Tech Science Press

DOI: 10.32604/cmc.2024.052147

ARTICLE

Improving Multiple Sclerosis Disease Prediction Using Hybrid Deep Learning

Model

Stephen Ojo1 , Moez Krichen2,3, *, Meznah A. Alamro4 , Alaeddine Mihoub5 ,

Gabriel Avelino Sampedro6 and Jaroslava Kniezova7, *
1
Department of Electrical and Computer Engineering, College of Engineering, Anderson University, Anderson, SC 29621, USA
2
Faculty of Computing and Information, Al-Baha University, Al-Baha, 65528, Saudi Arabia
3
ReDCAD Laboratory, University of Sfax, Sfax, 3038, Tunisia
4
Department of Information Technology, College of Computer & Information Science, Princess Nourah Bint Abdul Rahman
University, Riyadh, 11564, Saudi Arabia
5
Department of Management Information Systems, College of Business and Economics, Qassim University, P.O. Box 6640,
Buraidah, 51452, Saudi Arabia
6
School of Management and Information Technology, De La Salle-College of Saint Benilde, Manila, 1004, Philippines
7
Department of Information Management and Business Systems, Faculty of Management, Comenius University Bratislava,
Bratislava 25, 82005, Slovakia
*Corresponding Authors: Moez Krichen. Email: [Link]@[Link]; Jaroslava Kniezova.
Email: [Link]@[Link]
Received: 25 March 2024 Accepted: 31 July 2024 Published: 15 October 2024

ABSTRACT
Myelin damage and a wide range of symptoms are caused by the immune system targeting the central nervous
system in Multiple Sclerosis (MS), a chronic autoimmune neurological condition. It disrupts signals between
the brain and body, causing symptoms including tiredness, muscle weakness, and difficulty with memory and
balance. Traditional methods for detecting MS are less precise and time-consuming, which is a major gap in
addressing this problem. This gap has motivated the investigation of new methods to improve MS detection
consistency and accuracy. This paper proposed a novel approach named FAD consisting of Deep Neural Network
(DNN) fused with an Artificial Neural Network (ANN) to detect MS with more efficiency and accuracy, utilizing
regularization and combat over-fitting. We use gene expression data for MS research in the GEO GSE17048
dataset. The dataset is preprocessed by performing encoding, standardization using min-max-scaler, and feature
selection using Recursive Feature Elimination with Cross-Validation (RFECV) to optimize and refine the dataset.
Meanwhile, for experimenting with the dataset, another deep-learning hybrid model is integrated with different ML
models, including Random Forest (RF), Gradient Boosting (GB), XGBoost (XGB), K-Nearest Neighbors (KNN)
and Decision Tree (DT). Results reveal that FAD performed exceptionally well on the dataset, which was evident
with an accuracy of 96.55% and an F1-score of 96.71%. The use of the proposed FAD approach helps in achieving
remarkable results with better accuracy than previous studies.

KEYWORDS
Multi Sclerosis (MS); machine learning; deep learning; artificial neural network; healthcare

Copyright © 2024 The Authors. Published by Tech Science Press.

This work is licensed under a Creative Commons Attribution 4.0 International License, which permits
unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
644 CMC, 2024, vol.81, no.1

1 Introduction
In recent years, there has been a significant surge in medical advancements, particularly in success-
fully treating MS [1]. These breakthroughs mark a pivotal moment in the medical landscape, offering
new and innovative ideas to improve the management of MS. With continuous progress in medical
research, our understanding of this complex neurological disorder is becoming more sophisticated,
paving the way for more effective and targeted therapeutic methods. MS is a debilitating neurological
condition that affects the central nervous system. It occurs when the immune system mistakenly
attacks the protective covering of nerve fibers, known as the myelin sheath [2]. Consequently, this
disrupts communication between the brain and body. The results of this disruption can include fatigue,
visual impairments, tingling or numbness, weakness in muscles, coordination difficulties, and sensory
abnormalities. Diagnosis of MS typically involves imaging tests and clinical evaluations. While there
is currently no known cure for MS, various treatments, such as medication and physical therapy, are
utilized to manage symptoms, slow disease progression, and enhance the overall quality of life for
those living with MS. There exists a dire need to increase the detection efficiency of MS for efficiently
dealing with its diagnosis. Previous methods exist, but with several limitations; their complexity is high
to achieve higher accuracy. This high complexity and low results are the obstacles in the diagnosis
of MS. Studies show that multi-sclerosis is increasing with time in different areas of the world;
there is a crucial need to understand its dynamics to address this rising chronic condition. Previous
approaches were conventional and not efficient for diagnosing MS, showing that these techniques are
efficient. This approach of multi-sclerosis detection, including other attempts at diagnosis, emerges as
an authentic weapon, and quick detection and differentiation of the disease come at the pinnacle of
success. In addition, the development of this approach led to a more precise diagnosis. In addition,
early management is done so that the patient’s situation can be improved, and the prognosis will be
enhanced.

1.1 Motivation
Millions of people worldwide have MS, a chronic and disabling autoimmune disease [3]. Research
in this area is motivated by various factors, including the need to understand the disease better,
develop effective treatment strategies, and improve the quality of life for affected individuals. The
subtle intricacies of MS, an enduring and chronic autoimmune disease, have caused many researchers
to start exploring it due to the ever-growing rate of this condition [4]. Recent statistics indicate an
increasing trend necessitating deeper inquiry into the complexities of the disease. It is not just about
numbers; MS can be devastating for patients globally. This encompasses mobility issues, cognitive
decline, and emotional turmoil’s [5]. Hence, researchers must look beyond causation and create a new
approach that will provide promising solutions that will significantly enhance the quality of life among
victims. Although we have come to discover much about MS, there is still a lot of work that needs to
be done in terms of improving the current therapies and understanding how the disease progresses
[6]. In this way, our investigation represents a notable progression in science by integrating a FAD
using a multi-model approach. ANN supports complicated pattern identification through this, while
the deep-learning hybrid model offers a strong foundation for combining numerous data sources.
This effective merger exploits the advantages of each model. Therefore, through concerted efforts, we
expect to get profound insights into the intricacies of multi-sclerosis and provide steppingstones for
targeted and successful interventions [7]. Our studies employ expert computational models to widen
our comprehension of how multi-sclerosis is born and develops. This method of treatment gives hope
CMC, 2024, vol.81, no.1 645

for better care and results in patients with this multifaceted neurological disorder, thereby initiating a
new era of precision medicine specifically designed for the unique complexities associated with MS.

