2018 4th International Conference on Computing Communication and Automation (ICCCA)

Big Data Deep Learning Framework using Keras: A

Case Study of Pneumonia Prediction
Karan Jakhar Nishtha Hooda
Computer Science Engineering Computer Science and Engineering
Department, Chandigarh University Department Chandigarh University
Mohali, Punjab, Mohali, Punjab,
India India
[email protected] [email protected]

Abstract—Big Data predictive analytics using machine In this research work, the proposed prediction
learning techniques is currently a much active area of research model is implemented by convolutional deep neural
in medical science. With increasing size and complexity of networks (CNN) using Python programming language.
medical data like X-rays, deep learning gained huge success in CNNS, also known as ConvNets in deep learning.
prediction of many fatal diseases like pneumonia. In this
After pre-processing of data, different machine
research work, DCNN (deep convolutional neural networks)
learning algorithms are trained to measure the
an efficient predicting model for big data, having deep
performance of CNN with popular and modern
layers is a proposed, which can classify whether a person
classifiers. Promising results are achieved, when the
is having a pneumonia or not. The experiments are
results of the suggested framework is compared with
carried after extracting the features of high quality X-
the regular classifiers like SVM, random forest,
ray images data and achieved an prediction accuracy of
adaboost, etc. using different estimating metrics like
84% and AUC of
accuracy, specificity, area under the curve, and
Promising results are found, when the results of the
sensitivity, etc.
DCNN framework is compared with the regular
classifiers like SVM, random forest, etc. using different
evaluation metrics like accuracy, sensitivity, etc. With
the appearance of increasing cases of pneumonia, tactful
implementation of deep learning can play a big part in
improving the performance of prediction of many fatal
diseases in the future.

Keywords—big data; machine learning; prediction; deep

learning; pneumonia

I. INTRODUCTION

Around one million adults are diagnosis with pneumonia

and every year about fifty thousand die from this deadly in
the US alone [1]. Pneumonia is effecting a lot children who
are under age of five and also common cause of death of Fig. 1. 3 V's of Big Data [3]
them worldwide [2]. Predicting Pneumonia is important in
the medical field. Various tests can take some time but
Rest part of the paper is organized as follows:
predicting by X-ray of chest will help the doctor to get an
Section 2 discusses about the related work. Section 3
idea of the disease and steps can be taken accordingly.
presents brief discussion of classification methods
Detecting pneumonia by observing X-ray of chest is a
that are used in the suggested framework. Section 4
complex task and is an active area of research.
gives detail about the data, its features and
Deep learning as well as Big Data are two popular fields experimental setup. Section 5 explains summary of
in the rapidly growing digital world [3]. While Big Data has the experiment outcomes, graphs, and performance
numerous definitions, this research work refer it to the measures. Section 6 presents conclusion and about
veracity i.e. unstructured data as presented in the Figure 1, future scope.
defining important Vs of big data [4]. The medical data is
vast, complex, and difficult to analyze using conventional
data analysis techniques. Hence, deep learning offers a great
solution in harvesting valuable knowledge from such
complex medical data.

978-1-5386-6947-1/18/$31.00 ©2018 IEEE 1

 II. RELATED WORK
The data on which we want to predict should be in same
Researchers are utilizing the results of machine format as is training data. We tested our model on basis of
learning predictions for solving problems of medical various metrics (Accuracy, AUC , error rate, sensitivity, etc.)
science [12, 13, 14, 17, 18]. Medical images have large .
volume of information which can be extracted and used
for future prevention of dangerous diseases [19]. Many
researchers have implemented machine learning
algorithms using Python and R language for extracting
information from the medical images [17].
Use of ensemble methods for optimizing the results
of prediction accuracy is much in trend today. Ensemble
classifiers focuses on hybridization for improving the
results of machine learning prediction model [20].
Recently, deep learning is much active area of
research in medical science. Greenspan et al. has
reviewed the present and future perspectives of deep
learning in medical science [21]. Prediction model using
Convolutional Neural Networks (CNN) helps in
providing much better experimental results for high
dimensional image data [21].
Fig. 2 Proposed Prediction Framework
High dimensional data consists of the medical images
which has large number of feature descriptors. Feature Consider all these performance metrics, models can be
extraction techniques are applied on good quality X ray compared well because it is good to know what is TP rate as it
images to extract the numerous feature descriptors. Deep shows in how many cases the is going to do the right
learning neural networks are trained with the extracted classification.
data to build the prediction model [21]. Research is also
carried for prediction of pneumonia using machine
learning classifiers [22].

