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

net/publication/368845592

Prediction of Anemia using Machine Learning Algorithms

Article in International Journal of Computer Science and Information Technology · February 2023
DOI: 10.5121/ijcsit.2023.15102

CITATIONS READS

20 3,296

3 authors, including:

Prakriti Dhakal Rabindra Bista

3 PUBLICATIONS 26 CITATIONS
Kathmandu University
44 PUBLICATIONS 595 CITATIONS
SEE PROFILE
SEE PROFILE

All content following this page was uploaded by Rabindra Bista on 07 August 2023.

The user has requested enhancement of the downloaded file.

 International Journal of Computer Science & Information Technology (IJCSIT) Vol 15, No 1, February 2023

PREDICTION OF ANEMIA USING MACHINE

LEARNING ALGORITHMS
Prakriti Dhakal, Santosh Khanal, and Rabindra Bista

Department of Computer Science and Engineering, Kathmandu University,

Dhulikhel, Nepal

ABSTRACT
Anemia is a state of poor health where there is presence of low amount of red blood cell in blood stream.
This research aims to design a model for prediction of Anemia in children under 5 years of age using
Complete Blood Count reports. Data are collected from Kanti Children Hospital which consist of 700 data
records. Then they are preprocessed, normalized, balanced and selected machine learning algorithms were
applied. It is followed by verification, validation along with result analysis. Random Forest is the best
performer which showed accuracy of 98.4%. Finally, Feature Selection as well as Ensemble Learning
methods, Voting, Stacking, Bagging and Boosting were applied to improve the performance of algorithms.
Selecting the best performer algorithm, stacking with other algorithms, bagging it, boosting it are very
much crucial to improve accuracy despite of any time issue for prediction of anemia in children below 5
years of age.

KEYWORDS
Machine learning, Anemia, Children, Prediction, Algorithm, Accuracy

1. INTRODUCTION
Machine Learning is based on the idea that a system can learn from data, identifying key patterns
for better decision making that applies minimal human intervention [1]. Machine Learning
algorithms has proved to be an efficient tool for early prediction of fatal disease such as Anemia,
Hepatitis, Lung Cancer, Liver Disorder, Breast Cancer, Thyroid Disease, Diabetes etc. with
higher accuracy in order to save human life. In medical science, healthcare related data are being
used for predicting epidemics, for detecting various disease, for improving quality of life and
avoiding early deaths [2]. Thus, Machine Learning plays an important role in Health Informatics.

Anemia is a nutritional deficiency disorder, global public health problem affecting people of both
under developed and developed countries [3]. Anemia is a condition where the total
concentration of Red Blood Cells (RBC) or Hemoglobin (Hb) in the blood is low. According to
the World Health Organization (WHO), anemia is termed as ‘a condition in which the number of
red blood cells or their oxygen-carrying capacity is insufficient to meet physiologic needs’ [4].
Anemia disease can be classified on the basis of morphology and etiology. The most reliable
indicator of anemia is blood hemoglobin concentration, however, there are a number of factors
that can cause anemia including iron deficiency, chronic infections such as HIV, Tuberculosis,
vitamin deficiencies e.g. vitamins B12 and A, and acquired disorders that affect Red Blood Cell
production and Hemoglobin synthesis. Therefore, prediction of anemia plays most important role
in order to detect other associated diseases.

Children are the future of any country, the detection of anemia in early age helps to prevent other

DOI:10.5121/ijcsit.2023.15102 15
 International Journal of Computer Science & Information Technology (IJCSIT) Vol 15, No 1, February 2023

associated diseases in future which may seriously hamper their growth and development. This
issue emerges a social purpose to conduct this research. Furthermore, Anemia is typically
diagnosed on a complete blood count as it is the main test for effective diagnosis of anemia.
Henceforth, the main aim of this research is to design a model using different machine learning
algorithms and compare the performances of those algorithms on the basis of evaluation criteria
for prediction of Anemia using Complete Blood Count (CBC) for children under 5 years. The
section II presents the related survey, section III presents the methodology with experimental
setup, section IV shows results of experiments and section V concludes the paper.

