See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/342404666

In Depth Analysis of Lung Disease Prediction Using Machine Learning

Algorithms

Chapter · June 2020

DOI: 10.1007/978-981-15-6318-8_18

CITATIONS READS

3 4,115

5 authors, including:

Ishan Sen Md. Ikbal Hossain

Tampere University North South University
2 PUBLICATIONS 12 CITATIONS 1 PUBLICATION 3 CITATIONS

SEE PROFILE SEE PROFILE

Faisal Shakib Asaduzzaman Imran

North South University University of South Wales
1 PUBLICATION 3 CITATIONS 1 PUBLICATION 3 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Faisal Shakib on 22 March 2021.

The user has requested enhancement of the downloaded file.

 In Depth Analysis of Lung Disease Prediction
Using Machine Learning Algorithms

Ishan Sen, Md. Ikbal Hossain(&), Md. Faisal Hossan Shakib(&),

Md. Asaduzzaman Imran(&), and Faiz Al Faisal(&)

Department of Electrical and Computer Engineering (ECE),

North South University, Dhaka 1229, Bangladesh
{ishan.sen,ikbal.hossain,faisal.shakib,
asaduzzaman.imran,faiz.faisal}@northsouth.edu

Abstract. The objective of this examination is to investigate and foresee the

Lung Diseases with assistance from Machine Learning Algorithms. The most
common lung diseases are Asthma, Allergies, Chronic obstructive pulmonary
disease (COPD), bronchitis, emphysema, lung cancer and so on. It is important
to foresee the odds of lung sicknesses before it happens and by doing that
individuals can be causes and make fundamental strides before it occurs. In this
paper, we have worked with a collection of data and classified it with various
machine learning algorithms. We have collected 323 instances along with 19
attributes. These data have been collected from patients suffering from numerous
lung diseases along with other symptoms. The Lung diseases attribute contains
two types of category which are ‘Positive’ and ‘Negative’. ‘Positive’ means that
the person has a lung disease and so forth. The training of the dataset has been
done with K-Fold Cross Validation Technique and specifically, five Machine
Learning algorithms have been used which are Bagging, Logistic Regression,
Random Forest, Logistic model tree and Bayesian Networks. The accuracy for
the above mentioned machine learning algorithms are 88.00%, 88.92%, 90.15%,
89.23%, and 83.69% respectively.

Keywords: Machine learning algorithm
Lung diseases bagging
Logistic

Regression
Random Forest
Logistic model tree
Bayesian Networks

1 Introduction

Respiratory diseases are among the critical reasons for death around the world. Lung
contaminations (for the most part pneumonia and tuberculosis), carcinoma and chronic
obstructive pulmonary disease (COPD) together represented 9.5 million passing overall
during 2008, one-6th of the overall aggregate, the planet Health Organization evaluates
that a proportional four maladies represented one-tenth of disability-adjusted life-years
(DALYs) lost worldwide in 2008. Likewise among the 25 most fundamental causes
were COPD (positioned sixth in 1990 and ninth in 2010), tuberculosis (positioned
eighth in 1990 and thirteenth in 2010) and carcinoma (positioned 24th in 1990 and 22nd
in 2010) [1]. Forecast of Lung related malady is a difficult factor looked by specialists
and clinics. Right now, the exactness of lung sickness assumes an indispensable job.

© Springer Nature Singapore Pte Ltd. 2020

A. Bhattacharjee et al. (Eds.): MIND 2020, CCIS 1241, pp. 204–213, 2020.
https://doi.org/10.1007/978-981-15-6318-8_18
 In Depth Analysis of Lung Disease Prediction 205

Machine Learning methods are widely utilized in medical sectors. Data mining
holds great potential to explore the hidden patterns in huge information which will be
used for clinical diagnosis. Data mining allow health systems to use data systematically
and do the analysis for identifying inefficiencies, best practices that improve care and
reduce costs. Detection of lung disease is one among the vital issues and lots of
researchers are developing intelligent medical decision support systems to urge better
the power of the physicians. Therefore we conducted an enquiry supported machine
learning algorithms to seek out the probabilities of occurrence of lung disease before it
actually occurs, the only purpose of this analysis is that the identify the accuracy levels
of Bagging, Logistic Regression and Random Forest algorithms and to integrate this
leads to such a system which is user friendly in order that people can use it whenever
they feel convenient.