1.2 Contributions
• Proposed a novel approach named FAD consisting of ANN-DNN for multi-sclerosis detection
that significantly improves MS detection by using Machine Learning (ML) and Deep Learning
(DL) multimodal techniques. This work substantially improves the previous method as this
approach increases efficiency by increasing accuracy and precision.
• The gene expression data utilized for our research on MS is available to the public through the
GEO GSE17048 dataset. We preprocessed the data by encoding, standardizing, and selecting
features to improve and optimize the dataset.
• This study provides. It performed better than the previous studies and the other deep-learning
hybrid model approach. The results provide an impressive accuracy of 96.55%.

1.3 Organization
The article is divided into five main sections. Section 1 briefly introduces MS and its effects on
the human mind, and it also contains an abbreviation (Table 1) for understanding technical terms
in the research being conducted. Section 2 delves into the research surrounding MS detection and
its advancements. The following section, Section 3, outlines the specific methods and techniques
used to achieve improved outcomes. Moving on, Section 4 presents the Experimental Analysis and
Results. Section 5 provides the discussion and comparison. Lastly, Section 6 wraps up the article with
conclusions and recommendations for future research.

Table 1: List of abbreviations

Acronym Abbreviations
SVM Support Vector Machines
FAD Fused ANN DNN
DLHM Deep Learning Hybrid Model
RFECV Recursive Feature Elimination and Cross-Validation
DT Decision Tree
GB Gradient Boosting
DT Decision Tree
RF Random Forest
XGB XGBoost
KNN K-Nearest Neighbors
MS Multi Sclerosis
MRI Magnetic Resonance Imaging
DL Deep Learning
PBMC Peripheral Blood Mononuclear Cells
OCT Optical Coherence Tomography
ANN Artificial Neural Network
AI Artificial Intelligence
LSTM Long Short-Term Memory
(Continued)
646 CMC, 2024, vol.81, no.1

Table 1 (continued)
Acronym Abbreviations
DNN Deep Neural Network
NB Naive Bayes
ReLU Rectified Linear Unit

2 Literature Review
Traditional methods have long been the norm for diagnosing MS. However, groundbreaking
research from 2011 to 2024 has uncovered the potential of using ML and DL to revolutionize the
diagnostic process. This approach includes advanced algorithms like Convolutional Neural Networks
and Support Vector Machines, which have shown promise in accurately analyzing Magnetic Resonance
Imaging (MRI) scans and clinical data, outperforming human abilities. The goal is to improve
diagnostic accuracy. Still, this technological advancement also carries the potential for creating
personalized treatment plans that could lead to earlier interventions and, ultimately, reduce medical
costs. Despite these exciting prospects, significant challenges still need to be addressed. To fully
capitalize on the potential of ML and DL for diagnosing MS, it is crucial to continually seek
out large and reliable datasets. Equally important is making these advanced models accessible and
understandable for medical professionals.
De Leon-Sanchez et al. [8] introduced an innovative DL approach that utilizes a single hidden
layer ANN to diagnose MS accurately. It will set our model apart in utilizing gene expression profiles,
specifically 74 genes, as input features. Unlike traditional ML methods, this strategy improves accuracy
and reduces complexity. To further enhance the model’s performance and maintain its simplicity, we
have integrated a regularization term into the single hidden layer of the ANN architecture to prevent
over-fitting. By conducting a comprehensive assessment, the study effectively evaluated the proposed
ANN model against four prominent ML algorithms-RF, Support Vector Machine (SVM), KNN,
and Naive Bayes (NB). The dataset, consisting of 144 individuals (45 healthy controls and 99 MS
patients), underwent feature selection using a powerful dimensionality reduction technique to pinpoint
key genes. The results were remarkable, showcasing the ANN model’s superiority in recall (80.80%),
accuracy (83.33%), precision (85.71%), and F1-score (0.83%). Aslam et al. [9] delved into the multitude
of possibilities that ML and DL can bring to the field of MS diagnosis while also acknowledging and
addressing the challenges that must be overcome for their responsible and successful integration. This
acknowledgment serves as a launching point for further research and progress, ensuring that these
powerful tools are utilized ethically and effectively to enhance the lives of those affected by MS.
Maleki et al. [10] applied Artificial Intelligence (AI) to the analysis of retinal images, which is a
way to improve MS management. This method is based on Optical Coherence Tomography (OCT)
scans, with AI algorithms that are trained through ML and DL techniques to reveal and classify
the abnormalities that are MS indicators. Also, these algorithms can predict the disease progression,
individualize the treatment plans and classify the MS subtypes according to the retinal features;
thus, potentially, it will be easier for the physicians. Although the research shows the possibilities of
improving MS diagnosis, monitoring and treatment planning by increasing the image analysis accuracy
and efficiency of OCT, the study also recognizes the difficulties, such as the ethical considerations and
the need for more research to be done before the use can be applied to the whole clinical practice.
CMC, 2024, vol.81, no.1 647