2. LITERATURE SURVEY
Machine Learning has been an emerging tool for Prediction of Diseases. The work [5] has
figured out that each algorithm has its own strength as well as weakness and its own area of
implementation. The authors [6] identify those studies that applied more than one supervised
machine learning algorithm on one disease prediction. Algorithms include Random Forest,
Decision Tree, ANN, SVM, Logistic Regression, Naïve Bayes and K-nearest Neighbor. It shows
that Support Vector Machine (SVM) algorithm is applied most frequently and Random Forest
(RF) algorithm showed superior accuracy. The research [7] illustrates that many machine
learning algorithms have shown good results. It is so as they identify the related attributes
accurately.

The authors [8] investigated about supervised machine learning algorithms Naive Bayes, Random
Forest and Decision Tree algorithm for prediction of anemia using Complete Blood Count (CBC)
where Naive-Bayes technique performed well in terms of accuracy as compared to Decision Tree
and Random Forest. The work [9] determined which individual classifier or subset of classifier in
combination with each other achieves maximum accuracy in Red blood cell classification for
anemia detection showing unique idea of use of subset of classifier and use of ensemble learning
techniques. [10] specified anemia type for the anemic patients with dataset from the Complete
Blood Count (CBC) which showed J48 Decision Tree as best performer.

The research [11] predicted the anemia status of children under five years taking common risk
factors as features. The research concluded that ML methods in addition to the classical
regression techniques can be considered to predict anemia. The authors [12] constructed some
predictive models by using the identified risk factors through machine learning approach predict
the anemia status of children under 36 months. The work [13] constructed a prediction model to
predict the potential risk of anemia among infants from Multilayer Perceptron model (MLP)
which identified three risk factors for anemia including exclusive breastfeeding, maternal anemia
during pregnancy and non-timely supplementation of complementary food. The authors [14]
examined the prevalence of anemia in under-five years children taking Ghanian population which
showed higher prevalence below 2 years of age. The authors [1] investigates the prevalence of
anemia as children grew from infancy to preschool-age for check the dynamic anemia status of
children over time where children were at greater risk for developing anemia have persistent
anemia between toddlerhood and preschool-age.

From the survey, we could only find few technical predictions for children with technical results
that includes running the classifier algorithm and figuring out technical results that includes
accuracy, precision and other technical factors. Technical prediction of anemia for other age
groups, not for the case of children. Here, the prediction method only considered the risk factor,
social, economic factor, not blood reports. Therefore, there became a need of conducting
technical research that includes running classifier algorithm and producing technical output like
accuracy, precision for prediction of anemia in children considering a detail analysis of Blood
Report.
16
 International Journal of Computer Science & Information Technology (IJCSIT) Vol 15, No 1, February 2023

3. METHODOLOGY
We have proposed a research framework to answer our research questions. The research
framework is guided by a computational framework. The research started with a literature survey,
then advanced to data collection, algorithm processing, verification, validation and at last ended
with a result.

3.1. Data Collection

For the data collection, our main research site was Kanti Children Hospital, from where we
collected 700 data records of children below 5 years of age. Among the 18 attributes of Complete
Blood Count report, RBC counts are mainly used for classifying anemia is a person. So, we
selected 7 attributes in the RBC count section. The 7 attributes include Red Blood Cells (RBC),
Hemoglobin (Hb), Hematocrit (HCT), Mean Corpuscular Volume (MCV), Mean Corpuscular
Hemoglobin (MCH), Mean Corpuscular Hemoglobin Concentration (MCHC), RDW-cv. We
selected these attributes for prediction of Anemia. For children we consulted a doctor and
referred to the pediatric reference range to estimate the cut-off range for children below 5 years
of age.

After data collection, the raw data were preprocessed. The data has been recorded in
Hematological Analyzer from which data was collected manually, then we prepared dataset for
our research with different pre-processing techniques followed by data normalization.

3.2. Model Preparation

After the pre-processing of dataset, the dataset becomes ready to run in a classifier algorithm. For
anemia prediction we selected six classifier algorithms i.e. Random Forest, Decision Tree, Naïve
Bayes, Artificial Neural Network, Support Vector Machine and Logistic Regression. In the
model, we have used 10-fold cross validation for verification and validation. We used 10-fold
cross validation for separating data int o training set and testing set where data has been separated
into folds i.e. in 10 k-folds.

3.3. Performance Evaluation

The evaluation is based on confusion matrix. There are formulas related with confusion matrix to
calculate performance of any classifier algorithm. Along with performance measure from
confusion matrix, we have also evaluated other additional performance metrics related to time as
well. The performance evaluation was based on accuracy, precision, f-score, recall and area under
the curve. We also calculated the CPU time and Wall time required for running the algorithms.