III. MATERIAL AND METHODS

The main purpose of exploring the field of machine
learning is get a trained model for the classification and
prediction of pneumonia patients, considering available
X-ray data. The outcome of DCNN proposed framework
helps to predict whether a person has pneumonia or not.
A. Proposed framework
The outcome of proposed DCNN framework helps to
predict whether a person has pneumonia or not based on
the X-ray image of chest. Normally, in real scenario
problems, there is less control over the quality of images.
Some regular pre- processing like removal of corrupted
images, cleaning, etc. are always required [6]. Machine
learning aim is to adopt efficient techniques to process
large and complex data also considering cost. The
abstract and detailed view of DCNN framework is
displayed in the Figure 2 and Figure 3 respectively.
Image is first preprocessed in required format for feeding
to neural network and also checked for any corrupted
image and removed it. Then, the converted data goes
through the DCNN where various features are extracted Fig. 3 DCNN Framework
at each level. At the end of the DCNN there is fully
connected layer and then the last layer which is output
layer which expect ‘1’ or ‘0’ , ‘1’ for pneumonia and ‘0’
for normal. With the help of back- propagation the
network learns the right weights. After the model is
trained it can be used for predict output on data which it
has not seen earlier.

2
 The person having pneumonia should be classified as V. RESULTS AND DISCUSSION
positive but if a person not having pneumonia is
classified as positive is not a big issue as it can be This section discusses parameter evaluation metrics to
further rectified. measure the performance of various machine learning
algorithms. The results are discussed much in detail and are
B. Machine Learning Classifiers also presented graphically.
i. Neural Network: The idea is based on human brain A. Performance Evaluation
working, like neuron communicates in human brain The results and performance of the suggested
the same concept is applied here. There are framework is evaluated with different parameters shown in
different layer of neurons and they activate other confusion matrix Table 1. The various evaluation metrics
neurons, like this is learns right weight for calculated from the Table 1 are presented in Table 2. Based
prediction [8]. on various metrics DCNN performed better than other
ii. Random Forest : It ensembles results of different models as shown in Figure
4. As it was imbalanced data, we cannot totally depend on
decision trees and take their average, by doing so
accuracy so comparing other metrics results DCNN gives
improves accuracy and also avoid over-fitting [7].
good results. Neural Network and Random Forest also quite
iii. Support Vector Machine: Using vector, this method good and are strong competitors. Comparing TP rate and FP
finds a hyperplane between the datasets. The rate DCNN maintaining its stand. Overall DCNN is giving
hyperplane acts like a wall between the different efficient result on unseen data.
classes. Checked the category of new unseen data
(in which group it falls as all are separated by TABLE I. CONFUSION MATRIX
hyper- plane) accordingly and results are also True Reference
shown here. The dimension depends on the number
of features [9]. Predicted Condition Condition Positive Condition Negative
iv. Adaboost : It is an ensemble based method in which Pneumonia F P (C)
the output of one become input of next tree after Positive T P (A)
some changes. Doing so improves the accuracy and Pneumonia T N (B)
over-fitting [10]. Negative F N (D)
v. Logistic Regression : This is a classification
method which learn some link in the dependant TABLE II. PERFORMANCE METRIC FORMULA
variable (label) and independent variables (features)
by considering the probability [11,15]. Sensitivity A/(A + B)
vi. Decision Tree: It is a graph based machine Specificity B/(D + B )
learning classifier [16] Accuracy (A + B )/(A + C + D + B )
F Score (2 *A)/((2 *A) + (D + C))