Shoeibi et al. [11] delved into the potential of ML to differentiate individuals with MS from
healthy individuals utilizing data from an instrumented walkway. The researchers hypothesized that
ML algorithms could more effectively classify MS patients using raw gait data instead of traditional
methods relying on predetermined features. To put this theory to the test, they recruited a group of 72
participants and collected data, such as pressure distribution and foot trajectory, from the walkway’s
sensors. The results revealed that the ML models accurately identified MS patients. Specifically,
Support Vector Machines (SVM), Logistic Regression, and XGB demonstrated remarkable accuracy,
with SVM boasting an 81% success rate. Significantly, including novel features derived directly from
the raw gait data boosted the accuracy rate to 88%. This approach highlights the impressive capabilities
of ML in accurately detecting MS through careful analysis of raw data. This research suggests the
potential for surpassing traditional approaches that rely on predetermined features.
Daqqaq et al. [12] showed how DL can be the source of the early detection of MS by training
Convolutional Neural Networks (CNNs) to perform MRI scans for indicative patterns. Although
the correctness of the results is guaranteed, the problems that must be solved are the big, labeled
datasets, the way of explaining the decisions of the model and the proper organization of the data
from different sources. The patient’s age, race and MRI variations should be taken into consideration
while ensuring model compatibility with varied patient demographics. Despite difficulties, methods
like transfer learning and data augmentation are proving to be the future of AI. In the end, the
MS diagnosis through DL can be used to detect the disease at its early stage, to design personalized
treatments, and to follow the disease progression; thus, the patient outcomes are improved.
Denissen et al. [13] introduced an innovative approach to enhancing the accurate prediction of
distinctive cognitive changes in individuals with MS. A multidimensional ML model was developed
through federated learning methods by merging information from diverse sources, including brain
MRI scans, clinical assessments, and genetic data. The team successfully integrated these different
forms of data to anticipate the cognitive trajectory of individual patients in the future. This model
outperformed previous ones that solely relied on singular data sources, resulting in a more com-
prehensive comprehension of the disease’s progression and generating more precise predictions. The
utilization of federated learning played a pivotal role in addressing privacy issues by facilitating
efficient model training with decentralized patient data. The study’s results demonstrate the potential
of multimodal ML in accurately predicting the rate of cognitive decline for each patient with MS. This
offers promising prospects for implementing early intervention and personalized treatment plans to
minimize cognitive impairment in MS patients.
Imaging and analyzing MS through currently used methods are not very accurate, and that
is why a novel method of detection is required. Conventional approaches are not precise enough;
rather, they use features that are already known in advance. At the same time, ML and DL face
issues with the availability of data used and their interpretability. The new technique delivers exclusive
strategies, which are based on gene expression profiles for discriminating, regularization using FAD,
and analyzing the raw gait data for classification purposes. A new approach is proposed that uses FAD
for multi-sclerosis detection with more precise, better, reliable, and personalized ways to diagnose and
prognosis MS.

3 Proposed Approach
This paper proposes a fused approach named FAD consisting of fused ANN [14] and DNN [9]
algorithms to tackle the detection of MS with more efficiency and accuracy, utilizing regularization
and a single hidden layer in an ANN to combat over-fitting. This innovative approach incorporates
648 CMC, 2024, vol.81, no.1

AI to enhance MS classification accuracy, showcasing its immense potential. Using a FAD helps
achieve remarkable results with better accuracy than previous studies. Fig. 1 visually represents the
proposed approach. Integrating regularization techniques within the fused ANN-DNN framework
effectively tackles over-fitting concerns, while its streamlined hidden layers design balances efficiency
and complexity. This advanced approach has revolutionized our understanding of tackling complex
neurological disorders such as MS and the potential impact of AI in medicine. The experimental
validation of this approach further highlights its potential to improve diagnostic accuracy and
personalizing treatments for individuals with MS.
finaloutput = σ(a2dropout .Wfinal + b_final) (1)

In Eq. (1), the sigmoid activation function is symbolized by σ and a2dropout shows the output of
the second hidden layer after applying dropout, and Wfinal + bfinal are the weight and bias terms of the
output layer. Although previous techniques had shown success, they eventually became less precise,
prompting the adoption of more advanced methods such as single-hidden-layer ANN. This cutting-
edge approach aims to enhance accuracy and precision, ultimately improving results when addressing
current challenges. The ultimate objective is to elevate the model’s performance and generate more
refined and precise outcomes than previous methods by harnessing the DL capabilities of an ANN
and DNN framework. To stay ahead of the curve in the ever-evolving world of AI and neural network
technology, our strategic shift reflects a firm dedication to incorporating the latest advancements in
pursuit of optimal results.

Figure 1: Workflow of proposed FAD approach for MS detection

3.1 Dataset
This study uses a microarray dataset and an external blood T-cell microarray dataset that are
combined in a dataset, which pertains to Peripheral Blood Mononuclear Cells (PBMC) [15]. Gene
expression data for MS research can be found in the GEO GSE17048 dataset, which is accessible to
the public [16]. The dataset includes information from 144 people, 45 of whom are healthy and 99
of whom have MS. They concentrate on 74 genes, including 73 genes found in earlier studies and the
HLA-DRB1 gene, which is linked to MS riskThe GEO2R network tool was utilized for analysis, while
the GPL6947 platform was utilized for data acquisition.
CMC, 2024, vol.81, no.1 649

3.2 Feature Extraction and Selection

One of the basic dimensionality reduction methods in ML and DL is feature selection. The
selection of parameters has the potential to greatly enhance the learning model’s performance by
improving the computational efficiency of the regression or classification algorithm. The research
methodology is developed using the RFECV method, which formally applies to encourage the
flexibility of the models being used [17]. The key to changing feature selection with this new method
is carefully preserving features while iteratively training our model to create our feature landscape.
A thorough assessment of the model’s efficacy is ensured by including cross-validation and RFECV,
effortlessly integrated into our research methodology, more conventional methods, and dynamically
evolving feature subsets.
The ML model employed to investigate the importance of features in this line of code is called
the RF classifier, or clf. The computed logic is arranged neatly in descending order in a dataframe.
Then, the less important component is removed. The curated list generated and saved in the variable
selected_feature contains the most important features determined by the model, as shown in Fig. 2.
The features are shown on x-axis and importance of each feature is shown on y-axis where highest the
value the better the feature. This analytical stage can guide intelligent product selection for analysis on
the next try, facilitate fine-grained result interpretation, and help understand the model’s underlying
decision logic.