4. EXPERIMENTS AND RESULTS

All data related portion including data pre-processing, data analysis and building different
machine learning models were done using Python Programming Language. Some of the tools that
have been used for the experimentation part of this research include Google Colab, Python
Programming Language, SciPy, Scikit-learn, NumPy, Pandas, Matplotlib.

17
 International Journal of Computer Science & Information Technology (IJCSIT) Vol 15, No 1, February 2023

4.1. Comparative Performance Evaluation

We have selected six classifier algorithms for experimentation. These algorithms have their own
specific conditions for processing. These conditions i.e. the hyperparameters were normalized for
experimentation. We performed three experiments.

Table 1. Comparative Performance Analysis of Classifier Algorithms

Algorithm Accuracy Precision Recall F1- AUC CPU Wall

score time time
(ms) (ms)
Logistic 0.807 0.834 0.862 0.848 0.788 85.2 86.1
Regression
Support 0.951 0.980 0.940 0.960 0.955 270 273
Vector
Machine
Naïve 0.907 0.969 0.878 0.921 0.916 20.4 20.5
Baye
Decision 0.972 0.975 0.981 0.978 0.969 32.3 36.9
Tree
Artificial 0.961 0.970 0.967 0.969 0.959 9600 5110
Neural
Network
Random 0.984 0.981 0.988 0.985 0.979 1740 1750
Forest

From above Table 1, we can observe that Random Forest is the best performer in the case of all
the parameters i.e. Accuracy, Precision, Recall, F1-score, Area Under the Curve. Talking about
the time, the best performer took maximum CPU and Wall time to run the entire process. As the
prediction system is of medical data, so priority is given to the performance rather than time. The
fast performer here is Naïve Baye as it takes minimum CPU time and Wall time. Logistic
Regression, on the other hand showed minimum accuracy along with all the other parameters.
We can also analyze that as the accuracy improves concurrently there is improvement in
performance of other parameters as well.

4.2. Feature Analysis

Our aim in this research is to improve accuracy of the classifier algorithms for making the
accuracy more than that of the best performer. For that we conducted feature analysis i.e.
selecting best features starting from 3 best to 6 best.

Table 2. Feature Analysis (3 best, 4 best, 5 best, 6 best)

Algorithm 3 best 4 best 5 best 6 best

Logistic 0.764 0.797 0.801 0.821
Regression
Support Vector 0.906 0.946 0.948 0.947
Machine
Naïve Baye 0.892 0.925 0.912 0.915

Decision Tree 0.921 0.974 0.975 0.968

Artificial Neural 0.861 0.927 0.947 0.927

Network
Random Forest 0.938 0.975 0.982 0.981

18
 International Journal of Computer Science & Information Technology (IJCSIT) Vol 15, No 1, February 2023

Above Table 2 shows the corresponding accuracy from feature analysis of all the classifier
algorithms on the basis of 3 best, 4 best, 5 best and 6 best features. Overall, we can analyze that
the feature analysis and selection method was effective for improving accuracy only for Logistic
Regression, Naïve Bayes and Decision Tree whereas for SVM, ANN and Random Forest it was
not effective.

4.3. Ensemble Learning Methods

We performed experimentation through feature analysis but we were unable to improve the
accuracy more than of the best performer. Then, we conducted experiment by applying the
ensemble learning methods such as voting classifier, stacking, bagging, boosting. These ensemble
methods were applied to improve the overall accuracy of the model.

4.3.1. Voting Classifier

Below Table 3 shows the results after applying voting classifier for combination of different
algorithm along with its CPU time and Wall time. The accuracy of the best performer is 98.4%.
Voting classifier was not able to improve accuracy above 98.4% with any of the combination of
algorithms.

