in the FP rate as shown in Fig. 5.

Figure 5: Classifier output of Naïve Bayes algorithm.

Instance-based learning is a form of lazy learning

where the model memorizes data instances and

classification relies on measuring proximity (Witten,

2016). To address the issue of noisy data, it is

recommended to examine multiple nearest neighbors,

hence LCPS integrates the classic K-Nearest

Neighbor algorithm. Results using the same lung

cancer dataset are shown in Fig. 6 for the case of a

single neighbor (K=1). However, LCPS will

automatically try different values of K and retain the

optimal one. Explanations are given to the oncologist

who wishes to understand the significance of the

results. For instance, if the number of correctly

classified instances increases with K, it means that the

dataset is noisy, which in turn may require

examination. In this case results show that the dataset

is noise-free.

LCPS can display only the best results (achieved by

the best algorithm), or show all results, as well as a

detailed comparison, in tabular format. In this case, it

happens that J48 performs best on the sample lung

cancer dataset. There might be various reasons for

this, such as for instance, the fact that irrelevant gene

predictors have been pruned out, or because it can

process both nominal and numerical attributes better

than Naïve Bayes and KNN, or because decision trees

handle missing values well. It is important to note that

there is no classification algorithm that can perform

best for all available datasets. Hence LCPS will apply

all selected methods and build models anew for each

dataset.

Figure 6: Classifier output of the KNN algorithm (K=1).

9 DATASETS AND RESULTS

Since LCPS can handle generic lung cancer datasets,

we chose two databases available online for testing

and illustration purposes. The Genes database used to

predict lung cancer tumors based on the patient’s

genes was cleaned and used at Harvard University,

hence we knew it was mostly error free and would

give reliable results, making for a good tutorial. The

database consists of 12600 attributes and 203

instances, all numerical values except for the class. In

2001, lung carcinoma was the leading cause of cancer

Lung Cancer Prognosis System using Data Mining Techniques

365

death hence this dataset was created and exploited to

define distinct subclasses of lung adenocarcinoma

(Bhattacharjee, 2001). The dataset can be used by

doctors so long as they have a gene expression

profile, and LCPS will determine if the patient has the

genes to develop either cancer type.

Figure 7: Analysis of Veterans Admin. Lung Cancer Data

using a generalized F-regression Model (Venables, 2002).

The second dataset included for testing is the

Veteran’s Administration Lung Cancer Trial dataset

used by Venables and Ripley in 2002 to predict the

life expectancy of a patient with specific lung cancer

tumors (Venables, 2002). It was created using

questionnaires from doctors and patients. Initially, a

Generalized F-regression model was used to analyze

it, as per Fig. 7. Data collection and comparison of

statistical models was first done manually, then the

Cox model was applied to establish the baseline

hazards, as the dataset had several covariates.

The second dataset included for testing is the

Veteran’s Administration Lung Cancer Trial dataset

used by Venables and Ripley in 2002 to predict the

life expectancy of a patient with specific lung cancer

tumors (Venables, 2002). It was created using

questionnaires from doctors and patients. Initially, a

Generalized F-regression model was used to analyze

it, as per Fig. 7. Data collection and comparison of

statistical models was first done manually, then the

Cox model was applied to establish the baseline

hazards, as the dataset had several covariates.

The dataset obtained needed cleaning prior use:

outliers were removed and nonsensical or missing

values were replaced with mean values, irrelevant

attributes were discarded. In the end the dataset had

seven attributes: Treatment (standard or test), Age (in

years), Karnofsky Score (0-100), Time since

diagnosis (in months); Cell type (one of four types);

Prior therapy; Survival time (in days, later discretized

into 3 bins). The Karnofsky score is one of the

methods used to quantify a patient’s general health

and physical activities, ranging from 0 (dead) to 100

(perfect health). Results are illustrated in the

screenshots of the LCPS user interface hereafter.

