Diman
Siddiq Hassan
Computer
Science Department, College of Science, University of Zakho, Zakho, Kurdistan
Region, Iraq-
Corresponding
author email:diman.hassan@uoz.edu.krd
ABSTRACT:
Developing countries often face a high
incidence of breast cancer, making early detection vital for effective
treatment. The risk of developing breast cancer can be evaluated using machine
learning methods and regular diagnostic data. In cancer datasets, there is a
wealth of patient information, but not all of it is valuable for predicting
cancer. This highlights the significance of feature selection methods in
uncovering the relevant data. In this field, many studies have attempted to
predict the different types of breast tumours, since it is important to
diagnose breast cancer medication accurately. This paper aims to
perform a comparison such that to show the effect of different feature
selection methods on the accuracy of various existing machine learning
algorithms. The study focuses on seven machine learning algorithms: K-Nearest
Neighbors (KNN), Naive Bayes (NB), Decision Trees (DT), Support Vector Machines
(SVM), Logistic Regression (LR), Neural Network (NN), and Random Forest (RF).
The feature selection techniques examined include F-test Feature Selection,
Mutual Information (MI), and Spearman Correlation Coefficient. The dataset used
for the experiments is the Wisconsin Diagnostic Breast Cancer (WDBC) dataset,
which is publicly available from the UCI Repository. The findings reveal that
when feature selection is implemented, the LR and NN algorithms demonstrate
superior accuracy and perform exceptionally well across other metrics compared
to the other models.
KEYWORDS: Breast Cancer; Machine Learning; Feature Selection; Breast Cancer
Diagnostic Dataset.
1.
INTRODUCTION
Cancer is one of the deadliest
diseases in the world. The latest statistics about this disease were reported
in 2023 (Zhou et al., 2024), listing
ten types of cancers, including breast cancer diagnosed in women. Breast cancer
has been and still is the most common type of cancer that has affected a high
percentage of women around the world at approximately 31%. It is considered the
first type of cancer that causes deaths in women and is ranked fifth in terms
of all cancer deaths around the world. It was the reason for 685,000 deaths in
2020, and that number increased to around 963,000 deaths in 2021, exceeding
lung cancer with approximately 2.3 million new cases of this disease, according
to the World Health Organization (WHO) (Bray et al.,
2024). The percentage of these cancer cases was 25%, and the death cases
among women were 17% around the world (Zhou et al.,
2024). The abnormal growth of the breast cell is
called a tumour, which is divided into two types: malignant and benign. The former
is cancerous, while the latter is
non-cancerous. Despite the incomprehension of the causes of breast cancer in
women, several factors and attributes were contributed as the reasons for this
disease, such as family history, problems in the inside
uterine environment, adolescent exposures, pregnancy problems, gene mutation, alcohol
and tobacco consumption, and childbearing at advanced maternal ages,
specifically in developing countries (Uddin
et al., 2023).
Consequently, to reduce the rate of
breast cancer cases and to prevent mortality in women, it is important to make regular visits to health professionals for screening, treatment, and
accurate examination in clinical health. However, misdiagnosis may occur, which
reduces the opportunity for early recovery, or it may as well be that there is
a shortage in the number of health experts. Also, the medical examining of the
tumour is time-consuming and costly. Therefore, implementing techniques such as
Machine Learning (ML) that automatically diagnose breast cancer is crucial.
There are different examples of ML classification techniques that have been
used to determine whether the breast tumour is cancerous or not, such as Support
Vector Machine (SVM), Naïve Bayes (NB), and Logistic Regression (LR), among
others (Lappeenranta-, 2023) (Ak, 2020).
Furthermore, many researchers have
used different ML classification algorithms for the prediction of breast cancer,
underlining the
importance of using such techniques for predicting the disease and showing
challenges in this field (Ak, 2020; Mohammed
et al., 2020; Nemade et al.,
2022; Abunasser et al., 2023; and Ebrahim et al., 2023). Whereas some others,
namely (Chen et al., 2023) (Botlagunta et al., 2023) and (Laghmati et al., 2024) have analyzed
different breast cancer datasets, such as the Wisconsin Original Breast Cancer
and Wisconsin Diagnostic Breast Cancer datasets (WDBC)Click or tap here to enter text. for
this purpose and have obtained significant results
(Wolberg, 1995).
In this paper, the primary goal is
to implement different ML techniques to classify the patients as having cancer
or not using the WDBC dataset and to obtain the accuracies of the models. Then,
the following goal is to explore the influence of using the F-test, Mutual
Information (MI), and the Spearman correlation coefficient feature selection
techniques on the accuracy of the selected ML. This can be accomplished by
comparing the results of both implementations as well as comparing the results
obtained from implementing the feature selection methods with each other.
Seven different machine learning algorithms, K-Nearest Neighbors (KNN), Naive
Bayes (NB), Decision Trees (DT), Support Vector Machines (SVM), Logistic
Regression (LR), Neural Networks (NN), and Random Forest (RF) are used for this
investigation. Furthermore, some other methods are used to improve the
performance of the selected models. For instance, Synthetic Minority
Over-sampling (SMOTE) (Chawla et al., 2002), is used to prevent the
issue of imbalanced classes, using feature scaling to ensure that all features
contribute equally to the models, and cross-validation to reduce overfitting
and enhance model performance. The performance of the ML models is evaluated
using different evaluation metrics such as accuracy, F1 score, precision,
recall, ROC AUC, and Matthew’s correlation coefficient (MCC).
The rest of the paper is organized
as follows: in Section 2, the most relevant works to this study are included, specifically
those publications that have compared different ML methods using the WDBC
dataset. The methodology and the applications used in the study are presented
in Section 3. Section 4 illustrates the results of this study both prior to and
after the implementation of the feature selection methods. Section 5 compares
the outcomes of implementing the selected feature selection methods for each of
the seven models used in this study. Section 6 is the last section that includes
the conclusion and future works.
2.
RELATED WORKS
Recently, with advances in medical
research, different ML algorithms have been suggested to assess the classification
of breast cancer data. Breast cancer is one of the common medical data that researchers
have used for this purpose. These data can be obtained from breast cancer data
repositories. In this section, a review of the publications related to the
prediction of breast cancer, specifically the Wisconsin Diagnostic Breast
Cancer (WDBC), is presented and surveyed, as shown in Table 1.
Ak, M. F. (2020) accomplished a
comparative study to analyze the performance of different
machine learning techniques, LR, KNN, SVM, and DT, using a graphical program
named CITY for data visualization and samples from breast cancer patients in the
WDBC dataset. The results of the study indicated that LR outperformed the other
techniques, with the highest classification accuracy at 98.1%. In (Mohammed et al., 2020), the performance of
three different ML algorithms, DT, NB, and Sequential Minimal Optimization
(SMO), was compared using two different breast cancer datasets: the Wisconsin
breast cancer and the Original Wisconsin Breast Cancer datasets (WBC). The
study included a number of pre-processing steps to improve the performance of
the ML techniques further, such as discretization and
removing records that have missing data. The results showed that the algorithm SMO outperformed the other two classifiers with an accuracy
of 99.56% on the WBC dataset.