Table 3. Voting Classifier and corresponding accuracy

Voting Classifier
Algorithm Accuracy CPU time (ms) Wall time (ms)
RF+LR 0.982 1850 1800
RF+DT 0.972 1850 1810
RF+SVM 0.981 2000 2000
RF+NB 0.964 1790 1800
RF+ANN 0.975 1330 7490
DT+LR 0.972 148 149
DT+SVM 0.972 308 312
DT+NB 0.972 91.6 98.4
DT+ANN 0.972 10700 5.62
SVM+ANN 0.964 11000 5790
SVM+NB 0.937 293 255
SVM+LR 0.941 354 360
NB+ANN 0.942 10300 5470
NB+LR 0.922 129 130
ANN+LR 0.964 16800 5670
RF+LR+DT 0.978 1850 1870
RF+LR+DT+ANN 0.981 13200 7440
RF+LR+DT+NB 0.980 1860 1870
RF+LR+DT+NB+ANN 0.981 1360 7650
RF+DT+NB 0.984 2310 2770
RF+DT+ANN 0.984 7970 7960
RF+DT+ANN+NB+SVM+LR 0.978 14230 8210

19
 International Journal of Computer Science & Information Technology (IJCSIT) Vol 15, No 1, February 2023

4.3.2. Stacking

The Table 4 below shows results after stacking ensemble learning methods. Random Forest with
Logistic Regression or ANN when stacked produced accuracy of 98.7% as well as Random
Forest when stacked with SVM or Naïve Bayes produced accuracy of 98.6%. Furthermore, when
all the six algorithms considered for the study when stacked also produced accuracy of 98.6% i.e.
the accuracy improved.

Table 4. Feature Analysis (3 best, 4 best, 5 best, 6 best)

Stacking

Algorithm Accuracy CPU Wall

time (ms) time (ms)
RF+LR 0.987 10200 10200
RF+DT 0.981 9790 9800
RF+SVM 0.986 12700 13100
RF+NB 0.986 10000 10100
RF+ANN 0.987 100000 307000
DT+LR 0.972 719000 746000
DT+SVM 0.974 890 892
DT+NB 0.971 321 335
DT+ANN 0.972 31800 17400
SVM+ANN 0.967 56200 30300
SVM+NB 0.958 1.35 1.36
SVM+LR 0.957 1860 1950
NB+ANN 0.960 52.6 s 28.3 s
NB+LR 0.927 607 615
ANN+LR 0.962 32300 17400
RF+LR+DT 0.982 9890 9920
RF+LR+DT+ANN 0.981 66000 40600
RF+LR+DT+NB 0.984 10500 10600
RF+LR+DT+NB+ANN 0.982 65000 39600
RF+DT+NB 0.984 10000 10000
RF+DT+MLP 0.984 64000 38700
RF+DT+ANN+MLP+SVM+LR 0.986 65000 39700

4.3.3. Bagging

The Table 5 below shows value of n and the corresponding accuracy. We have to specify how
many random samples the data have to be separated i.e. the value of n. For Decision Tree
accuracy reached 98.6% when n is 20, which can be said as increase in accuracy than the best
performer. Finally, for Random Forest accuracy is 98.8% when value of n is 10, it was higher
than the best performer.

20
 International Journal of Computer Science & Information Technology (IJCSIT) Vol 15, No 1, February 2023
Table 5. Bagging with Different Values

Bagging LR SVM NB ANN DT RF

(value of n)

10 0.805 0.959 0.918 0.958 0.98 0.988

20 0.812 0.958 0.917 0.962 0.986 0.986
30 0.807 0.958 0.914 0.957 0.980 0.986
40 0.807 0.958 0.915 0.958 0.980 0.985
50 0.804 0.958 0.915 0.958 0.978 0.985
60 0.804 0.958 0.912 0.958 0.982 0.985
70 0.805 0.958 0.911 0.960 0.982 0.985
80 0.808 0.958 0.912 0.960 0.980 0.986
90 0.809 0.958 0.912 0.958 0.978 0.986
100 0.809 0.958 0.911 0.958 0.980 0.986

4.3.4. Boosting

We talk about two boosting approaches namely Adaptive Boosting and XGB Booster in this
section.

Adaptive Boosting

The table 6 below shows the accuracy and corresponding CPU and Wall time with Ada Booster.
The accuracy of Naïve Bayes was 90.7% and it reached up to 93.7% when applied Adaptive
Boosting. For other base estimator, accuracy was not increased, neither the improved accuracy
was better than best performer.

Table 6. Adaptive Boosting with Corresponding Accuracy

Boosting (Adabooster)
Algorithm Accuracy CPU time (ms) Wall time (ms)
Random Forest 0.984 1850 1850
Decision Tree 0.968 52 53
Artificial Neural Nan
Network
Naïve Bayes 0.937 1650 1650
Support Vector 0.942 4480 4490
Machine
Logistic Regression 0.624 3520 3530

XGB Booster

The Table 7 below shows the learning rate adjustment in Xgb Booster along with corresponding
accuracy it generates. We could see the fluctuation in the accuracy of the model. Finally, at
learning rate 0.07, 0.06, 0.05, 0.04 accuracy reached up to 99%.