IV. EXPERIMENTAL INVESTIGATION MCC (A * B )(D * C)/SQRT((A + D)*(A + C) * (B + D) * (B + C))

This section discusses about the dataset and

experimental setup. B. Experimental Results
After experimentation and testing of different models, the
A. Dataset
results of various metrics are represented in the Table 3. It
Chest X-ray Images (pneumonia) for classification can easily be observed that accuracy and other parameters of
from the medical database [4]. The dataset consists of the DCNN are the best among all other models. The results
5,863 X-Ray images with two labels i.e. Pneumonia or are also graphically depicted in Figure 4.
Normal. Random forest and Neural network are also showing
B. Experimental Setting good results but the performance of DCNN is better than the
Python’s sklearn library is used to perform the state-of-the art methods. False positive rate is low for DCNN
various tasks like pre-processing of images and model and True positive rate is high which is good as they show that
building techniques. For implementation of model is working good on unseen data. The patient having
convolutional neural network Keras library is used. The pneumonia is more likely be detected.
aim is to measure the classification accuracy of the The chance of cases of patients having pneumonia but
classifiers after training them then test on new samples classified as normal is low which is shown by False positive
which where were not shown to the model before and rate. Many models have accuracy close to one another but
checking the classification strength. To measure the when we consider other metrics then we can compare the
performance of the suggested framework, seven models easily. TP rate is also an important metric to consider
parameters namely accuracy, MCC, F measure, error when comparing the models. To check the robustness of
rate, TP Rate, FP rate, and also area under the curve proposed DCNN model, K fold cross validation method is
(AUC) are used. used.
With using 10 iterations, the stability of DCNN is
presented graphically in the Figure 5 and Figure 6.

3
 TABLE III. COMPARISON OF DCNN
PERFORMANCE WITH DIFFERENT STATE-OF-THE
ART METHODS USING MACHINE LEARNING
PERFORMANCE METRICS

Classifier Accur Erro TP FP F MCC

acy r Rate Rate Score
(%) rate
(%)
Decision 77 23 0.83 0.24 0.62 0.50
tree
Adaboost 78 22 0.92 0.25 0.60 0.53

Random 82 18 0.88 0.20 0.70 0.65

forest
SVM 76 23 0.91 0.26 0.56 0.50

Logistic 77 23 0.90 0.26 0.57 0.50 Fig. 6. K fold cross validation of AUC

DCNN 84 16 0.92 0.11 0.77 0.66 At last, K fold cross validation (with K=10) is performed
to test the robustness of DCNN framework. The result for K
Neural 81 18 0.72 0.15 0.76 0.62 fold validation for accuracy and AUC are depicted
Netwok graphically in Figure 5 and Figure 6 respectively.
Naive 72 27 0.63 0.22 0.62 0.40
Bayes As it can be observed from the graphs, the values of
accuracy and AUC are quite stable in all ten folds of cross
validation. Hence, promising results are achieved for the
prediction of pneumonia by the proposed framework.

VI. CONCLUSION AND FUTURE SCOPE

Pneumonia is life-threatening if it is not diagnosed
properly in patients. Around two third of the global
population lacks access to radiology diagnostics in India,
according to an estimate by the World Health
Organization. In this research, chest X ray image reports
are utilized to train an efficient deep machine learning
based prediction model for predicting pneumonia in
patients. Deep learning makes this task more effective as
deep learning is efficient in case of image data processing.

An efficient model is built using deep learning

algorithms in Python language which will help doctors to
Fig. 4 Comparison of Accuracy of DCNN with state-of-the-art detect this deadly disease. The proposed framework is
methods compared with state-of-the art methods of machine
learning and found to be more efficient in prediction with
an average accuracy of 84%, which is found to be better
than all other classifiers.
For future work, optimization of results will be done for
improving the performance of prediction. Further, more
volume of image data will be collected and data processing
is done on the top of Hadoop framwork.

