Available online at www.sciencedirect.

com
Available online at www.sciencedirect.com
Available online at www.sciencedirect.com

ScienceDirect
Procedia Computer Science 00 (2024) 000–000
Procedia
Procedia Computer
Computer Science
Science 00(2024)
233 (2024)45–55
000–000 www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia

5th International Conference on Innovative Data Communication Technologies and Application

5th International Conference on Innovative Data Communication
(ICIDCA 2024) Technologies and Application
(ICIDCA 2024)
Insightful
Insightful Clinical
Clinical Assistance
Assistance for
for Anemia
Anemia Prediction
Prediction with
with Data
Data
Analysis and Explainable AI
Analysis and Explainable AI
Kasthuri.Eaa , S.Subbulakshmibb , Rajasree Sreedharancc
Kasthuri.E
a
, S.Subbulakshmi , Rajasree Sreedharan
Research Scholar, Department of Computer Science and Engineering
a Research Scholar, Department of Computer Science and Engineering
b AssistantProfessor, Department of Computer Science and Applications, Amrita School of Computing, Amritapuri, India
b AssistantProfessor, Department of Computer Science
c Associate Professor, Division of Nephrology, Department of and Applications,
Pediatrics, Amrita
Medical School
College of Computing,
of Wisconsin, Amritapuri,
Milwaukee, India USA
Wisconsin,
c Associate Professor, Division of Nephrology, Department of Pediatrics, Medical College of Wisconsin, Milwaukee, Wisconsin, USA
[email protected]
[email protected]

Abstract
Abstract
Anemia among young people has positive correlation with susceptibility and impede cognitive development.As per WHO 2019
Anemia
statistics among
global young people
prevalence ratehas
of positive
anemia correlation withThis
is 33 percent. susceptibility
work aimsand impedeanemia
to predict cognitive development.As
with per WHO
inferences to identify the2019
root
statistics global prevalence
cause of anemia. Predictionrate of anemia
is done is 33 percent.
with machine learningThis work aims
algorithms withtoproper
predictpre-processing,
anemia with inferences to identifyand
data normalization, theclass
root
cause of anemia. Prediction is done with machine learning algorithms with proper pre-processing, data normalization,
imbalance approaches. Explainable AI, a transparent model, is used to enhance the prediction which enables to draw inferences. and class
imbalance
LGM Boostapproaches. Explainable
learning model prediction AI,has
a transparent model,
a 91% accuracy andisexplainable
used to enhance the prediction
AI framework whichofenables
predictions to draw inferences.
most contributing attribute
LGM Boostwhen
is accurate learning model prediction
compared has a 91%
to other models. Thisaccuracy
approach and explainable
helps AI professionals
healthcare framework predictions of mostdecisions
make informed contributing attribute
on the most
is accurate
optimal when compared
treatment for anemia.to other models. This approach helps healthcare professionals make informed decisions on the most
optimal treatment for anemia.
© 2024
© 2024 The
The Authors.
Authors. Published
Published by
by Elsevier
Elsevier B.V.
B.V.
© 2024
This is The Authors. Published by Elsevier B.V.
This is an
an open
open access
access article
article under
under the
the CC
CC BY-NC-ND
BY-NC-ND license
license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
(https://creativecommons.org/licenses/by-nc-nd/4.0)
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the 5th International
International Conference on Innovative Data
Data Communication
Communication
Peer-review under
Technologies and responsibility of the scientific committee of the 5th International Conference on Innovative Data Communication
and Application
Application.
Technologies and Application.
Keywords: Explainable Artificial Intelligence; Machine learning; Anemia; Data preprocessing.
Keywords: Explainable Artificial Intelligence; Machine learning; Anemia; Data preprocessing.