Chaurasia et
al. (2020) proposed a new method called
Mode to remove frequent features from the WDBC and then applied an ensemble
technique with stacking classifiers to categorize
records with all features in comparison to the reduced data subset and to enhance accuracy. Their results showed that their proposed
method increased the breast cancer accuracy to over 90%. Moreover, Islam et al. (2020) aimed to compare five ML
algorithms: SVM, KNN, RF, ANN, and LR to diagnose breast cancer using the Wisconsin Breast Cancer dataset.
The results of their study revealed that ANN outperformed other techniques,
achieving the highest accuracy, 98.57%.
Naji et al. (2021),
on the other hand, explored the ability of five ML algorithms to predict cancer
in the WDBC dataset. Their results showed that the SVM technique surpassed
other models by obtaining the highest accuracy, 97.2%. The authors revealed
that the prediction of breast cancer using ML algorithms is possible; however,
they acknowledged limitations and planned to explore larger datasets for
improved accuracy and ethical implications.
Furthermore,
Ara et al. (2021) explored ML algorithms for
categorizing breast tumours as cancer or not using the WDBC Dataset. Training
and testing techniques were used in the study, and the number of features was
reduced, keeping only the highly correlated
features to the target to improve the model’s performance. Among the ML
techniques used, the researchers concluded that RF and
SVM models outperformed the other models by obtaining an accuracy of 96.5%.
Sakib et al. (2022), used
two types of prediction techniques to predict and diagnose breast cancer in the
WDBC: ML and Deep Learning (DL). They used different evaluation
metrics to assess the performance of the models used for classification. The
metrics used were accuracy, recall, specificity, precision, false-negative rate
(FNR), false-positive rate (FPR), F1-score, and Matthews Correlation
Coefficient (MCC). The results showed that the performance of the RF classifier
was the highest based on the accuracy obtained, 96.66%. Chen et al. (2023),
studied machine learning algorithms — XGBoost, random forest, logistic
regression, and K-Nearest Neighbour (KNN) — for breast cancer classification,
emphasizing recall for early detection. Using a dataset named WDBC
from the UCI repository, they applied Z-score standardization and Pearson
correlation for feature selection and addressed data imbalance through
hierarchical sampling. Evaluating model performance with 80:20 and 70:30
splits, the XGBoost model outperformed others at 80:20, achieving a recall of
100%, precision of 96.0%, accuracy of 97.4%, and F1-score of 98.0%. The study
noted performance variability across splits and the limitations of a universal
machine learning approach in diagnostics.
The literature uses various methods and preprocessing
procedures to compare and obtain the best performance of the models used. Their
results are satisfactory in terms of accuracy and other metrics obtained.
However, in this study, several different experimental settings are implemented
to compare various machine learning algorithms and assess the impact of feature
selection methods on model performance, making the findings of this study
different from existing ones.
Table 1: A Survey of the Related Research
Used in This Study
|
Reference/ Authors
|
ML Algorithms
|
Feature Selection Methods
|
Dataset
|
Methodology Used
|
Results
|
|
(Ak, 2020)
|
LR, KNN, SVM, and DT
|
Features selected by
creating 3 datasets
1: data with all the
features
2: with highly
correlated features
3: with low correlated
features
|
WDBC
|
A comparative analysis
and new data visualization technique (CITY)
|
Accuracy: 98.1% for LR
for dataset1
97.4 for the dataset2
and 95.6% for the dataset 3
|
|
(Mohammed et al., 2020)
|
DT, NB, and SMO
|
None
|
WBC and original breast
cancer dataset
|
A comparison, discretization, and removing records with
missing data
|
Accuracy: 99.56% for SMO
using WBC dataset
|
|
(Chaurasia et al., 2020)
|
ensemble technique: AdaBoost, Gradient Boosting
Classifier, RF, Extra Tree (ET)
Bagging and Extra Gradient Boost (XGB).
stacking classifiers LR, DT, SVC, KNN, RF and NB
|
Statistical method of
feature selection ‘Mode’ to reduce the dataset to have 12 features only out
of 32 features
|
WDBC
|
proposed method named
Mode to reduce the dataset features
|
Accuracy over 90%
|
|
(Islam et al., 2020)
|
SVM, KNN, RF, ANN and LR
|
None
|
WBC
|
A comparison study
|
Accuracy: 98.57%. for ANN, precision of
97.82% and F1 score of 98.90%
|
|
(Naji et al., 2021)
|
SVM, RF, LR, DT and KNN
|
Feature extraction
method with no details
|
WDBC
|
A comparison study
|
Accuracy: 97.2% for SVM
|
|
(Ara et al., 2021)
|
SVM, LR KNN, DT, NB and RF
|
All the dataset features
used
|
WDBC
|
A comparison study and
the correlation between different features of the
dataset has been
analyzed for feature selection
|
Accuracy: 96.5% for RF and
SVM
|
|
(Sakib et al., 2022)
|
SVM, DT, LR, RF,
KNN, and a DL for classification using
cross-validation.
|
None
|
WDBC
|
A comparative study
|
Accuracy: 96.66% for RF
|
|
(Chen et al.,
2023)
|
XGBoost, RF, LR, and KNN
|
Z-score for
standardization and Pearson correlation for feature selection
|
WDBC
|
Predicting and
classifying along with data preprocessing and feature selection
|
Accuracy of 97.4%,
Recall of 100%, precision of 96.0%, and F1-score of 98.0%.
|
3.
METHODS AND APPLICATIONS
In this section, the dataset used
for conducting the experiments is explained. Then, each of the ML techniques
used in this study is presented, followed by the section presenting the feature
selection methods used. All of these techniques are presented to show the most
effective combinations of them for predicting breast cancer and to evaluate the
performance obtained using such combinations. The evaluation is accomplished
using the evaluation metrics to assess the effectiveness of each model to
provide valuable insight into enhancing predictive capabilities for breast
cancer diagnosis and treatment.
Dataset
Description
The dataset used in this study is called
Breast Cancer Wisconsin Diagnostic (WDBC) (Wolberg,
1995). It can be accessed from UCI Machine Learning Repository. WDBC
dataset consists of 569 samples designed for binary classification that is
distributed between 357 benign and 212 malignant breast tumours collected from
fine needle aspiration (FNA) biopsy images. In other words, the distribution of
the dataset is 62.7% non-cancerous and 37.3% cancerous breast lesions. The dataset
has 30 attributes, which represent the measurements of the shape of cell
nuclei, such as radius, roughness, and smoothness. Table 2 shows the dataset’s
features and their description as illustrated by (Kumar
et al., 2021). This paper uses this dataset because its
attributes can describe the symptoms effectively. Therefore, it is considered
as a good resource for diagnosing breast cancer and examining the feature
selection and model performance. However, the classes of the WDBC dataset are
imbalanced; therefore, they need to be resampled. In this work, the Synthetic
Minority Over-sampling (SMOTE) statistical technique is used for resampling the
dataset, and the results are shown in Figure 1. The figure shows the
distribution of the attributes to the target before and after resampling, which
is equal to 50% for each class.