Figure 8: Cumulative hazard functions for the cell types;

left graph is labelled by survival probability on log scale

and right graph is on log-log scale (Venables, 2002).

As mentioned the decision logic layer of our

system heavily relies on WEKA and its excellent Java

API, which offers many classes and methods,

specifically: to store all data instances, to preprocess

the data (discretize, sparse to non-sparse, and attribute

selection), to predict the class outcome (J48, Naïve

Bayes, KNN), to evaluate accuracy (statistical results,

TP/FP, ROC). To integrate WEKA in our system we

had to use beans to get information from our XHTML

pages and process the data using the provided

libraries.

Figure 9: LCPS dataset upload page.

Once logged in, a doctor can select a patient and

go to the Analysis page where he/she can upload any

lung cancer related dataset to the system, as depicted

in Fig. 9. Since a consistent, generic dataset format is

required, the doctor is able to view and download the

data template. Converters are available as well that

help with the formatting. Next, the doctor can view

all uploaded datasets, run data mining tools as

desired, examine results via tables, plots, and related

explanations. If technically inclined, the doctor can

customize a number of filters. While discretization

HEALTHINF 2018 - 11th International Conference on Health Informatics

366

Figure 10: Sample LCPS analysis screens.

happens automatically if needed, one can specify

which method and how many bins to use. As for

attribute selection, PCA is automatically be applied

but one can select other methods or even manually

specify which attributes to use for classification and

which to discard e.g., based on expert knowledge.

From this page, the doctor can modify the

database or view statistical information about the

dataset, as depicted in the sample screens of Fig. 10.

First, the analysis page shows the accuracy of correct

classifications using the highest scoring algorithm.

LCPS can also display statistics about how the

algorithms perform and how they differ in their

analysis and predictive model.

Figure 11: Sample Patient’s Prediction Page.

Next is shown a table listing the various correctly

and incorrectly classified instances in separate

classes. Next, the ROC curves and bar charts show

the true positives (TP) and false positives (FP) and

total of correct and incorrect values. Moreover, after

examining these graphs and results, the doctor can

still select any available algorithm, examine results in

further details, play “what if” scenarios by changing

some data and applying the algorithm anew. We aim

to implement automated explanatory mechanism later

on. Lastly, Fig. 11 shows the prediction page for a

fictitious patient using the life expectancy dataset. In

this case, the result reported is a bracket of 0–333

days to live. This can be regarded by the doctor as

fatal, and will be added to the patient’s record for later

use. The oncologist can always inspect the result and

calculation details.

LCPS heavily uses three types of visual tools to

help oncologists understand and appreciate the

generated results. Scatter plots as especially useful to

help identify anomalies in the dataset, which the

doctor can then decide to edit manually, or to show

how attribute values are distributed so the doctor can

determine if a patient profile is typical or not. Pie

charts show how many instances are correctly

classified in each class along with the total incorrectly

classified overall. Others show incorrect

classifications in each class. Pie charts are separate

Lung Cancer Prognosis System using Data Mining Techniques

367

for each predictive algorithm and can be used for

direct comparison as well. If the doctor wishes to see

if a patient falls into a specific class, he can examine

the details of each algorithm and the result values.

Line charts are also used often. For example, the

TP vs. FP rate line graph shows the ROC area which

can be used to identify visually how accurate each

prediction is. Explanations are given as well: in this

case, the more convex the ROC curve the better,

whereas a straight line would denote a random (hence

useless) classifier. The ROC curves are shown

separately for each algorithm, so that the doctor can

easily see how each algorithm performs. Lastly,

Column charts are best used to depict and compare

visually results such all the different TP and FP rate

for each algorithm and the overall correct and

incorrect classifications.