Figure 2: Important features of dataset

The RFECV model is built with 5-fold cross-validation, a step size of 1, and parameters that define
accuracy as the scoring metric. The base estimator that has been chosen is RF classifier. RFECV uses
a RF partition to iteratively find the least significant features and identify the best sub features. The
‘X _selected’ data set is then created by the code extracting the index of the least important features and
removing it from the original dataset. The least important feature, representing the selected features, is
removed before printing these complexes. Algorithms like backpropagation, which change the weights
between neurons to lower the prediction error, are used in the training process. Because DL can learn
650 CMC, 2024, vol.81, no.1

in isolation and represent complex data objects, it has been extremely successful in applications like
natural language processing, graphic speech recognition, and healthcare diagnostics.

Algorithm 1: Pseudocode of the proposed approach

Require: X 2 (Features), Y 2 (Labels), Test Size = 0.25, Random State = 0
1: Split the dataset into training and testing sets: X train , X test , ytrain , ytest = train_test_split (X, Y ,
test_size = 0.25, random_state = 0)
2: Initialize empty lists: accuracies, f1_measure_list, precision_list, recall_list
3: Number of Iterations = 3
4: for i in range (Number of Iterations) do
5: Initialize a model: model = training_model ()
6: Train the model using the current labeled data: [Link] (X train , ytrain ):
7: Get the uncertainty scores
8: Get the index of the most uncertain_sample
9: Add the uncertain sample to labeled data
10: Remove the uncertain sample from the test
11: X test = X test .drop(index = uncertain_sample)
12: ytest .pop(uncertain_sample)
13: ypred = model. predict (X test )
14: Calculate evaluation metrics:
15: accuracy = accuracy_score (ytest , ypred )
16: f 1_measure = f1_score (ytest , ypred )
17: prec = precision_score (ytest , ypred )
18: recall = recall_score (ytest , ypred )
19: Append the metrics to the respective lists:
20: accuracies. append(accuracy)
21: f1_score_list. append(f 1_Score)
22: precision_list. append(prec)
23: recall_list. append(recall)
24: Print (Iterationi+1: Accuracy, F1-score, Precision, Recall)
25: end for
26: Get the iteration with the best accuracy, F1-score, precision, and Recall
27: Print the best accuracy, F1-score, Precision, Recall

3.3 ML with Hybrid Model

In this segment on ML for MS detection, we employ an array of algorithms, including RF, XGB,
GB, KNN, and DT with a hybrid model. This comprehensive approach enables us to thoroughly
medical data, thus enhancing the accuracy and effectiveness of MS detection. We aspire to advance
diagnostic techniques and improve patient outcomes by combining these diverse models. Its architec-
ture consists of three layers: a merged features layer, a layer for RF predictions, and an input layer for
original features. A Long Short-Term Memory (LSTM) layer comes first in the sequence, followed by
two dense layers and an embedding layer. Finally, a sigmoid activation function in the output layer
makes binary categorization easier, as shown in Fig. 3.
CMC, 2024, vol.81, no.1 651

Figure 3: Model architecture of ML with DLHM model

3.3.1 Random Forest (RF) with Deep Learning Hybrid Model (DLHM)
This approach, known as RF with DLHM, blends the substantial representations gleaned by
DNN with the cooperative learning capabilities of RF. This hybrid model capitalizes on the strengths
of both methods to flexibly adapt to evolving data. Notably, it excels in optimizing diagnostic precision
in the context of MS detection, particularly when confronted with limited labeled data. The model
continually enhances its understanding by incorporating ambiguous predictions via DL, proving to
be a powerful tool for the accurate and timely identification of MS patterns in clinical settings. The
synergy of RF and DL in this hybrid model leads to superior performance in MS detection. The
RF classifier has been trained with default parameters While the neural network that we have used
comprises an LSTM layer with an 8-unit Rectified Linear Unit (ReLU) activation function, a dense
layer with 64 ReLU activation function, another dense layer with 32 ReLU activation function and
output layer with a single Sigmoid activation neuron. Although multi-layered structures are proposed,
dropout layers are recommended but not applied. This model is trained using Adam optimizer and
binary cross entropy loss across 50 epoch and batch size of 32 and then evaluated using accuracy,
precision, recall, F1-score criteria.

3.3.2 Gradient Boosting (GB) with DLHM

This approach blends with DLHM to enhance the boosting capabilities with refined feature
learning. The hybrid model’s flexibility boosts diagnostic accuracy, particularly when handling limited
labeled data for MS detection. The model iteratively refines its understanding by incorporating
uncertain predictions through active learning. It is a powerful tool for precise and timely recognition of
652 CMC, 2024, vol.81, no.1

MS patterns in clinical settings. This combination showcases a concise yet effective technique to elevate
the performance of ML models in medical contexts. The gradient boosting classifier is employed
with default parameters and the output generated from this is used as feature engineering. All these
features are padded and reshaped into a matrix of shapes that can be fed into the LSTM layer. This
neural network consists of an LSTM layer with 8 units, ReLU activation, and fully connected layers
with 64 and 32 units, ReLU activation, and the output layer with one neuron, sigmoid activation. It
was recommended to use dropout layers, but this layer was not applied in the network. This model
used the Adam optimizer together with binary cross-entropy loss function with the training process
being performed for 50 epochs with a batch size of 32, and performance evaluation metrics included
accuracies, precision, recolonization, and F1-score.