Moreover, a smaller dataset is created by
removing the attributes that have a weak correlation to the target to avoid
noise and the models’ inaccurate prediction, resulting in each record having a
patient ID, a diagnosis, and 23 real-valued attributes. To ensure that all the
features of the new dataset contribute equally to the prediction of the models,
the attributes are scaled using the feature scaling method ‘Standard Scaler’ to
have zero mean and a standard deviation.
Table
2: Summary of Wisconsin Diagnostic Breast Cancer (WDBC) Dataset (Kumar et al., 2021).
|
|
Measurement range
|
|
|
Attributes
|
Mean
|
Standard Deviation
|
Maximum
|
Attribute description
|
|
Radius
|
6.99–28.12
|
0.121–2.923
|
7.95–37.01
|
Calculated
as the average of distances from the center to points on the perimeter
|
|
Texture
|
9.80–40.02
|
0.37–4.90
|
112.10–50.01
|
Calculated
as the standard deviation of Gray-scale values.
|
|
Perimeter
|
44.02–189.09
|
0.80–22.01
|
50.48–252.03
|
The
total distance between consecutive points in a contour or outline.
|
|
Area
|
144.04–2503.01
|
6.90–543.10
|
186.01–4255.00
|
Calculated Number of conductive points in
an outline
|
|
Smoothness
|
0.054–0.164
|
0.003–0.035
|
0.072–1.102
|
calculated
as the local variation in radius lengths
|
|
Compactness
|
0.020–0.350
|
0.002–0.138
|
0.030–1.060
|
Calculated
as the ratio of perimeter squared to area minus 1
|
|
Concavity
|
0.001–0.501
|
0.000–0.400
|
0.000–1.255
|
The
severity of concave portions of the contour.
|
|
Concave
Points
|
0.0001–0.202
|
0.000–0.055
|
0.000–1.296
|
Number
of concave portions of the contour.
|
|
Symmetry
|
0.108–0.305
|
0.009–0.080
|
0.158–0.668
|
|
|
Fractal
dimension
|
0.051–0.098
|
0.001–0.031
|
0.057–0.210
|
Coastline
approximation minus 1
|
Figure
1: The Distribution of the WDBC before and after Resampling across the Target
Variable.
Machine Learning Algorithms
Decision
Tree (DT): the
DT algorithm (Mohammed et
al., 2020) is a supervised ML algorithm that is
mainly used for classification and regression. The input node is the main
feature of this technique. Its structure consists of a root node, where it is
at the top of the tree, an internal node representing the input features, and a
leaf node representing the decision node or the class of the dataset located at
the bottom of the DT, as shown in Figure 2. The hierarchical structure of the
DT is made up of a number of nodes at different levels, and the small trees
that can be extracted from the main tree are called subtrees. The larger the
tree, the more difficult the classification of data accurately is due to
problems such as overfitting and data splitting. These problems can be tackled
by using techniques such as pruning, cross-validation, and ensemble techniques
to integrate multiple trees.
Figure
2: The Structure of Decision Tree Algorithm (Mohammed et al., 2020)
Logistic Regression (LR): Logistic Regression (LR) (Ak,
2020; Dhanya R, 2019; and Hossin et al., 2023) is an ML technique used
for predicting two values, 0 or 1, and classifying data derived from a linear
combination of data. The parameter coefficient values can be calculated using
both linear regression and logistic regression. In the case of logistic
regression, gradient descent can be used for this purpose. The LR algorithm
utilizes some techniques to overcome the problems of overfitting and bias, such
as cross-validation and regularization. Generally, LR is a simple and strong
technique in solving classification problems.
Naïve Bayes (NB):
Naive
Bayes (NB) is a strong supervised classification
technique used for classifying large and complex data using a small size of
training data (Dhanya R, 2019; Hossin et al., 2023; and Kadhim
et al., 2023). It is an easy method and is based
on the theorem called Bayes, which assumes conditional independence between
each two features and a given class. NB calculates the probability theory in a
simpler way. It can also tackle the risk of data noise and overfitting for its
reliance on strong independence assumptions. The NB equation can be represented as follows:
(1)
Since
the features of WDBC data are integers and follow a normal distribution, in
this study Gaussian Naive Bayes type is used (Eq. 2).
(2)
Random Forest (RF): Random Forest is an ensemble
learning technique that is used for classification and regression (Dhanya R,
2019; Hossin et al., 2023; and Kadhim et al., 2023). The term
‘Random Forest’ refers to the group of decision trees that are created from
subsets of training data randomly instead of creating a single tree during the
preprocessing step. The created group helps to tackle noises in data, which in
turn reduces the effect of overfitting, improves the performance and the
generalization of the models, and obtains better accuracy results. Therefore,
RF is considered one of the best solutions for many ML applications.
Support Vector Machine (SVM): Support Vector Machine is a
supervised ML learning method utilized for classification and regression
problems (SVM) (Ak, 2020; Hossin et al., 2023). It is also known as a
powerful method to detect outliers and noises in data. It works by finding an
n-dimensional separation hyperplane that helps to classify data inputs into a
similar and non-similar class, as shown in Figure 3. The maximum the margin in the
SVM classifier between classes, the better the hyperplane to compare more than
two features for classification and then produce accurate findings.
Furthermore, the closer the support vectors are to the hyperplane, the more the
ability of the SVM classifier to reduce overfitting, ensuring the
generalization of the model to new data properly.
(3)
where 
is the loss function and 
is the regularization.
Figure
3: The Illustration of SVM Algorithm (Ak,
2020)
K-Nearest Neighbour (KNN): K-Nearest Neighbors (KNN) is
also a supervised learning technique used for classification and regression
tasks (Ak, 2020; Hossin et al., 2023). The term ‘nearest neigbors’ means
the numerical value of ‘k’, which represents the nearest data points determined
in a dataset for prediction using majority voting for classification or
averaging for regression. The value of the ‘k’ also determines the degree of
the model performance, and its low value causes overfitting due to the noise
capturing in the data, whereas the high value leads to the model generalization
and produces an accurate prediction. KNN predictions are based on distance
metrics such as Euclidean distance, which calculates the distance between each
of the two data points in the datasets to reduce noise and the risk of
overfitting. KNN is a direct and flexible method that requires accurate tuning
to obtain the best model performance. Figure 4 shows an example of the KNN
technique (Ak, 2020).
Figure 4: An Example of KNN Classifier (Ak,
2020).