21
 International Journal of Computer Science & Information Technology (IJCSIT) Vol 15, No 1, February 2023
Table 7. Learning Rate with Corresponding Accuracy

XGB Booster
Learning Rate Accuracy
2 0.972
1.9 0.974
1.8 0.982
1.7 0.978
1.6 0.980
1.5 0.980
1.5 0.981
1.4 0.978
1.3 0.986
1.2 0.983
1.1 0.980
1 0.980
0.9 0.986
0.8 0.982
0.7 0.986
0.6 0.978
0.5 0.978
0.4 0.982
0.3 0.984
0.2 0.984
0.1 0.983
0.09 0.984
0.08 0.983
0.07 0.99
0.06 0.99
0.05 0.99
0.04 0.99
0.03 0.99
0.02 0.984
0.01 0.972

4.4. Balanced Data

Data were balanced, trained, fit into different classifier algorithms model again and finally all the
experiments were performed.

4.4.1. Comparative Performance Evaluation

All the parameter, conditions as well as hyperparameters were same for this experiment section.

22
 International Journal of Computer Science & Information Technology (IJCSIT) Vol 15, No 1, February 2023
Table 8. Comparative Table Performance Analysis of Classifier Algorithms (Balanced Data)

Algorithm Accuracy Precision Recall F1- Area CPU Wall

score Under time time
Curve (ms) (ms)
Logistic 0.837 0.908 0.750 0.822 0.837 75.4 78
Regression
Support 0.954 0.978 0.929 0.953 0.954 357 362
Vector
Machine
Naïve 0.917 0.964 0.867 0.913 0.917 20.3 22.1
Baye
Decision 0.973 0.979 0.972 0.973 0.973 42.5 45.6
Tree
Artificial 0.964 0.967 0.956 0.962 0.964 11000 5890
Neural
Network
Random 0.986 0.984 0.988 0.986 0.986 1850 1830
Forest

Above Table 8 shows comparative performance analysis of algorithms from balanced dataset. We
can see the change in accuracy and also in other metrics. We can observe similar results, Random
Forest is the best performer whereas Logistic Regression is the weakest performer while
considering balanced data. However, Accuracy of Random Forest and Logistic Regression was
more than that of unbalanced data.

4.4.2. Feature Analysis

For balanced dataset we conducted feature analysis i.e. selecting best features starting from 3 best
to 6 best in the same manner as for unbalanced data.

Table 9. Feature Analysis (3 best, 4 best, 5 best, 6 best: Balanced Data)

Algorithm 3 best 4 best 5 best 6 best

Logistic Regression 0.831 0.841 0.836 0.839
Support Vector 0.929 0.951 0.948 0.956
Machine
Naïve Baye 0.878 0.926 0.915 0.923
Decision Tree 0.936 0.972 0.971 0.973
Artificial Neural 0.938 0.959 0.956 0.957
Network
Random Forest 0.955 0.979 0.980 0.979

The Table 9 shows results for feature selection for balance data. The result is varying in the case
of balanced data. For the case of unbalanced data feature selection was effective for Logistic
Regression, Naïve Bayes and Decision Tree whereas for balanced data, feature selection was
effective for Logistic Regression, SVM and Naïve Bayes. However, the overall accuracy could
not be improved.

4.4.3. Ensemble Learning Methods

We proceeded towards experiment by applying the ensemble learning methods i.e. voting
classifier, stacking, bagging, boosting for balance dataset.
23
 International Journal of Computer Science & Information Technology (IJCSIT) Vol 15, No 1, February 2023

Voting

The table 10 below shows accuracy for voting classifier with combination of different algorithm
in balanced dataset. Voting classifier was not able to improve accuracy above 98.6% with any of
the combination of algorithms for balanced data.