2 Related Works

In 2002, in the United Kingdom, Martin J Wildman et al. [2] and his team worked on to
assess whether the results found in terms of survival were the prognoses among
physicians of patients with intense poignancy with obstructive lung malady focused on
serious consideration. They have been worked on over 832 patients who were matured
45 years and more with a breathing problem, respiratory flunk, and variation of mental
health because of suffering from COPD and asthma, or a conjunction of both. Their
yielding is calculated by the physicians and they need around 180 days to predict this
outcome. D J Hole et al. [3] and his group worked on determining the connection
between the one-second forced expiratory volume (FEV1) and subsequent death. They
directed this research on 7058 men and 8353 women who were aged between 45–64
years. They wanted to find fatality rate from all causes, ischemic heart illness, lung
cancer, and other cancers, sudden stroke, disease of respiratory organ, and different
reasons for death following 15 years of development. They found that 2545 men and
1894 women have died in this period with critical patterns of expanding hazard with
lessening FEV1 for both genders, the reasons for death studied afterwards age change,
smoking, blood pressure, weight record and social groups are clear. Timor Kadir,
Fergus Gleeson [4] had given a review about fundamental lung disease expectation
ways to deal with date and feature a portion from their comparative qualities and
shortcomings in this paper. AI based lung malignant development expectation models
have been offered to help medical practitioners in overseeing unknown or recognized
uncertain aspiratory knobs. They have used CNN using trained data, which is a class of
deep neural network. While evaluating there were needed to be informed about some
limitations whether the patients were smoker or with history of lung nodule. Before
testing their system, their works concluded with some related questions for future
works that if it were to implement, which output or who should use this system.
Anuradha et al. [5] proposed a model, which uses machine learning that is to recognize
and analyze the lung disease as ahead of schedule as conceivable which will assist the
specialist with saving the patient’s life. They used binary classification with different
parameter like age, gender, X-ray images and view position. And the expected output is
to find whether the patient has lung disease or not. For their particular dataset, they
206 I. Sen et al.

have used mainly CNN and capsule network algorithm for extraction since their chosen
dataset of chest x-ray may have many unnecessary data. This paper portrays how lung
maladies were anticipated and controlled, utilizing Machine Learning. With pre-trained
model they would improve their accuracy level. Jason et al. [6] and his team in their
work they gathered information from aspiratory nodule(s) patients and they were about
1500 and their nodules were till 15 mm identified and distinguished on routinely
performed CT chest filters matured 18 years of age or else more established from three
scholastic focus in the United Kingdom were utilized for creating risk lamination
models. For developing their model, they have used Artificial Intelligence (AI) based
dataset. Aiyesha et al. [7] have examined the current clinical system and they have
suggested how AI based strategies might be utilized for analysis differentially on
Phthisis and Pneumonia. Here they presented a classification model using three
machine learning algorithm which are Naïve Bayes, Decision Tree and Random forest.
In this paper we also used these algorithms to show different results.

3 Methodology

The examination can be partitioned into four crucial stages which are given as follows:
• Data Collection
• Data Processing
• Data Training
• Application of ML algorithms

3.1 Data Collection

The data for this analysis has been collected from some hospitals in Dhaka. National
Institute of Diseases of the Chest and Hospital, National Institute of Cancer Research &
Hospital (NICRH) helped us by providing us with the majority of the data. A total of
323 instances and 19 attributes have been collected in which there is information for an
individual patient. The dataset contains the lung diseases attribute which is categorized
by two types; Positive (have lung diseases) and Negative (Doesn’t have any lung
diseases) (Table 1).