1. Introduction
1. Introduction
Despite the considerable progress made in economic, healthcare, and social development on a worldwide scale,
theDespite
issue ofthe considerable
anemia continuesprogress made
to persist in economic,
prominently healthcare,
among women and and social development
children. on a worldwide
Anemia mostly occurs duescale,
to a
the issue of anemia continues to persist prominently among women and children. Anemia mostly
reduction in the number of erythrocytes inside the bloodstream [1]. Based on the global statistics provided occurs duebytothea
reduction in the
World Health number of erythrocytes
Organization for the year inside
2019, the
it isbloodstream
seen that the[1]. Based onoftheanemia
prevalence globalamong
statistics provided
women by age
in the the
World Health Organization for the year 2019, it is seen that the prevalence of anemia among
group of 15-49 years is around 29.9 percent. Anemia may lead to fatigue, short breath, conjunctivitis[2], mucouswomen in the age
group of 15-49
membranes, coldyears is around
extremities due 29.9 percent.ofAnemia
to reduction oxygen may
in thelead to or
blood fatigue, short breath,
poor circulation conjunctivitis[2],
of blood, mucous
dizziness, cognitive
membranes, cold extremities due to reduction of oxygen in the blood or poor circulation of blood, dizziness,
impairment, weakened immune system, rapid heartbeat, heart failures due to extra strain to heart, pre-term delivery cognitive
impairment, weakenedamong
and weight reduction immune system,The
childbirth. rapid heartbeat,
United heart
Nations failures
2030 dueSustainable
goal for to extra strain to heart, pre-term
Development delivery
has included the
and
objective of increasing the incidence of anemia among women aged 15-49. Anemia affects a significant proportionthe
weight reduction among childbirth. The United Nations 2030 goal for Sustainable Development has included of
objective of increasing the incidence of anemia among women aged 15-49. Anemia affects a significant proportion of

1877-0509 ©
1877-0509 © 2024 The Authors.
2024 The Authors. Published
Published by
by Elsevier
Elsevier B.V.
B.V.
1877-0509
This isisan © 2024 Thearticle
Authors. Published by Elsevier B.V.
This anopen
openaccess
access under
article the CC
under the BY-NC-ND
CC BY-NC-ND licenselicense
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
(https://creativecommons.org/licenses/by-nc-nd/4.0)
This is an open access article under the scientific
CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review
Peer-review under
under responsibility
responsibility of the of the committee
scientific ofcommittee
the of the Conference
5th International 5th International Conference
on Innovative on Innovative
Data Communication Data
Technologies
Peer-review under responsibility of the scientific committee of the 5th International Conference on Innovative Data Communication Technologies
Communication
and Application. Technologies and Application
and Application.
10.1016/j.procs.2024.03.194
46 Kasthuri.E et al. / Procedia Computer Science 233 (2024) 45–55
2 Kasthuri.E / Procedia Computer Science 00 (2024) 000–000

the worldwide population, with about one-third 33 percent being affected [3]. The accuracy was impacted due to the
use of a limited dataset that did not undergo data pretreatment processes. The study uses machine learning and deep
learning models to detect anemia using photos of conjunctiva, fingernails, and palms. Deep learning models have
shown greater prediction ability, but their ability to provide insights into the fundamental mechanisms contributing
to anemia is limited. Twitter analysis is used to predict anemia by identifying terms related to it and finding the
strength of those links. However, most people affected by anemia fall into economically backward categories. Most
research has provided more accurate predictions about the existence of anemia among individuals using varied sets
of data. Prediction would be useful if it combines the inference about the reason for the prevalence of anemia, as
it would act as a clinical aid for taking preventive actions to overcome this disease. The proposed research aims to
create a different prediction model enabled with Explainable AI that provides insights into the features that influence
anemia prediction, helping practitioners assess the reason for anemia and provide better treatment. The model is
created using a publically available tabular data set with nearly 15300 records collected over a period of 5 years. A
comparative study is conducted to assess the model’s accuracy and the impact of various pre-processing techniques
on the effectiveness of machine learning algorithms. A meticulous process is devised to assess performance, which
may be applied to other diagnostic models.
This article is ordered as follows. Chapter 2 Literature survey, describes existing work; Chapter 3, elaborates
Methodology with pre-processing, machine learning models, and explainable AI used in this work. Chapter 4
investigates Evaluation Metrics, while Chapter 5 provides an overview of the contributions, The work performed in
developing this anemia disease diagnosis system is outlined in Chapter 6, and Chapter 7 concludes the study on the
identification system for anemia and suggests potential areas for future improvement.