Neural Networks (NN): Neural networks is another
supervised ML technique used for classification problems (Mahesh, 2020). The
simple structure of NN consists of an input layer, one hidden layer, and an output
layer, as shown in Figure 5 (Yadav et al., 2022). If more than one
hidden layer exists in the NN algorithm, then it will be defined as a deep
learning algorithm. The NN layers are connected to each other, which consist of
artificial neurons that work together to find a solution to a problem similar
to that in the human brain and are used to detect patterns in data. The neuron
of the NN layers works by processing and analyzing the data and then passing
its output to the hidden layer, which further processes the incoming output and
passes its output to the output layer. The final layer might have more than one
output neuron, depending on the problem solved. A function named a loss
function is used to evaluate the performance of the NN by calculating errors in
the estimation process. The lower the value of the loss function, the more
effective the NN prediction is. To avoid noises and overfitting in data,
several techniques can be used, such as dropout, early stopping, and
regularization.
Figure
5: The Structure of an Artificial Neural Network (Yadav Et Al., 2022).
Feature Selection Methods
F-test: In
statistics, the F-test is a statistical feature selection technique that
computes the difference amount between two or more subsets of data (Dhal et al., 2022). The F-test is known to
be an effective method to deal with data with high dimensionality, such as
medical datasets. Therefore, this method is useful to be used for selecting
features such as tumour characteristics to determine whether the patient has
cancer or not in a dataset like breast cancer. Moreover, selecting the only
relevant attributes by F-test helps to reduce the classes’ overfitting and
increase the models’ performance. The F-test formula is (s12/s22),
where s12 is the variance of the first sample set, and s22
is the variance of the second sample set.
Mutual Information (MI): Mutual Information (MI) is
another statistical feature selection method that is used to find linear and
non-linear associations in sophisticated datasets such as medical datasets (Dhal et al., 2022). Therefore, it is a useful
approach in many fields, such as ML for modeling and healthcare for diagnosing
and treatment. In MI, one feature provides valuable information about another
feature, i.e., it measures how dependent each of the two variables is, whereas
zero MI means no dependency is available. More than zero means a dependency
between the two attributes (Vergara et al., 2015). MI provides the most related features to
the target of the dataset used which helps to increase the ML model
performance. In medical datasets like breast cancer, this process is important in
clarifying the relation between the dataset variables, which are mostly
ambiguous.
Spearman Correlation Coefficients: In Spearman Correlation
Coefficients, the robustness of two-variable correlations is specified such
that it can have positive or negative and weak or
strong values (Dhal et al.,
2022). Statistically, it is a non-distribution rank
measure, i.e., it measures the correlation between variables without
considering the distribution of the data (Hauke et al., 2011). This
feature makes the Spearman
correlation coefficients valuable, specifically in medical datasets where the
frequency of the data distribution is not considered. In other words, it can
simply find the relation between the input variables and the target variable.
Based on that, the Spearman correlation coefficients method is a helpful
measure for researchers working in medical fields to decide treatment and
diagnosis for patients.
Evaluation
Metrics
Models’ performance, strengths, and
weaknesses are essential to be evaluated. Hence, in this study, the selected
classification models using different evaluation metrics commonly used in the
literature are evaluated. The metrics used are:
·
Accuracy (Patro et
al., 2021) , mathematically shown as follows:
(4)
·
F1 score (Lichtenwalter
et al., 2010), mathematically shown as follows:
(5)
where:
(6)
(7)
·
Matthew’s correlation coefficient (
) (Ali
et al., 2021),
(8)
·
The area under the curve (AUC) of
receiver operating characteristics (ROC) is another statistic that is used to
evaluate models (Shiny Irene et al., 2020).
It assesses the performance of classification models using
threshold values ranging from 0 to 1, signifying poor to excellent predictions.
Accuracy is the proportion of
participants accurately predicted by the classification model relative to the
total number of tested subjects. Precision and recall (Eqs. 6 and 7) are combined into
the F1 score (Eq. 5), which is frequently used in binary classification
problems. Precision is defined as the ratio of true positive forecasts to all
positive predictions, whereas recall is the ratio of true positive predictions
to all predictions of positive data that are actually seen. The F1 score
ranges within the interval [0, 1], with a score of 1 indicating perfect precision
and recall.
MCC is an important statistical
measure for assessing the accuracy of binary categorization. It only gives an
outstanding grade if the prediction performs well in all four elements of the
confusion matrix: true positive (TP), false positive (FP), true negative (TN),
and false negative (FN). As a result, it is frequently viewed as a balanced
metric that may be used even when the classes are of vastly different sizes.
Its value ranges from -1 to 1, with -1 indicating that all tested participants
were wrongly or correctly predicted and 0 indicating that the model prediction
is no better than a random guess.
4.
EXPERIMENTAL RESULTS
In this part of the study, two
different types of experiments are conducted on the created dataset from the WDBC
dataset as described in section 3.1. In the first experiment, the selected ML
algorithms are implemented on the dataset without the incorporation of the
feature selection methods F-test,
MI, and Spearman Correlation and without resampling the data. In the
second experiment, the ML models are implemented using the three feature
selection methods and the resampling method SOMTE. Moreover, the performance of
the models is analyzed and compared in Section 5 using the feature selection
methods as illustrated in Tables 7-13. Different assessment metrics are used
for the models’ performance evaluation: accuracy, F1 score, precision, recall,
ROC AUC, and MCC.
In this work, all the experiments are
accomplished using a computer system having the following features: Intel(R)
Core (TM) i3-2310M CPU @ 2.10GHz, 4 GB of RAM, and 64-bit Windows 10 Pro OS.
Also, the Python programming language is used to conduct the experiments with
its libraries or packages such as NumPy, Pandas, and Scikit-Learn. For each
experiment, the dataset is divided into ten parts of the same size in order to
train and test the models efficiently and to make the models more reliable for prediction.
The process of dividing the dataset is called cross-validation, which splits
the data into 10 equal parts. Moreover, the default hyperparameters are used
for each model, such as setting the value of k to 5 in KNN; one hidden layer
and a maximum of 5000 iterations are set in NN; and lastly, in LR, 1000
iterations are established.
Results without feature selection
In this section, the results of the
implementation of the selected ML models on the reduced WDBC dataset are presented.
This is done without using the resampling technique and the selected feature
selection methods. The findings of the classification are illustrated in Table 3
and Figure 6 for the selected evaluation metrics used in this study.
As shown in Table 3, the
performance of the NN and LR models shows the highest accuracy values of 0.978910 and 0.977153 compared to those of the
NB model, which is poor. Also, the two models provide a high balance between
true positive and false positive scoring, with high F1 values of 0.978869 and 0.977096, respectively. In terms of precision and
recall, NN, LR, and SVM achieve the highest discrimination of the class labels.
Moreover, the highest ROC AUC value is also for NN, with a score of 0.975530
compared to the DT’s, which is the lowest value of 0.907424. The obtained
results reveal that NN and LR are the best models in performing classification
on the WDBC dataset compared to the other models. Lastly, the MCC values show
that NN and LR also outperform the other models, specifically NB and DT, in
providing a balanced score, achieving 0.954827 and 0.951067, respectively, while NB and DT have the lowest values of
MCC.