Table 1. The hospitals which helped us to collect data are mentioned below
1. National Institute of Diseases of the Chest and Hospital
2. National Institute of Cancer Research & Hospital (NICRH)
 In Depth Analysis of Lung Disease Prediction 207

3.2 Data Table 2. Features list

Preprocessing Features Subcategory Data distribution
Sex Male 46.62%
In the wake of gathering the
Female 53.08%
information, we did the prepro- Age Lowest: 7 Mean ± SD
cessing inside the preprocessing Highest: 85 43.412 ± 17.475
stage, we utilized the two unsu- Hemoglobin (Hb) Lowest: 5 gm/dl 11.526 ± 1.491
pervised filters inside the exten- Highest:
sively utilized AI stage WEKA 16.9 gm/dl
3.8.3 (Waikato Environment for Erythrocyte Lowest: 2 mm 59.105 ± 33.893
sedimentation rate (ESR) Highest:
Knowledge Analysis). From the
165 mm
start, we applied the Replace
White Blood Cell (WBC) Distinct: 102 Unique: 22(07%)
MissingValues filter on our data- Platelet Count (PC) Distinct: 131 Unique: 43(13%)
set. This replaces all the missing Hematocrit (HCT) Lowest: 9.3% 33.979 ± 4.718
qualities for ostensible and Highest: 54.6%
numerical traits utilizing the Neutrophils Lowest: 6% 68.006 ± 14.591
modes and means. Besides, we’ve Highest: 96%
utilized the Randomize filter Lymphocytes Lowest: 2% 24.035 ± 13.588
which replaces the missing data Highest: 90%
without surrendering a great part Monocytes Lowest: 1% 5.168 ± 2.82
Highest: 17%
of the exhibition.
Eosinophil Lowest: 0% 4.342 ± 2.662
Highest: 16%
3.3 Data Training Basophils Lowest: 0% 0
Highest: 0%
The preparation of the data has Blood-Glucose-After- Lowest: 106.099 ± 48.443
been finished utilizing the K-Fold Meal 40 mg/dl
Cross Validation method of Highest:
WEKA. It is a resampling proce- 312.25 mg/dl
dure to evaluate the forecast Serum-creatinine Lowest: 0.868 ± 0.236
0.3 mg/dl
model by parting the first dataset
Highest: 2 mg/dl
into two sections preparing the set Serum-bilirubin Lowest: 1.085 ± 2.249
and a test set. The parameter K 0.2 mg/dl
decides the number of groups the Highest:
dataset will be separated into by 12 mg/dl
rearranging the dataset haphaz- Serum glutamic pyruvic Lowest: 6 u/l 33.049 ± 25.767
transaminase (SGPT) Highest: 178 u/l
-ardly.
Height Lowest: 130 cm 160.157 ± 9.608
Highest: 176 cm
3.4 Application of Weight Lowest: 20 kg, 50.231 ± 8.989
Machine Learning Highest: 75 kg
Algorithms Cancer-Test-Result Negative: 307 94.17%
{Negative, Positive} Positive: 18 5.83%
After the training phase, classifi- Lung-Disease-Result Positive: 285 87.423%
cation has been done using vari- Negative: 40 12.577%
ous machine learning algorithms
among them Bagging, Logistic
208 I. Sen et al.

Regression and Random Forest Logistic model tree and Bayesian Networks outper-
formed. Therefore, we determined those five algorithms to be our model.

4 Workflow

Figure 1 represents the overall workflow of the entire analysis illustrated in brief. We
collected a dataset comprising of 323 instances from hospitals mentioned in Table 2
along with 19 attributes. After that, we preprocessed the data and used the feature
selection option in WEKA. The next thing we did was to train the dataset using K-fold
Cross-Validation Technique. Afterward, we applied various machine learning algo-
rithms among which three of the above mentioned algorithms stood out. In the end, we
concluded our analysis by comparing the performances of the five algorithms.