2. Literature Survey

Wemyss, T. A et.al., [9] developed a predictive technique for the identification of anemia in children and newborns
residing in Ghana. The objective of this study is to examine and assess several techniques for modifying the
surrounding illumination and assessing the degree of redness in images. The study primarily concentrated on a
specific area of interest and aimed to conduct colorimetric measurements from two distinct regions of interest, namely
the lower eyelid and the lower lip. The provided sample did not consist of a substantial percentage of persons with
severe anemia. As a result, the accuracy of predicting the severity of anemia was not improved. Research elaborated
in [8] aims to predict anemia among pregnant women in Ethiopia by using ensemble machine-learning algorithms.
The Ethiopian population health survey data underwent appropriate preprocessing techniques to improve the data
quality provided to the prediction model. The prediction model was constructed using decision trees, random forests,
cat boost, and severe gradient boosting techniques, both with and without class decomposition. In the study [4],
researchers investigated the relationship between symptoms of illness and the emotions expressed by patients on the
Twitter network. Although there has been accurately predicting anemia, the widespread adoption of these methods
across a broad consumer base has been unsuccessful. Iron deficiency is a prominent etiological factor contributing
to the development of anemia in humans. Kwak et al.[5] conducted a comprehensive investigation to illustrate the
relationship between iron deficiency and the incidence of seizures in children. The meta-analysis findings indicated
a moderate correlation between the variables under investigation. However, the experimental study conducted was
limited by many constraints, such as the exclusion of randomized controlled trials from the analysis. ThalPred [7]
was identified in a cohort of Thai people using a retrospective analysis of laboratory data. A discriminant model
was constructed to evaluate the presence of thalassemia and anemia diseases, using five distinct machine learning
techniques: k-nearest neighbour, decision tree, random forest, artificial neural network, and support vector machine.
The evaluation and comparison of performance were conducted to thirteen pre-existing discriminant equations and
indicators A novel methodology was developed for the non-invasive diagnosis of anemia, using a combined approach
strategy that integrates modified Kalman filtering (KF) with penalty regression. The Kalman Filter (KF) has under-
gone modifications to accommodate temporal variations in medical data. The use of the suggested Kalman Filter (KF)
[6] has the potential to decrease the variance of the dataset, hence potentially facilitating the subsequent task of pattern
identification. After conducting a comprehensive analysis of existing research studies on the prediction of anemia,
using diverse approaches and datasets, we have determined that our study would specifically concentrate on the
 Kasthuri.E et al. / Procedia Computer Science 233 (2024) 45–55 47
Kasthuri.E / Procedia Computer Science 00 (2024) 000–000 3

accurate prediction of anemia via the use of explainable artificial intelligence methods. Previous studies successfully
predicted the presence or absence of anemia in individuals, but they were unable to ascertain the underlying cause of
anemia. Work done by Yagin FH, Cicek ˙IB reveals the importance of Explainable AI for identifying COVID gene
biomarkers[? ]. With reference to the above research this study aims to provide clinical support in the identification
of the underlying etiology of anemia, enabling the administration of appropriate therapeutic interventions for its
resolution

3. Methodology

Anemia is a prevalent disease among children and women worldwide, with a 40%-52% chance of further seizures.
Epilepsy is one of the diseases most common among kids and it is a disorder where unprovoked seizures frequently
happen. Iron deficiency anemia is linked to more seizures in some groups[10]. An efficient anemia identification
system is essential for improving society’s healthcare system. The research aims to perform data analysis, create
models for anemia predictions, and generate inferences about the main causes using explainable AI. This will help
identify treatment processes for individuals with anemia and improve the overall healthcare system.

Fig. 1. The architecture of an Anemia Prediction System using Machine Learning Models and Explainable AI

3.1. Dataset description

The publicly accessible dataset which is created with the data obtained from Faculty of Medicine, Tokat
Gaziosmanpaşa University, Turkey related to anemia is used in this research work. It contains whole blood count
test results gathered from 15,300 individuals from 2013-2018 and it reveals that 1,019 (7%) of the patients had
HGB-anemia, 4,182 (27%) went through iron insufficiency, 199 (1%) encountered B12 deficiency, and 153 (1%) was
having folate deficit, while the majority, 9,747 (64%) did not have any anemia.