10 CONCLUSIONS

As mentioned earlier, cancer is one of the leading

causes of death in the UAE. In the emirate of Abu

Dhabi alone, 1,729 new cases were detected in 2012,

with 28 per cent among Emiratis and 72 per cent

among expatriate residents. This made cancer the

third leading cause of death in Abu Dhabi, accounting

for 13%of all fatalities (HAAD, 2016). This death toll

will keep on increasing if we do not devise a method

to find cancer at an early, curable stage. The UAE has

recently launched an initiative to cut down the

number cancer cases by 18% by 2021 (HAAD, 2016),

hence we hope that a system such as LCPS will be a

highly contributing factor.

The early detection of any type of cancer is of

paramount importance to pave the way for successful

cancer treatment. Unfortunately, most cancers are

only detected once they reach an advanced and

incurable stage. Individuals who are victims of many

types of cancer do not know about it until it’s too late.

Many people do not even go to the doctor to get

themselves checked for several reasons which can

include, affordability, fear, travelling cost or even

time. Everyone is so absorbed in their work that they

disregard the possibility of having cancer, even after

the symptoms start to manifest. Hence, this was one

of the driving factors for us to make such an

application. To be able to measure how much of a

societal impact it can have we need to evaluate its

impact on both general users and doctors. Therefore,

by focusing on just UAE-based users, our proposed

application is predicted to have a significant impact

on the country as a whole. By focusing on the patient-

specific aspect of LCPS, it becomes more apparent

why having a tailored prediction, that is specific to the

patient’s health, is important. It gives the oncologist

and patient room to forecast and define a targeted

treatment process, helping in better patient care and

improving the patient’s chances of survival and

remission. Thus, having such a system will decrease

lung cancer death cases significantly, especially since

there are no similar systems in the market.

REFERENCES

American Cancer Society, 2017. "Cancer Facts and Figures

2017". Atlanta, Ga: American Cancer Society.

American Cancer Society, 2016. "Lung Cancer Prevention

and Early Detection", Atlanta, Ga: American Cancer

Society.

Khaleej Times, 2016. "Lung Cancer Among Top 5 Fatal

Diseases in UAE", Khaleej Times, UAE.

WAM, 2015. "Tawam Hospital Now Offers Lung Cancer

Screening, a First for the UAE", WAM Emirates News.

HAAD, 2016. "What is Lung Cancer?", Health Authority

Abu Dhabi, UAE. [online: https://www.haad.ae/]

HMW, 2013. "GenieMD Announces First Place Finish in

ONC Blue Button Co-Design Challenge," Health &

Medicine Week, p. 476.

WebMD, 2017. "Datamonitor", WebMD Corporation.

Khaleej Times, 2014. "DHA to launch patient-centric app

Sehhaty at Gitex," Khaleej Times, UAE.

Mesko, B., 2015. "The Lungscreen App: Find out your

risk," ScienceRoll.

OFWW, 2011. "Lung Cancer; New Lung Cancer Findings

from University of Toronto Described," Obesity,

Fitness & Wellness Week, p. 846.

Park, A., 2001. "Cancer spotter," Time, vol. 158, no. 7.

Witten, I., Frank, E., 2016. Data Mining: Practical

Machine Learning Tool and Techniques, Amsterdam:

Morgan Kaufman.

Rangra, K., Bansal, D., 2014. "Comparative Study of Data

Mining Tools", International Journal of Advanced

Research in Computer Science and Software

Engineering, vol. 4, no. 6.

Zhu, Z., Ong, Y., 2007. "Markov Blanket-Embedded

Genetic Algorithm for Gene Selection", Pattern

Recognition, vol. 49, no. 11, pp. 3236-3248.

Bhattacharjee, A., 2001. "Classification of Human Lung

Carcinomas by Mrna Expression Profiling Reveals

Distinct Adenocarcinoma Subclasses", Proceedings of

the National Academy of Sciences, 98(24), 13790-

13795.

Venables, W., Ripley, B., 2012. Modern Applied Statistics

with S, 4th ed. New York: Springer.

HEALTHINF 2018 - 11th International Conference on Health Informatics

368