Table 10. Voting Classifier and corresponding accuracy (Balanced Data)

Voting Classifier
Algorithm Accuracy CPU time Wall time
(ms) (ms)
RF+LR 0.978 1930 1940
RF+DT 0.972 1900 1900
RF+SVM 0.982 2190 2200
RF+NB 0.962 1860 1860
RF+ANN 0.979 13500 7690
DT+LR 0.972 141 147
DT+SVM 0.972 406 410
DT+NB 0.972 68.1 71.6
DT+ANN 0.972 10300 5.42
SVM+ANN 0.964 11.2 s 5.97 s
SVM+NB 0.938 374 378
SVM+LR 0.951 435 442
NB+ANN 0.942 10300 4160
NB+LR 0.922 102 109
ANN+LR 0.962 10500 5570
RF+LR+DT 0.978 1940 1950
RF+LR+DT+ANN 0.982 13700 7820
RF+LR+DT+NB 0.982 1970 1980
RF+LR+DT+NB+ANN 0.979 13600 7790
RF+DT+NB 0.982 1900 1910
RF+DT+ANN 0.981 13600 7800
RF+DT+SVM 0.982 2240 2240
RF+DT+ANN+NB+SVM+LR 0.975 13900 8140

Stacking

The Table 11 shows the combination of algorithms when applied stacking ensemble methods
with corresponding accuracy. For unbalanced dataset stacking had proved effective to improve
accuracy. Stacking ensemble learning method did not proved to be effective for improving
overall accuracy than the best performer in the case balanced dataset.

24
 International Journal of Computer Science & Information Technology (IJCSIT) Vol 15, No 1, February 2023
Table 11. Stacking and corresponding accuracy (Balanced Data)

Stacking

Algorithm Accuracy CPU time (ms) Wall time (ms)

RF+LR 0.986 10800 10800
RF+DT 0.982 10600 10600
RF+SVM 0.986 11900 11900
RF+NB 0.986 10500 10500
RF+ANN 0.986 72000 42400
DT+LR 0.972 701000 725
DT+SVM 0.978 1820 1830
DT+NB 0.972 349 349
DT+ANN 0.978 58800 31200
SVM+ANN 0.965 61000 32800
SVM+NB 0.957 1720 1730
SVM+LR 0.957 2020 2040
NB+ANN 0.965 58700 31100
NB+LR 0.928 518 521
ANN+LR 0.964 58800 31300
RF+LR+DT 0.981 10900 10900
RF+LR+DT+ANN 0.982 71000 42600
RF+LR+DT+NB 0.980 11000 11000
RF+LR+DT+NB+ANN 0.982 71000 42600
RF+DT+NB 0.982 10700 10700
RF+DT+ANN 0.982 71000 42100
RF+DT+SVM 0.981 12100 12100
RF+DT+ANN+NB+SVM+LR 0.982 73000 44500

Bagging

The Table 12 shows the value of n and the accuracy it gives when it executed for balanced data.
From the table we can elaborate that when the value of n was adjusted, the accuracy was
improved for all the algorithms. Random Forest when bagged gave accuracy higher than best
performer. Bagging proved to be the effective method to improve accuracy for best performer in
the case of balanced data. In the case of unbalanced data also bagging proved to be effective.

25
 International Journal of Computer Science & Information Technology (IJCSIT) Vol 15, No 1, February 2023
Table 12. Bagging with Different Values (Balanced Data)

Bagging LR SVM NB ANN DT RF

(value of n)

10 0.844 0.956 0.916 0.962 0.978 0.986

20 0.845 0.956 0.918 0.966 0.981 0.987
30 0.841 0.957 0.916 0.964 0.982 0.987
40 0.844 0.956 0.915 0.965 0.982 0.987
50 0.842 0.956 0.915 0.966 0.983 0.986
60 0.844 0.956 0.915 0.965 0.982 0.986
70 0.844 0.956 0.915 0.966 0.982 0.986
80 0.842 0.956 0.915 0.969 0.982 0.986
90 0.844 0.956 0.914 0.969 0.982 0.986
100 0.842 0.956 0.915 0.970 0.982 0.986

Boosting

Adaptive Boosting

The Table 13 shows the accuracy and corresponding CPU and Wall time with Ada Booster with
balanced data. For Random Forest as a base estimator the accuracy remained the same. For other
base estimators’ accuracy was not increased. In this case, Adaptive Boosting did not prove to be
effective for increasing accuracy than the best performer for balanced dataset.