3.2. Data preprocessing and analysis

3.2.1. Outlier removal

Datapoints in a dataset that significantly deviate from other observations [11, 12] are called outliers. The acceptable
data range is between Q2 -1.5 * IQR(lower limit)and Q3 - 1.5 * IQR(upper limit) , where IQR is calculated us-
ing the formula given below. Points beyond this range which are considered to be outliers are dropped from the dataset.

IQR = Q3 − Q4 (1)
48 Kasthuri.E et al. / Procedia Computer Science 233 (2024) 45–55
4 Kasthuri.E / Procedia Computer Science 00 (2024) 000–000

3.2.2. Univariate analysis

Data Normalization: Normalization is a technique that transforms attributes to have a mean value of 0 and a stan-
dard deviation of 1, improving their effectiveness when the attribute distribution follows a Gaussian distribution[13].
It reduces bias, improves numerical stability, speeds up training, and simplifies feature comparisons. Each numeric
attributes in the dataset is normalised using the Z-score method, with the following equation:

 Y − Y mean 
Y new = (2)
S tandardDeviation
Skewness: Distribution curves are used to assess skewness in a dataset, a term referring to the asymmetry in the data
distribution curve when inclined towards either side. Skewness can affect model assumptions and performance. It
helps understand the spread of the data-points in the dataset[14]. Skewness in dataset should be corrected using cube
root transformation in order to produce normalized data. Skewness is calculated using a formula.

 
mean − median
S kewness = 3 (3)
S tandardDeviation

Kurtosis: It is used to illustrate the nature of data distribution in the data set. It represents the degree to which a
distribution tails are heavier than those of a normal distribution[15]. The formula for calculating the kurtosis is given
below where, represents an ith variable, mean of the distribution, n total number of variables in the distribution

 n 
 Σ (zi − z̄)4 
Kurtosis = n ∗  ni  (4)
Σi (zi − z̄2 )2

3.2.3. Bivariate analysis

Covariance: Variables can show a linear association [20], which can be explained by examining the covariance
between two variables and it does not reveal the variable’s relationship strength. When values are not centrally
located, they are calculated by the average of the product of each sample’s values. The computation is determined
using the below formula, in which A and B are variables.

1
Covariance(A,B) = S um(A − mean(a) ∗ (B − mean(b)) ∗ (5)
n−1
Correlation: It helps to assess the strength of relationship between the two variables. Positive correlations occur when
two variables pass through in the same direction. A negative correlation occurs when one variable rises and others
decline [19]. No correlation or association means no relationship between variables. To calculate, use the formula
below, where m and n are variables.

cov(m, n)
Correlation = (6)
σm ∗ σ n

Multicollinearity: Multicollinearity is significant when the variance inflation factor (VIF) for variables is rela-
tively large. It gives a quick approximation of how much a variable adds to the standard error of categorization.
There are several methods for reducing or combining collinear variables to address the multicollinearity problem. In
anemia dataset, we found multicollinearity and removed it with feature engineering. The formula for calculating VIF is

1
VIF = (7)
1 − R2i
 Kasthuri.E et al. / Procedia Computer Science 233 (2024) 45–55 49
Kasthuri.E / Procedia Computer Science 00 (2024) 000–000 5

3.3. Machine Learning Models

3.3.1. Logistic Regression

It is a classification method designed to determine the relationship between inputs and the probability of a specific
result. When diagnosing different medical disorders based on specific symptoms and traits, it is highly helpful [17].
The Logistic Regression is a kind of predictive analysis, just like other regression analysis. It determines the expected
likelihood of the result. By including the logit function, the influence of the outlier in a logistic regression is lessened.

3.3.2. XGBoost
XGBoost, also known as Extreme Gradient Boosting, is a machine-learning technique used to create gradient-
boosting decision trees (GBDT). It effectively manages sparsity patterns and tackles overfitting in complex models.
The methodology uses LASSO and Ridge regularization approaches and features built-in cross-validation capabilities
for selecting the most suitable number of boosting iterations. The distributed weighted quantile sketch approach is
used to determine the optimal number of split points among weighted datasets.