3.3.3 XGB with DLHM

This approach merges the strengths of two methods: XGB’s enhanced boosting techniques and
DLHM enriched feature representations. The result is a hybrid model that adapts dynamically, leading
to better diagnostic accuracy, especially when dealing with a scarcity of labeled data for MS detection.
The model’s ability to integrate uncertain predictions through active learning continuously enhances
its understanding, making it a reliable tool for accurately and timely identifying MS patterns in clinical
settings. Combining XGB and DLHM improves the performance and versatility of ML models.
Specifically, the XGB classifier is trained by default, and its outputs are utilized as extra categories.
These combined features are padded and reshaped as required for LSTM input. They applied an
LSTM layer with 8 neurons; the next layers were dense layers, with 64 and 32 neurons activating ReLU;
and the output layer had a neuron with sigmoid activation. However, the recommended dropout layers
are not utilized here. The proposed model uses the Adam optimizer and the binary cross-entropy loss
function implemented for 50 epochs for 32 batch size and uses accurate, precise, recall, and F1-score
measurement frameworks.

3.3.4 K-Nearest Neighbors (KNN) with DLHM

This approach seamlessly blends the strengths of KNN and DLHM to achieve superior perfor-
mance in MS detection. The model’s adaptability allows it to excel even in scenarios with limited
labeled data. Its continuous learning capability enhances diagnostic accuracy by actively incorporating
uncertain predictions, making it a dependable tool for the timely and precise identification of MS pat-
terns in clinical settings. This ingenious integration of KNN and DLHM showcases a straightforward
yet effective technique for boosting the efficiency and adaptability of ML models, demonstrating its
potential for widespread use across various domains. The KNN model is trained with default hyper
parameters, and the predictions resulting from this are added as new variables. All these features are
padded and reshaped to suit the input dimension expected of the LSTM layer. The architecture of
the neural network comprises of an LSTM layer that has 8 neurons with ReLU activation, two fully
connected layers with ReLU activation, one with 64 neurons and the other with 32 neurons, and an
output layer with one neuron having sigmoid activation. While it may be daunting to see that dropout
layers are recommended, none of it is implemented in the final model. The neural network is compiled
with Adam optimizer, binary cross-entropy loss function to train the model through 50 epochs, 32
batch size, and then test the model on the test data set using accuracy, precision, recall, F1-score.

3.3.5 Decision Tree (DT) with DLHM

This improved hybrid model merges DT and DLHM for diagnosing MS efficiently and accurately.
The model’s strength lies in combining the interpretability of DT with the sophisticated feature
CMC, 2024, vol.81, no.1 653

learning abilities of DLHM. Its flexibility allows it to handle sporadic labeled data and generate
reliable predictions by incorporating uncertain predictions through active learning. This makes it
a valuable tool for clinical settings, where precise detection and clear-cut, understandable methods
are crucial. It offers a powerful machine-learning solution that aids in accurate MS detection while
being comprehensible for medical professionals. The DT classifier is then fitted with the default
parameters and the predictions obtained are used as the features. These combined features are padded
and reshaped to fit the dimensions of the LSTM layer’s input. The network is comprised of LSTM
with 8 neurons (ReLU activation), and then two dense layers, with 64 and 32 neuronal units (ReLU
activation), as well as the final output layer with a single neuron (sigmoid activation). However, for this
setup there are some dropout layers suggested but they are not used in this context. To train the model
we use the Adam optimizer and binary cross-entropy loss, we train the model for 50 epochs utilizing
a batch size of 32 and the model is evaluated by the accuracy, precision, recall, F1-score metrics.

3.4 Deep Learning with Hybrid Model

In ML, DL is a crucial task that entails developing and deploying applications, particularly deep
roots. Adding additional layers of deep architectures to these networks is what the “Deep” aspect
means. Because of the way the human brain is structured, neurons are interconnected, multilayered
networks of inputs, stores, and outputs. Enabling the model is the main objective of DL. These models
are very good at extracting the hierarchy of objects and images through deep architecture, making it
possible to view intricate shapes.
Neurons in the input layer of the architecture are oriented according to the dataset’s properties.
The 53 neurons in Hidden Layers 1 and 2 add to the intricacy. Overfitting is prevented by dropout
layers in both layers with a 0.1 rate. One neuron in the output layer designed for binary classification
has a sigmoid activation function. This approach uses dropout layers to improve generalization
while balancing resilience and complexity, as shown in Fig. 4. It has two hidden layers, which have
53 neurons, each having ‘tanh’ activation with normal kernel initializer and has L2 = 0.0 on the
regularizer. This is followed by adding ‘1’ to each of the nodes in the succeeding hidden layers: The
hybrid model used the last fully connected layer followed by a ‘sigmoid’ activation function for binary
classification. They use the Adam optimizer with a learning rate of 0.01, binary cross-entropy loss
function for training and 50 epochs with a batch size of 32. Model selection is performed using model
checkpoint through the best validation accuracy, and the metrics for the best model include accuracy
and log loss.