Fig. 5 K fold cross validation of accuracy

4
 REFERENCES [18] Magoulas GD, Prentza A. Machine learning in medical applications.
[1] P. Rajpurkar et al. "Chexnet: Radiologist-level pneumonia detection InAdvanced Course on Artificial Intelligence 1999 Jul 5 (pp. 300-
on chest x-rays with deep learning."arXiv preprint 307). Springer, Berlin, Heidelberg.pneumonia pattern using RNA-Seq
arXiv:1711.0522,2017. and machine learning: challenges and solutions. BMC genomics.
[2] WHO. Pneumonia, 2016 [Online] Available: 2018 May;19(2):101.
http://www.who.int/news- room/fact-sheets/detail/pneumonia [19] Wernick MN, Yang Y, Brankov JG, Yourganov G, Strother SC.
[Accessed: May 24, 2018] Machine learning in medical imaging. IEEE signal processing
[3] X. Chen. "Big data deep learning: challenges and magazine. 2010 Jul;27(4):25-38.
perspectives."IEEE access", pp. 514-525, 2014. [20] Dietterich, Thomas G. "Ensemble methods in machine learning."
[4] A. Gandomi et al. "Beyond the hype: Big data concepts, methods, International workshop on multiple classifier systems. Springer,
and analytics."International Journal of Information Management Berlin, Heidelberg, 2000.
vol 35(2), pp.137-144, 2014. [21] Greenspan H, Van Ginneken B, Summers RM. Guest editorial deep
[5] Chest XRay data, 2018, [ONLINE] Available: learning in medical imaging: Overview and future promise of an
http://dx.doi.org/10.17632/rscbjbr9sj.2#file-41d542e7-7f91-47f6- exciting new technique. IEEE Transactions on Medical Imaging.
9ff2- dd8e5a5a7861 [Accessed: May 24, 2018] 2016 May;35(5):1153-9.
[6] H. Nishtha et al. "B2FSE framework for high dimensional [22] Choi Y, Liu TT, Pankratz DG, Colby TV, Barth NM, Lynch DA,
imbalanced data: A case study for drug toxicity Walsh PS, Raghu G, Kennedy GC, Huang J. Identification of usual
prediction."Neurocomputing vol. 276, pp.31-41, 2018. interstitial pneumonia pattern using RNA-Seq and machine learning:
[7] Liaw, Andy, and Matthew Wiener. "Classification and regression by challenges and solutions. BMC genomics. 2018 May;19(2):101.
randomForest."R news vol. 2(3), pp. 18-22, 2002.
[8] A. Rowley et al. . "Neural network-based face detection."IEEE
Transactions on pattern analysis and machine intelligence vol.
20(1), pp. 23- 38, 1998.
[9] M. Hearst, et al. "Support vector machines."IEEE Intelligent
Systems and their applications vol. 13(4), pp. 18-28, 1998.
[10] R.. Takashi Onoda, and K-R. Müller. "Soft margins for
AdaBoost."Machine learning, vol. 42(3), pp. 287-320,2001.
[11] Hosmer Jr, David W., Stanley Lemeshow, and Rodney X.
Sturdivant. Applied logistic regression. Vol. 398. John Wiley &
Sons, 2013.
[12] P. Pedro et al. . Community-acquired pneumonia: identification and
evaluation of non responders. Therapeutic advances in infectious
disease, 1(1), pp. 5-17, 2013.
[13] M. Aydogdu et al. Mortality prediction in community-acquired
pneumonia requiring mechanical ventilation; values of pneumonia
and intensive care unit severity scores. Tuberk Toraks, vol. 58(1),
pp. 25–34, 2010.
[14] D. Mollura et al. White paper report of the rad-aid conference on
international radiology for developing countries: identifying
challenges, opportunities, and strategies for imaging services in the
developing world. Journal of the American College of Radiology,
vol. 7(7), pp. 495– 500, 2010.
[15] Press, S. James, and Sandra Wilson. "Choosing between logistic
regression and discriminant analysis." Journal of the American
Statistical Association 73.364 (1978): 699-705
[16] Safavian, S. Rasoul, and David Landgrebe. "A survey of decision
tree classifier methodology." IEEE transactions on systems, man,
and cybernetics 21.3 (1991): 660-674.
[17] Kononenko I. Machine learning for medical diagnosis: history, state
of the art and perspective. Artificial Intelligence in medicine. 2001
Aug 1;23(1):89-109.