3.3.3. Decision tree

The hierarchical tree-structured approach is utilized to extract significant results from diverse inputs, contributing
significantly to decision-making processes[18]. The model construction process identifies hidden patterns within the
input data, determining information gain, index, and dividend growth rate based on optimal partitioning of nodes. The
dataset is divided based on the parameters demonstrating the lowest entropy or highest gain in data, ensuring accurate
and efficient data analysis.

3.3.4. Random Forest

It is a meta-estimator fitting several decision tree classifiers on various dataset subsamples and uses averaging
to improve prediction accuracy and control over-fitting. A random forest with a specific number of trees is created
with any supervised classification techniques. The classifier won’t overfit the model if the forest has more trees. The
ensemble learning technique classifies outcomes based on individual trees and uses many decision trees to train. The
random forest reduces the likelihood of overfitting by using many trees.

3.3.5. LGM Boost

The algorithm discussed is an enhanced iteration within the Gradient Boosting Decision Tree (GBDT) framework,
utilizing histogram-based approaches. It divides continuous eigenvalues into smaller intervals and creates a histogram,
identifying the most suitable points for feature segmentation. This technique reduces computational time and memory
consumption[21], making it promising for real-time anemia prediction due to its strong theoretical foundations in data
feature extraction and parallel processing, the LightGBM method has promise for real-time anemia prediction.

3.4. Explainable AI

Explainable artificial intelligence is increasingly important in the healthcare sector to make machine learning mod-
els less opaque[22, 23]. There are two main methods to explain AI models: model-specific, which is limited to specific
versions, and model-agnostic, which applies to all types. Early AI adoption focused on building human-understandable
models, rather than providing operational descriptions, and now our aim main objective is to make ML models more
interpretable.

3.4.1. Locally interpretable model agnostic explanations (LIME)

It is a white box model providing interpretations to clearly explain individual predictions. LIME acts locally which
is observation-specific and provides explanations for the prediction about each observation[24]. It aims to fit a local
model using sample data points that are comparable to the observation being explained. This can fall under the um-
brella of interpretable models, which also cover linear models, decision trees, etc. Feature selection can be done using
explainable AI for increasing accuracy. LIME explanations for each observation c is calculated as follows:
∅(c) = argminBb∈B (L(a, b, πd) + Ω(b)) (8)
50 Kasthuri.E et al. / Procedia Computer Science 233 (2024) 45–55
6 Kasthuri.E / Procedia Computer Science 00 (2024) 000–000

where B is the class of potentially interpretable models, bϵB: a statement that is considered as a model.πd (e): Proximity
measure of an instance e from D. Ωb (b): a measure of the explanation’s complexity bϵB. Such an explanation helps to
interpret the results and aids in making appropriate decisions relevant to the specific situations.

4. Evaluation Metrics

The evaluation of the efficacy of various machine learning models is conducted via the assessment of their accuracy,
precision, recall, and F1-score. All the above evaluation scores are retrieved from the confusion matrix generated for
every model separately. It takes into consideration classification results of both training and test dataset. It considers
correct and wrong predictions made by the models. Accuracy is a metric that quantifies the proportion of correctly
classified samples to the total number of samples inside a specified model. Precision is a quantitative measure that
evaluates the proportion of correctly detected positive cases among all positive predictions made by various models.
The classification of recall is established based on the proportionate distribution of true positives and false negatives
in the whole dataset. The F1-score may be computed by obtaining the weighted mean of accuracy and recall.

5. Contribution

i SMOTE techniques and hyper-parameter adjustment to address class disparity and improve disease prediction.
ii LIME is an Explainable AI algorithm developed to provide physicians with a comprehensive understanding of
disease factors, including feature importance, visualization, and decision-making importance.
iii An explainable AI technique is used to determine the model’s prediction outcome, making it easier to identify
key factors.
iv Internal explainability is also utilized to enhance model reliability, balancing accuracy and interpretability. This
approach is beneficial for physicians and healthcare professionals.

6. Result and analysis

The experiment of this work is carried out with a data set of nearly 15,300 data records. The relevance of work
evaluated on datasets without preprocessing and with preprocessing is explained in the following section.

6.1. Anemia Classification without Preprocessing

The original dataset without preprocessing fed into the models gives 100% accuracy for Logistic Regression,
XGBoost, and LGM Boost as shown in Table 1. Decision tree and Random forest have an accuracy of 93% and 98%
respectively. 100% accuracy reveals the presence of overfitting issues in those models. Class label imbalance might
have led to this issue and this work performs preprocessing on the dataset to overcome it.