3.4.1 Long Short-Term Memory (LSTM)

DL approach, particularly with Long Short-Term Memory (LSTM) networks, is a game-changing
tool for handling data that unfolds over time. LSTMs, a specialized form of recurrent neural
networks, excel at capturing long-term dependencies and patterns hidden within sequential data.
This powerful technology is being leveraged in the medical realm to advance the diagnosis of MS.
LSTM, an ANN, improves the ability to detect MS by optimizing sensitivity and precision. This
research establishes a new benchmark for using LSTMs and DL in the intricate task of MS pattern
recognition. It demonstrates how using these sophisticated models can greatly enhance the accuracy
of neurodegenerative disease diagnoses.
654 CMC, 2024, vol.81, no.1

Figure 4: Model architecture of ANN with hybrid model

3.4.2 Fused ANN-DNN (FAD)

FAD is a fusion of two DL models, ANN and DNN. The combination of these two models helps
in better data processing and gives better results than other ML models used for MS detection. In this
technique, the output of the DNN model is the input of the ANN model. This technique helped us
achieve better accuracy because of its robustness and efficiency. LSTM layers within ANN architecture
enable the FAD model transformation of temporal data since they can detect sequential patterns in
the data, thereby helpful when dealing with time-stamped or sequential data sets. Moreover, proper
utilization of innovations like L2 regularization and drop-out properly addresses the problem of over
fitting and, therefore, the development of a model that has a high degree of generality. Finally, the
ideas of the careful standardization or padding of inputs need to be adhered to for uniform processing
while preserving temporal interactions and dependencies, as it was mentioned before. The interaction
between these factors leads to the FAD model’s greater accuracy and F1-scores than the other models.
ANN and DNN have a significant role by using the characteristic whose objective is to identify
patterns in medical data. Another aspect illustrates how these neural networks can be applied to
increase the level of MS identification. First, the features are normalized utilizing the standard scaler,
which helps balance the measures’ contributions to the model’s performance. The next step in the
construction of the model is to create the ANN model with two hidden layers, each of the hidden layers
contains 53 neurons and uses the ‘tanh’ activation function and dropout layer to reduce overfitting of
the model. To increase generalization regularized methods are used. The last dense layer activation
function, sigmoid function is used in the final layer to produce binary classification such as presence
of MS after the integrated outputs of ANN have been further processed. The given hybrid model is
CMC, 2024, vol.81, no.1 655

created using the Adam optimizer which modifies the learning rate during training, and the binary
cross entropy loss function which is perfect for the classification problems. To rightly select the
best model, a Model Checkpoint callback is applied during the training process for the viewport of
saving the model with the maximum validation accuracy. From the training process, the validation
performance is checked, and the weights must be tweaked to minimize the loss and thus maximize the
accuracy. After training the model which has the best accuracy is selected and is tested on the test set
to check its efficiency. The performance of the model is evaluated using metrics of accuracy and of the
log loss in the case of the binary classification.

3.4.3 Artificial Neural Network (ANN)

ML has transformed significantly with the introduction of ANNs, drawing inspiration from the
intricate framework of the human brain. In this section, we delve into the fundamental compo-
nents of ANNs, unraveling their tiered arrangement encompassing input and output layers, hidden
layers, weights, biases, and activation functions. Our attention is then turned towards the training
mechanisms, highlighting the iterative adjustments made through algorithms like backpropagation.
Understanding these mechanisms is crucial in grasping the adaptability of ANNs. The versatility of
ANNs is evident through their numerous applications. They have greatly improved image recognition
techniques and have been critical in developing natural language processing with sentiment analysis
and machine translation tools. Beyond that, ANNs have made significant strides in facial recognition
and medical image analysis, transforming how we approach these tasks. Furthermore, their presence
is felt in the financial world, where they play a crucial role in evaluating credit risks, predicting stock
prices, and detecting fraudulent activity.

3.4.4 Deep Neural Network (DNN)

DNNs are a revolutionary development in AI characterized by a complex, multilayered hierar-
chical structure. In the framework of my work, this section sheds light on the architectural details
of DNNs, describing them as a mathematical assembly of non-linear transformations that can extract
subtle patterns and features from large, intricate datasets. The investigation goes into the revolutionary
uses of DNNs in various fields, which aligns with my research’s particular emphasis. DNNs play a key
role in computer vision tasks like medical image analysis, object detection, and image classification,
all of which are directly related to the main goals of my research. DNNs have proven to be an
interesting subject in the research, with applications in robotics, healthcare, finance, and natural
language processing offering promising avenues for future investigation. By exploring the intricacies of
DNNs, including interpretability and ethical issues, I can better match the possible difficulties of using
these cutting-edge technologies with my research objectives. This subsection thoroughly explains DNN
applications and demonstrates their relevance to my research. Understanding the complex nature of
DNNs is essential for researchers as the field of DL and AI develops, and the knowledge acquired
from this section will surely be a useful tool in navigating this cutting-edge area within the research
framework being conducted.

4 Experimental Analysis and Results

The section dedicated to the experimental analysis and results evaluates the effectiveness of
different classifiers for MS detection. It investigates conventional models such as DLHM-RF, DLHM-
XGB, and a recently developed FAD. The analysis compares the performance of these classifiers
and provides insights into their advantages and disadvantages. It aids in the ongoing development
656 CMC, 2024, vol.81, no.1

of accurate and reliable tools for MS diagnosis. In ML, a performance matrix encompasses an array
of metrics that guide evaluating and gauging a model’s ability to predict outcomes on a given dataset.
These metrics include accuracy, precision, recall, F1-score, and other relevant measures, using which
one can also analyze the model’s generalization capabilities on novel data. Performance metrics aim
to objectively assess a model’s strengths and weaknesses, enabling practitioners to make informed
decisions. Ultimately, this empowers them to ensure that the model fulfills the requirements of the
given application. Model accuracy measures how well an ML model classifies instances, expressed as
the percentage of correctly classified ones. Model loss is a numerical value reflecting the discrepancy
between a model’s predicted outputs and targets. Lower values indicate less error and better model
performance. In ML, particularly in detection tasks, the performance of models is measured by their
accuracy (correct predictions) and loss (total errors). Striking a balance between high accuracy and
low loss is crucial for building detection systems that are both accurate and efficient. The model’s
loss is minimized, as shown in Fig. 5, resulting in a smaller difference between predicted and actual
values. This dual trend suggests the FAD continuously learns and optimizes its parameters, leading to
improved overall performance.