Table
3: Evaluation Metrics Comparison for Wisconsin Diagnostic Breast Cancer Without
Feature Selection
|
Model
|
Accuracy
|
F1 Score
|
Precision
|
Recall
|
ROC_AUC
|
MCC
|
|
DT
|
0.913884
|
0.913842
|
0.913807
|
0.913884
|
0.907424
|
0.815636
|
|
KNN
|
0.964851
|
0.964707
|
0.964976
|
0.964851
|
0.958578
|
0.924663
|
|
LR
|
0.977153
|
0.977096
|
0.977202
|
0.977153
|
0.973171
|
0.951067
|
|
NB
|
0.933216
|
0.933015
|
0.933036
|
0.933216
|
0.925704
|
0.856551
|
|
NN
|
0.978910
|
0.978869
|
0.978931
|
0.978910
|
0.975530
|
0.954827
|
|
RF
|
0.968366
|
0.968303
|
0.968343
|
0.968366
|
0.964253
|
0.932183
|
|
SVM
|
0.973638
|
0.973599
|
0.973618
|
0.973638
|
0.970370
|
0.943512
|
Figure
6: Model Performance across Different Metrics without Feature Selection
Methods.
Results
with Feature Selection
F-test: This
section presents the results of combining F-test feature selection methodology
with machine learning techniques. The dataset is resampled and scaled, and the
features are selected. The classification results show accuracy, precision,
recall, MCC, and ROC-AUC assessment metrics, as presented in Table 4 and Figure
7.
Table 4 shows a detailed analysis
of how well different models classify data based on the metrics produced from
the F-test feature selection procedure. Among the models tested, LR and NN have
the maximum accuracy, scoring 0.985994 and 0.981793, with the F-test technique
selecting only 20 and 23 features, respectively. The obtained high performance
of these two models reveals that both models are reliable for accurate
prediction of the classes and also that the number of selected attributes is enough
to obtain such good performance, which means it is not necessary to use all the
features of the dataset. On the other hand, the F test method selected 18, 20,
and 20 features for the KNN, RF, and SVM models of 0.976190, 0.976190, and
0.977591, respectively. This also indicates that the selected features are good
enough to obtain such good performance; however, these performances are still
lower in accuracy compared to those in NN and LR. In terms of precision and
recall, both LR and NN outperform the other models in detecting the true
positives and reducing the false negatives, with LR’s recall of 0.988796 and
precision of 0.983287, providing a balance to distinguish between positive and
negative records.
The NB model, on the other hand,
underperforms the other models by obtaining the lowest accuracy of 0.946779 and
F1 score of 0.947514. This indicates that NB is not suitable enough for
predicting cancer in the WDBC dataset.
Moreover, LR and NN provide a
strong F1 score, which balances between false positives and false negatives,
with LR receiving the best score of 0.986034 and NN, with 0.981818.
Furthermore, LR, NN, and three other models KNN, RF, and SVM produce a good
score in terms of ROC AUC values. The metrics values are over 90%, indicating
that the models are capable of distinguishing between classes. The final metric,
MCC, is considered as a reliable metric for the prediction process and for
considering all categories of the confusion matrix since it provides the highest
values for LR with 0.972004 and NN with 0.963589.
Table
4: Evaluation Metrics Comparison for Wisconsin Diagnostic Breast Cancer
with F-test Feature Selection Method.
|
Model
|
No.
of Selected Features
|
Accuracy
|
F1
Score
|
Precision
|
Recall
|
ROC_AUC
|
MCC
|
|
DT
|
16
|
0.959384
|
0.958982
|
0.968571
|
0.949580
|
0.959384
|
0.918944
|
|
KNN
|
18
|
0.976190
|
0.976224
|
0.974860
|
0.977591
|
0.976190
|
0.952385
|
|
LR
|
20
|
0.985994
|
0.986034
|
0.983287
|
0.988796
|
0.985994
|
0.972004
|
|
NB
|
3
|
0.946779
|
0.947514
|
0.934605
|
0.960784
|
0.946779
|
0.893908
|
|
NN
|
23
|
0.981793
|
0.981818
|
0.980447
|
0.983193
|
0.981793
|
0.963589
|
|
RF
|
20
|
0.976190
|
0.976157
|
0.977528
|
0.974790
|
0.976190
|
0.952385
|
|
SVM
|
20
|
0.977591
|
0.977591
|
0.977591
|
0.977591
|
0.977591
|
0.955182
|
Figure
7: Model Performance Across Different Metrics with F-test Feature
Selection Method
Mutual Information (MI): The second feature selection method used in this study is MI, which
is combined in this section with the ML techniques to show its impact on the
performance of the models. The prediction process includes resampling the
dataset, scaling, and then using the MI for selecting features. The
classification results are then evaluated using the evaluation metrics selected
for this study, as shown in Table 5 and Figure 8.
In the context of classification tasks,
each metric in Table 5 provides a different perspective on the model's ability
to make accurate predictions. The best-performing model is LR, which achieves
the highest scores across practically all metrics using all the dataset's
attributes, as selected using the MI approach. With an accuracy of 0.983193 and
an F1 score of 0.983287, LR demonstrates outstanding ability to balance
precision with recall, which is critical for effective classification. LR’s
precision of 0.977839 and recall of 0.988796 shows that it not only generates
accurate positive predictions but also correctly recognizes a large number of
true positive events. The ROC_AUC score of 0.983193 highlights its ability to
discriminate across classes, while the MCC of 0.966447 indicates a significant
connection between predicted and actual results. KNN is a reliable model, with
an accuracy of 0.964986; however, it falls short of LR in terms of overall
effectiveness. The NN model achieves an impressive accuracy of 0.978992,
demonstrating its ability to sustain high precision (0.975000) and recall
(0.983193). The SVM and RF models perform well, with SVM reaching an accuracy
of 0.976190 and RF scoring 0.970588. Despite their competitiveness, these
models do not outperform the other ones in any specific category, implying that
they may be more appropriate for certain contexts rather than being the top
choices overall. Furthermore, the investigated models indicate that the DT
model has the lowest performance, with an accuracy of 0.959384. On the other
hand, the NB model dramatically underperforms, with the lowest scores across
most metrics and an accuracy of 0.929972. This suggests that NB may not be
appropriate for this dataset.