Table 13. Adaptive Boosting with Corresponding Accuracy (Balanced Data)

Boosting (Adabooster)
Algorithm Accuracy CPU time (ms) Wall time (ms)
Random Forest 0.986 2040 2040
Decision Tree 0.973 63.8 75.9
Artificial Neural Nan
Network
Naïve Bayes 0.940 206 210
Support Vector 0.900 1530 1530
Machine
Logistic Regression 0.624 403 403

XGB Booster

The table 14 shows the learning rate adjustment in Xgb Booster along with corresponding
accuracy it generates. We could observe the fluctuation in the accuracy of the model. At learning
rate 0.03 accuracy reached up to 98.6% which was just equal to the best performer.

26
 International Journal of Computer Science & Information Technology (IJCSIT) Vol 15, No 1, February 2023
Table 14: Learning Rate with Corresponding Accuracy (Balanced Data)

XGB Booster
Learning Rate Accuracy
2 0.979
1.9 0.980
1.8 0.978
1.7 0.975
1.6 0.978
1.5 0.977
1.4 0.979
1.3 0.981
1.2 0.981
1.1 0.981
1 0.974
0.9 0.980
0.8 0.982
0.7 0.980
0.6 0.983
0.5 0.982
0.4 0.982
0.3 0.981
0.2 0.982
0.1 0.983
0.09 0.982
0.08 0.983
0.07 0.983
0.06 0.982
0.05 0.982
0.04 0.982
0.03 0.986
0.02 0.983
0.01 0.981

We can see some variations in results of unbalanced and balanced data. For both the cases, the
best performer was Random Forest. For unbalanced data accuracy was 98.4% whereas accuracy
increased up to 98.6% for balanced data. The accuracy for all the algorithms increased when data
was balanced. In the case of unbalanced data, the F1-score was higher than the accuracy.
However, F1-score should be lower than the accuracy as it is one of the major parameters in the
case of medical data. For the balanced data this issue was solved as F1-score was not more than
accuracy, rather it was less or equal to accuracy for all the classifier algorithms. The same
scenario of trade off factor for time was observed for both the data nature. The fast performer for
both the cases was Naïve Baye taking minimum CPU time and Wall time. Overall, ensemble
methods proved to increase overall accuracy of the prediction system for anemia in children.

27
 International Journal of Computer Science & Information Technology (IJCSIT) Vol 15, No 1, February 2023

4.5. Proposed Prediction Method

The Fig. 1 and Fig. 2 show our proposed prediction method for unbalanced and balanced data
respectively. For unbalanced data, Random Forest showed the highest accuracy. Stacking all the
six-classifier algorithm, bagging Random Forest 10 times, Decision Tree 20 times and adjusting
XGB Booster’s Learning Rate from 0.06 to 0.03 the accuracy was improved. However, for
balanced data the scenario was different. Accuracy was successful to increase by 0.3% when
Random Forest was bagged where number of random samples equals to 20, 30 and 40. For other
considered ensemble learning methods accuracy couldn’t increase more than the best performer.
For some cases accuracy was similar to the best performer for Prediction of Anemia in children
below 5 years of age.

Fig. 1. Proposed Prediction Model (Unbalanced Data)

28
 International Journal of Computer Science & Information Technology (IJCSIT) Vol 15, No 1, February 2023

Fig. 2. Proposed Prediction Model (Balanced Data

5. CONCLUSION
The research study aims to predict anemia in children below 5 years of age. From the literature
survey we found out that we could not find technical prediction for children. Moreover, the
prediction method has only considered the risk factor, social, economic and demographic factor
rather than giving emphasis for blood reports. So, we found a need for conducting our research
study to technically predict anemia for children considering the blood report.

We ran the pre-processed data into classifier algorithms and conducted experiments for
unbalanced as well as balanced data. For unbalanced data, results showed Random Forest with
accuracy of 98.4%. We performed various experiments to improve accuracy of the model than
the best performer. We conducted feature analysis from which we could not increase accuracy
more than that of the best performer. After that, we applied ensemble learning methods in our
experiment. The accuracy was increased up to 98.8% when applied with Stacking and Bagging.
The accuracy increased by 0.2% i.e. it reached 98.6% when stacking Random Forest with SVM
or Naïve Bayes, stacking all the six algorithms and bagging Decision Tree 20 times. Stacking
Random Forest with Logistic Regression or ANN increased the accuracy by 0.3% i.e. it was
increased by 98.7%. Accuracy was increased by 0.4% i.e. it reached 98.8% when Random Forest
was bagged and the number of random samples equals to 10. Then, applying Extreme Gradient
Boosting accuracy reached up to 99%. This was the case for unbalanced data. Then for balanced
data, accuracy was increased by 0.1% which reached 98.7% when Random Forest was bagged
where number of random samples equals to 20, 30 and 40. For other ensemble learning methods
accuracy couldn’t increase or accuracy remained the same.