Table 1. Classification Results using dataset without and with Preprocessing

Comparison Results without preprocessing Results with preprocessing
ML Models Precision Recall F1-Score Accuracy Precision Recall F1-Score Accuracy
Logistic reg. 100 100 100 100 73.6 74.2 73.8 74
Decision tree 78.6 79.4 79.2 93 79.8 80 79.8 80
XGBoost 100 100 100 100 85.2 85.6 85.4 86
Random forest 97.6 84.6 89.6 98 87 87.2 86.8 87
LGM Boost 100 100 100 100 90.8 90.8 90.6 91
 Kasthuri.E / Procedia Computer Science 00 (2024) 000–000 7
Kasthuri.E et al. / Procedia Computer Science 233 (2024) 45–55 51

Fig. 2. (a) Unbalanced Data without SMOTE; (b) Balanced Data with SMOTE.

Fig. 3. (a) Before skewness; (b) After skewness.

6.2. Preprocessing in Anemia Dataset

6.2.1. Imbalance class

When one class label has fewer samples than the other, that is known as an imbalance class. In this dataset, the class
label no anemia has higher data than all other classes. It is challenging to train the model and generate trustworthy
user evaluation predictions because of the class imbalance.
The SMOTE method is used to select instances closest to the feature space by drawing a new instance along the
line connecting them. To ensure balanced classes, a representative sample from a minority class is randomly selected,
followed by the creation of neighboring instances for the same in a synthetic manner. Fig.2a and Fig. 2b show the
imbalance dataset and SMOTE balanced dataset.
Exploratory data analysis techniques are used for outlier removal, data normalization, skewness, kurtosis, covari-
ance, correlation, and variance inflation factor. Fig. 3a with box plot shows the dataset with positive skewness in
the variable FERRITTE. It is eliminated through cube root transformation and its result is shown in Fig. 3b. Multi-
collinearity is identified using the variance inflation factor and eliminated using feature engineering. After the above
preprocessing steps overfitting is removed. Anemia dataset after preprocessing is used for training and testing of ML
models to produce classification outputs. Comparative results of the model used for classification on non preprocessed
dataset exhibits the existence of overfitting problem and it is shown in Fig. 4a. This problem is overcome with prepro-
cessing and the results of those models are shown in Fig 4b. LGM Boost produces the highest score for the evaluation
metrics Precision, Recall, and F1 score. Model output is given to LIME explainable AI which provides inferences
for the clinical professionals about the main factor causing the prediction output and it is depicted in Fig. 5. The
52
8 Kasthuri.E
Kasthuri.E et al. /Computer
/ Procedia Procedia Computer
Science 00Science 233 (2024) 45–55
(2024) 000–000

Fig. 4. (a) Accuracy of ML Models with and without Preprocessing; (b) ML models efficiency for preprocessed data.

Fig. 5. Anemia Prediction System Using Explainable AI Visualization

Fig. 6. (a) Explainable AI for Logistic regression with interpretation ; (b) Explainable AI for Random forest with interpretation.

explanation for LIME in explainable AI for the 9th instance in the dataset is shown in Figures 6a,6b,7a,7b and 8. It
confirms that the class label’s prediction probabilities are 4 in all ML models with the use of Explainable AI. Most
machine learning algorithms use a black box approach, whereas Explainable AI uses a white box method to explain
why the machine learning system predicts in this way. Explainable AI shows the interpretation regarding the attributes
that have more impact in selecting the specific class label. In short, it demonstrates the main causes for that predicted
anemia type. It may help the doctor in providing proper treatment based on this information.
Anemia prediction with Explainable AI results reveals that variables SDTSD squared, LY squared, TSD squared,
MPV squared,EO are listed in one or two model inferences, and RBC Squared and PWD are listed in 4 ML models.
Inference made with LGMBoost model which has the highest accuracy included the above 2 variables RBC squared
 Kasthuri.E et al. / Procedia Computer Science 233 (2024) 45–55 53
Kasthuri.E / Procedia Computer Science 00 (2024) 000–000 9

Fig. 7. (a) Explainable AI for Decision tree with interpretation ; (b) Fig.8.Explainable AI for XGBoost with interpretation.