Table
5: Evaluation Metrics Comparison for Wisconsin Diagnostic Breast Cancer with MI
Feature Selection Method
Model |
No of Selected Features |
Accuracy |
F1 Score |
Precision |
Recall |
ROC_AUC |
MCC |
DT |
23 |
0.959384 |
0.959327 |
0.960674 |
0.957983 |
0.959384 |
0.918771 |
KNN |
23 |
0.964986 |
0.965132 |
0.961111 |
0.969188 |
0.964986 |
0.930005 |
LR |
23 |
0.983193 |
0.983287 |
0.977839 |
0.988796 |
0.983193 |
0.966447 |
NB |
23 |
0.929972 |
0.931694 |
0.909333 |
0.955182 |
0.929972 |
0.861039 |
NN |
23 |
0.978992 |
0.979079 |
0.975000 |
0.983193 |
0.978992 |
0.958017 |
RF |
23 |
0.970588 |
0.970547 |
0.971910 |
0.969188 |
0.970588 |
0.941180 |
SVM |
23 |
0.976190 |
0.976290 |
0.972222 |
0.980392 |
0.976190 |
0.952415 |
Figure 8: Model Performance Across
Different Metrics with MI Feature Selection Method
Spearman Correlation Coefficient: This section presents the results of combining Spearman Correlation
Coefficient feature selection methodology with machine learning techniques. The
dataset is resampled, scaled, and the features selected. The classification
results show accuracy, precision, recall, MCC, and ROC-AUC assessment metrics,
as shown in Table 6 and Figure 9. In Table 6, the results of the assessment
metrics used to evaluate the models’ performance in this study are presented.
This is to show the impact of combining the Spearman method with the ML models.
In the table, the LR model outperforms the other models, obtaining the highest
accuracy and F1 score using 23 features. The accuracy is 0.984594, and the F1
score is 0.984658. In terms of other metrics such as precision and recall, as
well as the MCC, LR show competitive values with a precision of 0.980556, a recall
of 0.988796, and an MCC of 0.969222. These findings indicate that the
capability of LR to correctly classify the positive cases and almost all of
them are correct. They also provide a strong association between the predicted
and real outcomes. Also, LR demonstrates the high capability of differentiation
between the target classes by providing a ROC AUC value of 0.984594.
On the other hand, the table shows the
results of both NN and KNN, which are considered to be close to each other in
terms of accuracy and F1 score. NN has the highest values of accuracy of
0.981793 and F1 score of 0.981818 compared to KNN, which has the accuracy value
of 0.970588 and F1 score of 0.970711. This indicates that KNN provides a good
balance between precision and recall according to the value of the F1 score as
well as the capability of NN to provide a stronger balance between precision of
0.980447 and recall of 0.983193. Despite the good results of both NN and KNN,
they still fall below the performance of LR. Regarding the RF and SVM models,
their assessment metrics results show that they are accurately measured in
terms of accuracy and F1 score. Furthermore, the results of the DT model
highlight that its performance is the lowest compared to the other models, with
an accuracy of 0.950980 and an F1 score of 0.950495, indicating that DT is not
suitable for undertaking the process of modelling incorporating the Spearman
feature selection method. However, the performance of the NB model is the worst
compared to DT and other models, achieving an accuracy of 0.938375 and lower
precision and recall values, leading to the least favorable overall metrics.
Table
6: The Comparison of Evaluation Metrics for WDBC dataset using Spearman Feature
Selection Method
Model |
No of Features |
Accuracy |
F1 Score |
Precision |
Recall |
ROC_AUC |
MCC |
DT |
21 |
0.950980 |
0.950495 |
0.960000 |
0.941176 |
0.950980 |
0.902134 |
KNN |
23 |
0.970588 |
0.970711 |
0.966667 |
0.974790 |
0.970588 |
0.941210 |
LR |
23 |
0.984594 |
0.984658 |
0.980556 |
0.988796 |
0.984594 |
0.969222 |
NB |
23 |
0.938375 |
0.939560 |
0.921833 |
0.957983 |
0.938375 |
0.877426 |
NN |
21 |
0.981793 |
0.981818 |
0.980447 |
0.983193 |
0.981793 |
0.963589 |
RF |
20 |
0.978992 |
0.979021 |
0.977654 |
0.980392 |
0.978992 |
0.957987 |
SVM |
23 |
0.978992 |
0.979021 |
0.977654 |
0.980392 |
0.978992 |
0.957987 |
Figure
9: Model Performance Across Different Metrics with Spearman Feature
Selection Method
5.
COMPARISON OF THE ML MODELS USING THE
FEATURE SELECTION METHODS
In this section, Tables 7-13
illustrate the results of comparing the performance of the ML models - DT, KNN,
LR, NB, NN, RF, and SVM that are analyzed and selected for this study. The
analysis and comparison are done using the feature selection techniques:
F-test, MI, and Spearman, and are evaluated using the performance metrics:
accuracy, F1 score, precision, recall, ROC_AUC, and Matthews Correlation
Coefficient (MCC).
FS method/ DT |
No of Features |
Accuracy |
F1 Score |
Precision |
Recall |
ROC_AUC |
MCC |
F-test |
16 |
0.959384 |
0.958982 |
0.968571 |
0.949580 |
0.959384 |
0.918944 |
MI |
23 |
0.959384 |
0.959327 |
0.960674 |
0.957983 |
0.959384 |
0.918771 |
Spearman |
21 |
0.950980 |
0.950495 |
0.960000 |
0.941176 |
0.950980 |
0.902134 |
FS method/ KNN |
No of Features |
Accuracy |
F1 Score |
Precision |
Recall |
ROC_AUC |
MCC |
F-test |
18
|
0.976190
|
0.976224
|
0.974860
|
0.977591
|
0.976190
|
0.952385
|
MI |
23 |
0.964986 |
0.965132 |
0.961111 |
0.969188 |
0.964986 |
0.930005 |
Spearman |
23 |
0.970588 |
0.970711 |
0.966667 |
0.974790 |
0.970588 |
0.941210 |
Table
7: The Performance of Decision Tree Classifier Using F-test, MI, and Spearman
Feature Selection Methods
Table
8: The Performance of KNN Classifier Using F-test, MI, and Spearman Feature
Selection Methods
Table
9: The LR Classifier Performance Using F-test, MI, and Spearman Feature
Selection Methods
FS method/ LR |
No of Features |
Accuracy |
F1 Score |
Precision |
Recall |
ROC_AUC |
MCC |
F-test |
20
|
0.985994
|
0.986034
|
0.983287
|
0.988796
|
0.985994
|
0.972004
|
MI |
23 |
0.983193 |
0.983287 |
0.977839 |
0.988796 |
0.983193 |
0.966447 |
Spearman |
23 |
0.984594 |
0.984658 |
0.980556 |
0.988796 |
0.984594 |
0.969222 |
Table
10: The NB Classifier Performance Using F-test, MI, and Spearman Feature
Selection Methods
FS method/ NB |
No of Features |
Accuracy |
F1 Score |
Precision |
Recall |
ROC_AUC |
MCC |
F-test |
3
|
0.946779
|
0.947514
|
0.934605
|
0.960784
|
0.946779
|
0.893908
|
MI |
23 |
0.929972 |
0.931694 |
0.909333 |
0.955182 |
0.929972 |
0.861039 |
Spearman |
23 |
0.938375 |
0.939560 |
0.921833 |
0.957983 |
0.938375 |
0.877426 |
Table
11: The Performance of NN Classifier Using F-test, MI, and Spearman Feature
Selection Methods
FS method/ NN |
No of Features |
Accuracy |
F1 Score |
Precision |
Recall |
ROC_AUC |
MCC |
F-test |
23
|
0.981793
|
0.981818
|
0.980447
|
0.983193
|
0.981793
|
0.963589
|
MI |
23 |
0.978992 |
0.979079 |
0.975000 |
0.983193 |
0.978992 |
0.958017 |