29
 International Journal of Computer Science & Information Technology (IJCSIT) Vol 15, No 1, February 2023

Finally, we developed a new proposed prediction framework for both unbalanced as well as
balanced data which improves accuracy of existing algorithm. Therefore, we claim that selecting
the best performer algorithm, stacking with other algorithms, bagging it, boosting it are very
much crucial to improve accuracy despite of any time issue for prediction of anemia in children
below 5 years of age.

REFERENCES

[1] L. Wang, M. Li, S. E. Dill, Y. Hu, and S. Rozelle, "Dynamic Anemia Status from Infancy to
Preschool-Age: Evidence from Rural China," International journal of environmental research and
public health, vol. 16, no. 15, pp. 2761, 2019.
[2] V. Arun, V. Shyam, and S. K. Padma, "Privacy of health information in telemedicine on private
cloud," Int J Family Med Med Sci Res, vol. 4, no. 189, pp. 2, 2015.
[3] N. Soundarya and P. Suganthi, "A review on anaemia–types, causes, symptoms and their treatments,"
Journal of science and technology investigation, vol. 1, no. 1, 2017.
[4] N. Alli, J. Vaughan, and M. Patel, "Anaemia: Approach to diagnosis," SAMJ: South African Medical
Journal, vol. 107, no. 1, pp. 23-27, 2017.
[5] H. Bhavsar and A. Ganatra, "A comparative study of training algorithms for supervised machine
learning," International Journal of Soft Computing and Engineering (IJSCE), vol. 2, no. 4, pp. 2231-
2307, 2012.
[6] S. Uddin, A. Khan, M. E. Hossain, and M. A. Moni, "Comparing different supervised machine
learning algorithms for disease prediction," BMC Medical Informatics and Decision Making, vol. 19,
no. 1, pp. 1-16, 2019.
[7] M. Fatima and M. Pasha, "Survey of machine learning algorithms for disease diagnostic," Journal of
Intelligent Learning Systems and Applications, vol. 9, no. 01, pp. 1, 2017.
[8] M. Jaiswal, A. Srivastava, and T. J. Siddiqui, "Machine Learning Algorithms for Anemia Disease
Prediction," in Recent Trends in Communication, Computing, and Electronics, Springer, Singapore,
2019, pp. 463-469.
[9] P. T. Dalvi and N. Vernekar, "Anemia detection using ensemble learning techniques and statistical
models," in 2016 IEEE International Conference on Recent Trends in Electronics, Information &
Communication Technology (RTEICT), 2016, pp. 1747-1751.
[10] M. Abdullah and S. Al-Asmari, "Anemia types prediction based on data mining classification
algorithms," Communication, Management and Information Technology–Sampaio de Alencar (Ed.).
[11] J. R. Khan, S. Chowdhury, H. Islam, and E. Raheem, "Machine learning algorithms to predict the
childhood anemia in Bangladesh," Journal of Data Science, vol. 17, no. 1, pp. 195-218, 2019.
[12] P. Anand, R. Gupta, and A. Sharma, "Prediction of Anemia among children using Machine Learning
Algorithms," International Journal of Electronics Engineering, vol. 11, no. 2, pp. 469-480, 2019.
[13] J. Zhang and W. Tang, "Building a prediction model for iron deficiency anemia among infants in
Shanghai, China," Food Science & Nutrition, 2019.
[14] J. E. Ewusie, C. Ahiadeke, J. Beyene, and J. S. Hamid, "Prevalence of anemia among under-5
children in the Ghanaian population: estimates from the Ghana demographic and health survey,"
BMC public health, vol. 14, no. 1, p. 626, 2014.

AUTHORS

Prakriti Dhakal, Recent Graduate of Masters in Computer Engineering Department of Computer Science
and Engineering, Kathmandu University, Dhulikhel, Nepal

Santosh Khanal, Assistant Professor, Department of Computer Science and Engineering, Kathmandu
University, Dhulikhel, Nepal

Rabindra Bista*, Associate Professor, Department of Computer Science and Engineering, Kathmandu
University, Dhulikhel, Nepal ,*- Corresponding Author

30

View publication stats