Fig. 1. Flow chart of the overall analysis

 In Depth Analysis of Lung Disease Prediction 209

Performance Parameters
The outcomes of this analysis are based on the following performance parameters:
A. Seed:
It indicates changing numbers haphazardly and getting an alternate outcome.
B. Correctly Classified Instances (CCI):
The precision of the model relies upon the test information.

Tp þ Tn
Accuracy ¼
Tp þ Tn þ Fp þ Fn

Here, Tp = True Positive, Fp = False Positive, Fn = False Negative, Tn = True

Negative.
C. Kappa Statistics (KS):
The Kappa Statistics is used to quantify the proclamation among predicted and
observed provisions of a dataset.

R0  Re
K¼
1  Re
Here, R0 = Relative watched understanding among raters, Re = Theoretical likeli-
hood of chance assertion.
D. Mean Absolute Error (MAE):
Without assessing the sign it normalizes the size of individual mistakes.

jp1  b1j þ


þ jpn  bn j

MAE ¼
n
Here, p = Predicted Value, b = Actual Value.
E. Relative Absolute Error (RAE):
It is the absolute error with the similar manner of normalization.

jp1  b1j þ


þ jpn  bnj

RAE ¼
jb1  bj þ


þ jbn  bj

F. Specificity:
It is the proportion of patients without the disease who test negative.

Tn
Specificity ¼
Fp þ Tn

G. Precision (PRE):
Tp
It is denoted as PRE ðPRE ¼ Tp þ FP
Þ:
210 I. Sen et al.

H. Recall (REC):
It is denoted as REC and the cohesion between PRE and REC is MCC.

Tp
REC ¼
Tp þ Fn

I. F-Measure:
It is denoted by FM.

PRE  REC 2  Tp
FM ¼ 2  ¼
PRE þ REC 2  Tp þ Fp þ Fn

5 Results Analysis

The examination was finished in 3 seeds utilizing every calculation for 2 classes titled
as Positive and Negative. In Figs. 2 and 3, where the x-axis shows to the False positive
rate (Fp) and the y-axis connotes the True positive rate (Tp). Tp is likewise perceived as
Recall and it is utilized to quantify the real level of accurately ordered cases. Fp is the
proportion between the number of negative cases that are inaccurately delegated
positive and the quantity of really negative examples. (0, 1) a point on the diagram is
where the classifier gives the ideal outcome for example it orders the genuinely positive
and negative cases effectively.
In Fig. 3a, 3b given below illustrates the percentage of correctly classified instances
and incorrectly classified instances respectively from 5 different algorithms. From
Fig. 3a, it has clearly seen that the Random Forest (RF) gives the highest accuracy
results whereas Bayes Net gives the lowest accuracy; the figure is approximately
90.1538% and 83.6929% successively. On the other hand, Logistic Regression
(LR) gives us nearly 88.9231% which is higher than Bagging which gives us almost
88% accuracy. However, the logical model tree (LMT) is slightly higher than LR and
the figure is around 88.2308%. From Fig. 3b it can be said that Bayes Net has given us
the highest incorrect instances and the percentage is around 16.30 whereas the RF and
LMT give us almost lowest incorrect instances and the figure are almost 10%. Table 3
represents the comparison of five different algorithms on various evaluation matrix. It
can be observed that the KS value for Bagging, LR, RF, LMT, Bayes Net are 0.2821,
0.3803, 0.501, 0.4052, 0.2974 respectively.
The MAE for Bagging, LR, RF, LMT, Bayes Net are 0.1785, 0.1592, 0.1551,
0.197 successively whereas the Root Mean Squared Error are 0.3102, 0.3038, 0.2888,
0.3015 and 0.3258 respectively for these algorithms. In terms of RAE, it can be said
that the Bayes Net gives the highest percentage of error, which is 90.4575% whereas
 In Depth Analysis of Lung Disease Prediction 211