Spearman |
21 |
0.981793 |
0.981818 |
0.980447 |
0.983193 |
0.981793 |
0.963589 |
Table
12: The Performance of RF Classifier Using F-test, MI, and Spearman Feature
Selection Methods
FS method/ RF |
No of Features |
Accuracy |
F1 Score |
Precision |
Recall |
ROC_AUC |
MCC |
F-test |
20
|
0.976190
|
0.976157
|
0.977528
|
0.974790
|
0.976190
|
0.952385
|
MI |
23 |
0.970588 |
0.970547 |
0.971910 |
0.969188 |
0.970588 |
0.941180 |
Spearman |
20 |
0.978992 |
0.979021 |
0.977654 |
0.980392 |
0.978992 |
0.957987 |
Table
13: The SVM Classifier Performance Using F-test, MI, and Spearman Feature Selection
Methods
FS method/ SVM |
No of Features |
Accuracy |
F1 Score |
Precision |
Recall |
ROC_AUC |
MCC |
F-test |
20
|
0.977591
|
0.977591
|
0.977591
|
0.977591
|
0.977591
|
0.955182
|
MI |
23 |
0.976190 |
0.976290 |
0.972222 |
0.980392 |
0.976190 |
0.952415 |
Spearman |
23 |
0.978992 |
0.979021 |
0.977654 |
0.980392 |
0.978992 |
0.957987 |
By utilizing the F-test, the method
in Table 9 achieves an impressive accuracy of 0.985994 and an F1 score of
0.986034, proving its robustness. In addition, it has comparable metrics to MI
and Spearman. LR exhibits a balanced performance in classification tasks, as
evidenced by its consistently high precision and recall. Likewise, the NN
exhibits impressive results, especially when utilizing F-test (achieving an
accuracy of 0.981793) as depicted in Table 11. It consistently achieves high
precision and recall scores in both MI (accuracy of 0.978992) and Spearman
(accuracy of 0.981793), validating its capability to handle complex patterns.
Table 8 demonstrates that KNN
produces impressive results, particularly when compared to F-test (with an
accuracy of 0.976190). However, it exhibits more variability when used with MI
and Spearman (yielding accuracies of 0.964986 and 0.970588, respectively). While
KNN demonstrates good performance, it does not outperform the top two models
(LR and NN). RF algorithm consistently demonstrates strong performance,
especially when paired with F-test (achieving an accuracy of 0.976190). Despite
showing strong metrics, it still falls behind LR and NN, but performs similarly
to KNN with F-test. In Table 13, SVM demonstrates good performance in both
F-test and Spearman, with an accuracy of approximately 0.977591 and 0.978992,
respectively. However, the accuracy achieved by MI is slightly lower – 0.976190.
Additionally, SVM's robust recall scores validate its effectiveness in
accurately detecting positive instances. On the other hand, when comparing the
models, the DT is found to be the least effective performer, as indicated in
Table 7. It achieves an accuracy of 0.959384 using F-test. Despite achieving
high accuracy scores of 0.959384 in MI and 0.950980 in Spearman, the
performance of the model does not match that of the other models. NB performs
weakest among all models, consistently scoring below 0.943 in accuracy across
all feature selection methods. The performance metrics as presented in Table 10
reveal a significant drop in precision and F1 scores, indicating that NB is
less suited for this classification task compared to other models.
Based on the feature selection
techniques used, the overall analysis of Tables 7–13 identifies NN and LR as
the best models. SVM, RF, and KNN are reliable alternatives; however, they fall
short of LR and NN in terms of performance. On the other hand, it is evident
that DT and NB have limits when it comes to accurately capturing the underlying
data complexity, which further confirms their lower suitability for this
particular problem domain.
CONCLUSION
AND FUTURE WORKS
In this study, a comparison was conducted
to show the effect of various feature selection methods, F-test, MI, and
Spearman correlation coefficients, to the performance of seven different
machine learning techniques: KNN, NB, DT, SVM, LR, NN, and RF. The study
evaluated the results using different evaluation metrics such as accuracy, F1
score, precision, recall, ROC AUC, and MCC with the breast cancer dataset WDBC.
It is concluded that the high performance of the models can be obtained by not
using all the features of the dataset for prediction and improving modelling.
Thus, feature selection methods were employed to select the features that
really influence the model’s performance and can predict whether a patient has
cancer or not. Tables 7-13 presented the number of the selected features using
the feature selection methods that yielded the best performance according to
the evaluation metrics used.
The results indicate that LR and NN are the
best models. SVM, RF, and KNN are reliable alternatives; however, they do not
match the performance of LR and NN. On the other hand, it appears that DT and
NB are not as effective in this particular problem domain as they could be when
it comes to accurately capturing data complexity.
Future research needs to explore more
feature selection methods, deep learning models, hyperparameter optimization,
and diverse data types to improve breast cancer prediction and enhance machine
learning model effectiveness in healthcare. Furthermore, the future work also
needs to propose new hybrid algorithms from the ML models based on the results
obtained in this paper. This can be accomplished by integrating different ML
models to enhance the prediction of breast cancer in the healthcare sector and
then use different feature selection methods to select the optimal features
from the breast cancer datasets.
REFERENCES
Abunasser, B. S., AL-Hiealy, M. R. J., Zaqout, I. S.,
& Abu-Naser, S. S. (2023). Convolution Neural Network for Breast Cancer
Detection and Classification Using Deep Learning. Asian Pacific Journal of
Cancer Prevention, 24(2), 531–544. DOI: 10.31557/APJCP.2023.24.2.531
Ak, M. F. (2020). A comparative analysis of breast
cancer detection and diagnosis using data visualization and machine learning
applications. Healthcare (Switzerland),8(2).DOI:10.3390/healthcare8020111
Ali, M. M., Paul, B. K., Ahmed, K., Bui, F. M., Quinn,
J. M. W., & Moni, M. A. (2021). Heart disease prediction using supervised
machine learning algorithms: Performance analysis and comparison. Computers
in Biology and Medicine, 136, 104672. DOI: 10.1016/j.compbiomed.2021.104672
Ara, S., Das, A., & Dey, A. (2021). Malignant and
Benign Breast Cancer Classification using Machine Learning Algorithms. 2021
International Conference on Artificial Intelligence, ICAI 2021, 97–101. DOI: 10.1109/ICAI52203.2021.9445249
Botlagunta, M., Botlagunta, M. D., Myneni, M. B.,
Lakshmi, D., Nayyar, A., Gullapalli, J. S., & Shah, M. A. (2023).
Classification and diagnostic prediction of breast cancer metastasis on
clinical data using machine learning algorithms. Scientific Reports, 13(1).