Figure 5: Accuracy and loss curves

Evaluating the performance of a model is crucial, where accuracy serves as a fundamental

indicator that measures the percentage of accurately predicted cases. Throughout the model’s training
process, it is guided by the concept of model loss. This metric represents the difference between the
expected and actual values and aids in improving its precision. The Receiver Operating Characteristic
(ROC) curve visually represents a model’s ability to distinguish between classification thresholds,
highlighting the delicate balance between true positives and false positives. Additionally, the confusion
matrix offers a comprehensive breakdown of the model’s predictions, providing insights on false
positives, false negatives, true positives, and true negatives.
This approach relies on the confusion matrix as a crucial tool for evaluating the model’s classi-
fication performance. The confusion matrix offers a detailed breakdown of the predicted and actual
classifications, allowing for a thorough understanding of the strengths and weaknesses of the model.
Notably, observations reveal incidences of misclassifications of DLHM-RF in Fig. 6. DLHM-GB and
DLHM-RF models contributing to a lower overall accuracy rate. It could be attributed to the data’s
complexity or specific challenges these models face in accurately recognizing certain patterns. The data
in the dataset consists of binary classification, consisting of 2 classes for multi-sclerosis detection. FAD
CMC, 2024, vol.81, no.1 657

is remarkable for its unusually high precision in classification. It can discern and adapt to complex data
patterns better than DLHM-RF and DLHM-GB in this application. The multi-sclerosis detection
framework using FAD underscores its potential as a robust and accurate tool for classification tasks.
Understanding the unique features contributing to FAD’s exceptional performance can offer valuable
insights for optimizing model selection in similar diagnostic scenarios. In Fig. 7, Receiver Operating
Characteristic (ROC) curves explain model performance at various classification levels. For instance,
the ROC curve for FAD showed strong performance for all classes.

Figure 6: Confusion matrix for ML classifiers

658 CMC, 2024, vol.81, no.1

Figure 7: ROC for ML classifiers

We assessed the efficiency of various classifiers in detecting MS using a MS dataset. Models like
RF, GB, XGB, KNN and DT were integrated with a deep-learning hybrid model that provided the
highest accuracy of 89.65% by XGB. Meanwhile, ANN was fused with another DNN (FAD), which
provided even better results, as shown in Table 2. The accuracy, precision, recall, and F1-scores were
used to measure each classifier’s performance. The results revealed that the efficiency of the classifiers
differed. Impressively, the FAD exceeded traditional ML classifiers, achieving an outstanding accuracy
of 96.55%. This outcome shows how DL approaches can enhance the accuracy and effectiveness of
MS detection systems.

Table 2: Evaluation results

Model Accuracy (%) Precision (%) Recall (%) F1-score (%)
RF-DLHM 79.31 80.00 88.89 84.21
GB-DLHM 79.31 75.00 100 85.71
XGB-DLHM 89.65 87.5 100 93.33
KNN-DLHM 82.76 84.21 98.89 86.49
DT-DLHM 86.21 90.48 90.48 90.48
FAD 96.55 96.48 96.55 96.71

With an accuracy of 96.55%, FAD demonstrated the best overall performance among the
traditional classifiers. Given its high recall (96.55%) and precision (96.48%), it is a good option for
MS detection. With a balanced precision-recall trade-off and an accuracy of 89.65% accuracy, XGB-
DLHM also demonstrated competitive performance.
CMC, 2024, vol.81, no.1 659

5 Discussion
The hybrid model’s success in integrating ANN can be attributed to its combination of depth
and traditional DL strengths. Two methods are proposed for binary classification tasks. First is a
traditional neural network approach with dropout layers and standardization for better adaptation
and generalization, The FAD model aims to reduce over-fitting, meanwhile achieving an exceptional
accuracy of 96.55% surpassing the base approach with the margin of 10.35%. The second method
presents a hybrid model that combines different ML models with deep learning hybrid model, which
can detect subtle patterns for classification. The choice of method depends on the dataset’s features,
allowing for a customized approach based on specific requirements. This combination allows the
Hybrid Model to identify complex patterns in MS-related data, adapting and providing consistently
high performance across various datasets. The ML models implemented with the hybrid model include
RF-DLHM, GB-DLHM, XGB-DLHM, KNN-DLHM and DT-DLHM with the highest achieved
accuracy of 79.31%, 79.31%, 89.65%, 82.76% and 86.21%. The XGB-DLHM results also surpassed the
base paper results with a margin of 3.45%. The model’s resilience and adaptability make it more reliable
and accurate, cementing its role as a valuable tool in MS detection. The model’s results are evident
as shown in Table 3. In essence, the progress of MS detection relies on continuous advancements
in modelling, embracing new technologies, and forging collaborations to guarantee our findings’
clinical relevance and impact. Through these unified efforts, our goal is to significantly contribute to
developing accurate and reliable methods that aid in the early diagnosis and management of MS. The
provided approach has the potential to be improved for more accurate results by using this approach
with other various approaches. It can be implemented in different domains, including healthcare,
finance, natural language processing, and others. The proposed algorithms also provided better results
than other papers as well that include the results of de Leon-Sanchez’s article [8] that yielded results
of 86%, which our proposed models passed efficiently with an accuracy of 96.55% and 89.65% from
FAD and XGB-DLHM, respectively.