Fig. 2. (a) BAG curve for positive class (b) LR curve for positive class (c) RF curves for
positive class (d) LMT curve for positive class (e) Bayes net curve for positive class

the lowest proportion of these errors is 70.1852% percentage, which is for RF. On the
other hand, this proportion is slightly lower which is for LMT and the figure is
71.1941%. If it comes to Bagging the error is 81.9378% which is the second-lowest
error after Bayes Net. Now if we talk about the Root Relative Squared Error then the
Bayes Net gives 99.1762% whereas RF gives 87.9163%. On the other hand, Bagging,
LR, KMT give 94.4328%, 92.4839%. 91.7746% respectively.
212 I. Sen et al.

Fig. 3. (a) Correctly classified instances (b) Incorrectly classified instances

Table 3. Statistical metrics comparison for seed 1

Evaluation Metrix Bagging Logistic Random LMT Bayes
Regr. (LR) Forest (RF) Net
Kappa Statistic 0.2821 0.3803 0.501 0.4052 0.2974
Mean absolute 0.1785 0.1592 0.1529 0.1551 0.197
Error
Root Mean Squared 0.3102 0.3038 0.2888 0.3015 0.3258
Error
Relative absolute 81.93% 73.09% 70.18% 71.19% 90.45%
Error
Root Relative 94.43% 92.48% 87.91% 91.77% 99.17%
Squared error
TR Rate (Weighted 0.880 0.889 0.902 0.892 0.837
Avg.)
FR Rate (Wei. 0.662 0.574 0.444 0.552 0.517
Avg.)
Precision (Wei. 0.856 0.873 0.894 0.878 0.849
Avg.)
Recall (Wei. Avg.) 0.880 0.889 0.902 0.892 0.837
F-Measure (Wei. 0.861 0.877 0.897 0.881 0.842
Avg.)
 In Depth Analysis of Lung Disease Prediction 213

6 Conclusion

In conclusion, our study shows that lung diseases can be correctly classified using the
machine learning techniques. However, obtaining the real time data was one of the
primary concerns that we had faced at the initial stages. In addition to that, we could
not able to get similar data from existing works to compare our results. However it is
worth to mention that our dataset has many attributes, which is rare to find from online
sources. Among the five algorithms, Random Forest gives the best performance than
LR, Bagging, LMT and Bayes Net. The accuracy level are 88.00%, 88.9231%,
90.1538%, 89.2308% and 83.6929% respectively. In addition to work, future resear-
ches like, study with deep learning methods like - Neuron Advanced Ensemble
Learning, Fuzzy Inference System, and Convolution Neural Network would be useful
and beneficial.

References
1. The burden of lung disease. https://www.erswhitebook.org/chapters/the-burden-of-lung-
disease/
2. Wildman, M.J., et al.: Implications of prognostic pessimism in patients with chronic
obstructive pulmonary disease (COPD) or asthma admitted to intensive care in the UK within
the COPD and asthma outcome study (CAOS): multicenter observational cohort study (2007).
https://doi.org/10.1136/bmj.39371.524271.55
3. Hole, D.J., et al.: Impaired lung function and mortality risk in men and women: findings from
the Renfrew and Paisley prospective population study (1996). https://doi.org/10.1136/bmj.
313.7059.711
4. Kadir, T., Gleeson, F.: Lung cancer prediction using machine learning and advanced imaging
techniques (2018). https://doi.org/10.21037/tlcr.2018.05.15
5. Gunasinghe, A.D., Aponso, A.C., Thirimanna, H.: Early prediction of lung diseases.
Conference Paper, March 2019
6. Oke, J.L., et al.: Development and validation of clinical prediction models to risk stratify
patients presenting with small pulmonary nodules: a research protocol. Diagn. Progn. Res. 2,
22 (2018)
7. Sadiya, A., et al.: Differential diagnosis of tuberculosis and pneumonia using machine
learning. Int. J. Innov. Technol. Explor. Eng. (IJITEE) 8(6S4), 245–250 (2019). ISSN 2278-
3075

View publication stats