Fig. 8. Explainable AI for LGBoost with interpretation

and PWD. So, clinical professionals can decide their treatment based on those 2 variables, so that anemia can be
cured easily as their treatment process will be more specific and accurate.

7. Conclusion

Anemia, a common and potentially life-threatening illness, can be identified and predicted using machine learning
algorithms. These models can assess patient data, medical history, demographic information, and behavioral variables
to identify those at risk but successful implementation requires careful data quality analysis and model interpretability.
Continuous validation and development are necessary to ensure accuracy and usefulness in clinical settings. The use
of machine learning and explainable AI in anemia prediction is a significant step forward in healthcare, potentially
improving early diagnosis, and patient outcomes, and reducing the burden of this common illness. The dataset is fine-
tuned with EDA and five machine-learning models are used to predict the type of anemia. LGMBoost predicts anemia
with the highest accuracy and the inference of the variables predicted by this model with explainable AI were listed
in 4 out of 5 models used for prediction.More research on this subject would be required since the scope of this study
is limited to the development and deployment of a mobile application to increase its utilization. In order to enhance
accessibility and usefulness, future research endeavors may include developing a mobile application that integrates
the recommended models and is suitable for rural populations.
 54 Kasthuri.E et al. / Procedia Computer Science 233 (2024) 45–55
10 Kasthuri.E / Procedia Computer Science 00 (2024) 000–000