DOI: 10.1038/s41598-023-27548-w
Bray, F., Laversanne, M., Sung, H., Ferlay, J., Siegel,
R. L., Soerjomataram, I., & Jemal, A. (2024). Global cancer statistics
2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in
185 countries. CA: A Cancer Journal for Clinicians,74(3),229263.DOI:10.3322/caac.21834
Chaurasia, V., & Pal, S. (2020). Applications of
Machine Learning Techniques to Predict Diagnostic Breast Cancer. SN Computer
Science, 1(5). DOI: 10.1007/s42979-020-00296-8
Chawla, N. V, Bowyer, K. W., Hall, L. O., &
Kegelmeyer, W. P. (2002). SMOTE: Synthetic Minority Over-sampling Technique. In
Journal of Artificial Intelligence Research (Vol. 16).
Chen, H., Wang, N., Du, X., Mei, K., Zhou, Y., &
Cai, G. (2023). Classification Prediction of Breast Cancer Based on Machine
Learning. Computational Intelligence and Neuroscience, 2023, 1–9.
DOI: 10.1155/2023/6530719
Dhal, P., & Azad, C. (2022). A comprehensive survey
on feature selection in the various fields of machine learning. Applied
Intelligence, 52(4), 4543–4581. DOI:
10.1007/s10489-021-02550-9
Dhanya R, I. R. P. S. S. A. M. S. and J. J. N. (2019). A
Comparative Study for Breast Cancer Prediction using Machine Learning and
Feature Selection.
Ebrahim, M., Sedky, A. A. H., & Mesbah, S. (2023).
Accuracy Assessment of Machine Learning Algorithms Used to Predict Breast
Cancer. Data, 8(2). DOI: https://doi.org/10.3390/data8020035
Hauke, J., & Kossowski, T. (2011). Comparison of
values of pearson’s and spearman’s correlation coefficients on the same sets of
data. Quaestiones Geographicae,30(2),8793.DOI:10.2478/v10117-011-0021-1
Hossin, M. M., Javed Mehedi Shamrat, F. M., Bhuiyan, M.
R., Hira, R. A., Khan, T., & Molla, S. (2023). Breast cancer detection: an
effective comparison of different machine learning algorithms on the Wisconsin
dataset. Bulletin of Electrical Engineering and Informatics, 12(4),
2446–2456. DOI: 10.11591/eei.v12i4.4448
Islam, M. M., Haque, M. R., Iqbal, H., Hasan, M. M.,
Hasan, M., & Kabir, M. N. (2020). Breast Cancer Prediction: A Comparative
Study Using Machine Learning Techniques. SN Computer Science, 1(5).
DOI: 10.1007/s42979-020-00305-w
Kadhim, R. R., & Kamil, M. Y. (2023). Comparison of
machine learning models for breast cancer diagnosis. IAES International
Journal of Artificial Intelligence,12(1),415421.DOI:10.11591/ijai.v12.i1.pp415-421
Kumar, S., & Singh, M. (2021). Breast Cancer
Detection Based on Feature Selection Using Enhanced Grey Wolf Optimizer and
Support Vector Machine Algorithms. Vietnam Journal of Computer Science,8(2),177197.DOI:10.1142/S219688882150007X
Laghmati, S., Hamida, S., Hicham, K., Cherradi, B.,
& Tmiri, A. (2024). An improved breast cancer disease prediction system
using ML and PCA. Multimedia Tools and Applications, 83(11),
33785–33821. DOI: 10.1007/s11042-023-16874-w
Lappeenranta-. (2023). BREAST CANCER DIAGNOSTIC USING
MACHINE LEARNING Applying Supervised Learning Techniques to Coimbra and
Wisconsin Datasets.
Lichtenwalter, R. N., Lussier, J. T., & Chawla, N.
V. (2010). New perspectives and methods in link prediction. Proceedings of
the 16th ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining, 243–252. DOI: 10.1145/1835804.1835837
Mahesh, B. (2020). Machine Learning Algorithms - A
Review. International Journal of Science and Research (IJSR), 9(1),
381–386. DOI: 10.21275/art20203995
Mohammed, S. A., Darrab, S., Noaman, S. A., & Saake,
G. (2020). Analysis of breast cancer detection using different machine learning
techniques. Communications in Computer and Information Science, 1234
CCIS, 108–117. DOI: 10.1007/978-981-15-7205-0_10
Naji, M. A., Filali, S. El, Aarika, K., Benlahmar, E.
H., Abdelouhahid, R. A., & Debauche, O. (2021). Machine Learning Algorithms
for Breast Cancer Prediction and Diagnosis. Procedia Computer Science,191,487492.DOI:10.1016/j.procs.2021.07.062
Nemade, V., & Fegade, V. (2022). Machine Learning
Techniques for Breast Cancer Prediction. Procedia Computer Science, 218,
1314–1320. DOI: 10.1016/j.procs.2023.01.110
Patro, S. P., Nayak, G. S., & Padhy, N. (2021).
Heart disease prediction by using novel optimization algorithm: A supervised
learning prospective. Informatics in Medicine Unlocked, 26,
100696. DOI: 10.1016/j.imu.2021.100696
Sakib, S., Yasmin, N., Tanzeem, A. K., Shorna, F., Md.
Hasib, K., & Alam, S. B. (2022). Breast Cancer Detection and
Classification: A Comparative Analysis Using Machine Learning Algorithms. Lecture
Notes in Electrical Engineering, 844, 703–717. DOI: 10.1007/978-981-16-8862-1_46
Shiny Irene, D., Sethukarasi, T., & Vadivelan, N.
(2020). Heart disease prediction using hybrid fuzzy K-medoids attribute
weighting method with DBN-KELM based regression model. Medical Hypotheses,143(March),110072.DOI:10.1016/j.mehy.2020.110072
Uddin, K. M. M., Biswas, N., Rikta, S. T., & Dey, S.
K. (2023). Machine learning-based diagnosis of breast cancer utilizing feature
optimization technique. Computer Methods and Programs in Biomedicine Update,3.DOI: 10.1016/j.cmpbup.2023.100098
Vergara, J. R., & Estévez, P. A. (2015). A Review
of Feature Selection Methods Based on Mutual Information. DOI: 10.1007/s00521-013-1368-0
Wolberg, W. M. S. and S. (1995). Breast Cancer
Wisconsin (Diagnostic). UCI Machine Learning Repository. DOI: 10.24432/C5DW2B
Yadav, R. K., Singh, P., & Kashtriya, P. (2022).
Diagnosis of Breast Cancer using Machine Learning Techniques -A Survey. Procedia
Computer Science,218,14341443.DOI:10.1016/j.procs.2023.01.122
Zhou, S., Hu, C., Wei, S., & Yan, X. (2024). Breast
Cancer Prediction Based on Multiple Machine Learning Algorithms. Technology
in Cancer Research and Treatment, 23. DOI: 10.1177/15330338241234791