References

[1] Asare JW, Appiahene P, Donkoh ET, Dimauro G. Iron deficiency anemia detection using machine learning models: A comparative study of
fingernails, palm and conjunctiva of the eye images. Engineering Reports. 2023 Feb 6:e12667
[2] Jaiswal M, Srivastava A, Siddiqui TJ. Machine learning algorithms for anemia disease prediction. InRecent Trends in Communication, Com-
puting, and Electronics: Select Proceedings of IC3E 2018 2019 (pp. 463-469). Springer Singapore.
[3] Meena K, Tayal DK, Gupta V, Fatima A. Using classification techniques for statistical analysis of Anemia. Artificial intelligence in medicine.
2019 Mar 1;94:138-52.
[4] Sarsam SM, Al-Samarraie H, Alzahrani AI, Shibghatullah AS. A non-invasive machine learning mechanism for early disease recognition on
Twitter: The case of anemia. Artificial Intelligence in Medicine. 2022 Dec 1;134:102428.
[5] Kwak BO, Kim K, Kim SN, Lee R. Relationship between iron deficiency anemia and febrile seizures in children: A systematic review and
meta-analysis. Seizure. 2017 Nov 1;52:27-34.
[6] Chen YM, Miaou SG. A Kalman filtering and nonlinear penalty regression approach for noninvasive anemia detection with palpebral conjunc-
tiva images. Journal of healthcare engineering. 2017 Jul 30;2017.
[7] Laengsri V, Shoombuatong W, Adirojananon W, Nantasenamat C, Prachayasittikul V, Nuchnoi P. ThalPred: a web-based prediction tool for
discriminating thalassemia trait and iron deficiency anemia. BMC medical informatics and decision making. 2019 Dec;19:1-4.
[8] Dejene BE, Abuhay TM, Bogale DS. Predicting the level of anemia among Ethiopian pregnant women using homogeneous ensemble machine
learning algorithm. BMC Medical Informatics and Decision Making. 2022 Dec;22(1):1-1.
[9] Wemyss TA, Nixon-Hill M, Outlaw F, Karsa A, Meek J, Enweronu-Laryea C, Leung TS. Feasibility of smartphone colorimetry of the face as
an anaemia screening tool for infants and young children in Ghana. Plos one. 2023 Mar 3;18(3):e0281736.
[10] Sharathappriyaa V, Gautham S, Lavanya R. Auto-encoder based automated epilepsy diagnosis. In2018 International Conference on Advances
in Computing, Communications and Informatics (ICACCI) 2018 Sep 19 (pp. 976-982). IEEE.
[11] Krithika R, Narayanan J. Learning to grade short answers using machine learning techniques. InProceedings of the Third International Sym-
posium on Women in Computing and Informatics 2015 Aug 10 (pp. 262-271).
[12] Zhang P, Jia Y, Shang Y. Research and application of XGBoost in imbalanced data. International Journal of Distributed Sensor Networks. 2022
Jun;18(6):15501329221106935.
[13] Aditya A, Vinod R, Kumar A, Bhowmik I, Swaminathan J. Classifying Speech into Offensive and Hate Categories along with Targeted
Communities using Machine Learning. In2022 International Conference on Inventive Computation Technologies (ICICT) 2022 Jul 20 (pp.
291-295). IEEE.
[14] Md AQ, Kulkarni S, Joshua CJ, Vaichole T, Mohan S, Iwendi C. Enhanced Preprocessing Approach Using Ensemble Machine Learning
Algorithms for Detecting Liver Disease. Biomedicines. 2023 Feb 16;11(2):581.
[15] Ribeiro MT, Singh S, Guestrin C. ” Why should i trust you?” Explaining the predictions of any classifier. InProceedings of the 22nd ACM
SIGKDD international conference on knowledge discovery and data mining 2016 Aug 13 (pp. 1135-1144).
[16] Emsley D. Variance analysis and performance: two empirical studies. Accounting, Organizations and Society. 2000 Jan 1;25(1):1-2.
[17] Aravind T, Reddy BS, Avinash S, Jeyakumar G. A comparative study on machine learning algorithms for predicting the placement information
of under graduate students. In2019 Third International conference on I-SMAC (IoT in Social, Mobile, Analytics and Cloud)(I-SMAC) 2019
Dec 12 (pp. 542-546). IEEE.
[18] Tallapureddy G, Radha D. Analysis of Ensemble of Machine Learning Algorithms for Detection of Parkinson’s Disease. In2022 International
Conference on Applied Artificial Intelligence and Computing (ICAAIC) 2022 May 9 (pp. 354-361). IEEE.
[19] Asiyah S, Tritoasmoro II, Sa’idah S. Anemia Detection Through Conjunctiva on Eyes Using Principal Component Analysis Method and
K-Nearest Neighbor. In2022 8th International Conference on Science and Technology (ICST) 2022 Sep 7 (Vol. 1, pp. 1-5). IEEE.
[20] Browne MW. Covariance structures. Topics in applied multivariate analysis. 1982 Apr 22:72-141.
[21] Cui Z, Qing X, Chai H, Yang S, Zhu Y, Wang F. Real-time rainfall-runoff prediction using light gradient boosting machine coupled with
singular spectrum analysis. Journal of Hydrology. 2021 Dec 1;603:127124.
[22] Tantithamthavorn CK, Jiarpakdee J. Explainable ai for software engineering. In2021 36th IEEE/ACM International Conference on Automated
Software Engineering (ASE) 2021 Nov 15 (pp. 1-2). IEEE.
[23] Samek W, Müller KR. Towards explainable artificial intelligence. Explainable AI: interpreting, explaining and visualizing deep learning.
2019:5-22.
[24] Çelik TB, İcan Ö, Bulut E. Extending machine learning prediction capabilities by explainable AI in financial time series prediction. Applied
Soft Computing. 2023 Jan 1;132:109876.
[25] ] Yagin FH, Cicek İB, Alkhateeb A, Yagin B, Colak C, Azzeh M, Akbulut S. Explainable artificial intelligence model for identifying COVID-19
gene biomarkers. Computers in Biology and Medicine. 2023 Mar 1;154:106619.

Acknowledgements

First and foremost, we would like to express our gratitude to my Guru Sri Mata Amritanandamayi Devi, Chancellor
of Amrita Vishwa Vidyapeetham, for her unconditional love and blessings. Dr. Subbulakshmi S, Assistant Professor
of Computer Science and Applications at Amrita School of Engineering in Amritapuri, and Rajasree Sreedharan, MD.
 Kasthuri.E et al. / Procedia Computer Science 233 (2024) 45–55 55
Kasthuri.E / Procedia Computer Science 00 (2024) 000–000 11

Associate Professor, Division of Nephrology, Department of Pediatrics, Medical College of Wisconsin, Milwaukee,
Wisconsin USA has been an immense help throughout the research work.