Table 3: Performance comparison with base study [8]

Reference Accuracy
[8] 86.20%
FAD 96.55%

6 Conclusion and Future Scope

This paper proposed a FAD approach to diagnose MS based on the data collected from the
patients. Various machine-learning models and one deep-learning model were integrated with the
hybrid deep-learning model approach, including RF, GB, XGB, KNN, DT and ANN. After evaluation
of the dataset on these models, few of them performed better than the previous approaches, which
were FAD, XGB-DLHM and DT-DLHM, with accuracies of 96.55%, 89.65% and 86.21%. The other
models, including GB-DLHM, KNN-DLHM and RF-DLHM, provided an accuracy of 79.31%,
82.76%, and 79.31%, respectively. The results imply that our approach can provide more accurate
results on this dataset. The disadvantage is that if a user applies combining ML models and LSTM
neural networks, then the possible computation complexity will bring along universal training or
inference actions. Thus, it will take a long time and many resources to finish. On the same front,
pre-trained models or ensemble approaches also receive interpretability challenges. Hence, they need
660 CMC, 2024, vol.81, no.1

in-depth re-training to make them redundant on newly inputted data and domains. These problems
can only be solved by the implementation of optimization techniques, which are very complicated and
require considerable effort. The approach can be further improved by using other ensemble methods
and by implementing different sizes of the bidirectional LSTM network as well as their parameters
according to the changed dataset. We provide an opportunity to perfect the knowledge gained in the
part of consistently refining and expanding the application range of this innovative method in a world
where a variety of situations are possible.

Acknowledgement: Princess Nourah bint Abdulrahman University Researchers Supporting Project

number (PNURSP2024R503), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia.

Funding Statement: This study was supported by Princess Nourah bint Abdulrahman University
Researchers Supporting Project number (PNURSP2024R503), Princess Nourah bint Abdulrahman
University, Riyadh, Saudi Arabia.

Author Contributions: The authors confirm contribution to the paper as follows: study conception
and design: Stephen Ojo, Moez Krichen, Meznah A. Alamro, Jaroslava Kniezova; data collection:
Stephen Ojo; analysis and interpretation of results: Stephen Ojo, Moez Krichen, Meznah A. Alamro,
Alaeddine Mihoub Author; draft manuscript preparation: Stephen Ojo, Moez Krichen, Meznah A.
Alamro, Alaeddine Mihoub, Gabriel Avelino Sampedro, Jaroslava Kniezova. All authors reviewed the
results and approved the final version of the manuscript.

Availability of Data and Materials: Data available within the article.

Ethics Approval: Not applicable.

Conflicts of Interest: The authors declare that they have no conflicts of interest to report regarding the
present study.

References
[1] J. J. Cerqueira et al., “Time matters in multiple sclerosis: Can early treatment and long-term follow-up
ensure everyone benefits from the latest advances in multiple sclerosis?,” J. Neuro., Neuro. Psy., vol. 89, pp.
844–850, 2018. doi: 10.1136/jnnp-2017-317509.
[2] H. Tahernia, F. Esnaasharieh, H. Amani, M. Milanifard, and F. Mirakhori, “Diagnosis and treatment of
MS in patients suffering from various degrees of the disease with a clinical approach: The original article,”
J. Pharm. Negat. Results, pp. 1908–1921, 2022.
[3] E. Waubant et al., “Environmental and genetic risk factors for MS: An integrated review,” Anna. Clin.
Trans. Neuro., vol. 6, no. 9, pp. 1905–1922, 2019.
[4] B. Solomon, P. Boruta, D. Horvath, and J. MacKellar, “The multiple sclerosis stress equation,” J. Med.
Stat. Inf., vol. 11, p. 1, 2023. doi: 10.7243/2053-7662-11-1.
[5] K. O. Mohammed Aarif, A. Alam, Pakruddin, and J. Riyazulla Rahman, “Exploring challenges and
opportunities for the early detection of multiple sclerosis using deep learning,” Arti. Intell. Auto. Dis.: Appl.
Diagn., Prognosis, Ther., pp. 151–178, 2024.
[6] C. S. Constantinescu, N. Farooqi, K. O’Brien, and B. Gran, “Experimental Autoimmune Encephalomyelitis
(EAE) as a model for Multiple Sclerosis (MS),” Br. J. Pharmacol., vol. 164, no. 4, pp. 1079–1106, 2011. doi:
10.1111/j.1476-5381.2011.01302.x.
[7] C. M. Hersh, Multiple Sclerosis, An Issue of Neurologic Clinics. vol. 42. Philadelphia, Pennsylvania: Elsevier
Health Sciences, 2023.
CMC, 2024, vol.81, no.1 661

[8] E. R. P. De Leon-Sanchez, A. M. Herrera-Navarro, J. Rodriguez-Resendiz, C. Paredes-Orta, and J. D.

Mendiola-Santibañez, “A deep learning approach for predicting multiple sclerosis,” Micro, vol. 14, no. 4.
2023, Art. no. 749.
[9] N. Aslam et al., “Multiple sclerosis diagnosis using machine learning and deep learning: Challenges and
opportunities,” Sensors, vol. 22, no. 20, 2022, Art. no. 7856. doi: 10.3390/s22207856.
[10] S. F. Maleki et al., “Artificial Intelligence for multiple sclerosis management using retinal images: Pearl,
peaks, and pitfalls,” Semi. Ophtha., Tay. Fra., vol. 39, no. 4, pp. 271–288, 2024.
[11] A. Shoeibi et al., “Applications of deep learning techniques for automated multiple sclerosis detection
using magnetic resonance imaging: A review,” Comp. Bio. Med., vol. 136, 2021, Art. no. 104697. doi:
10.1016/[Link].2021.104697.
[12] T. S. Daqqaq, A. S. Alhasan, and H. A. Ghunaim, “Diagnostic effectiveness of deep learning-based
MRI in predicting multiple sclerosis: A meta-analysis,” Neuro. J., vol. 29, no. 2, pp. 77–89, 2024. doi:
10.17712/nsj.2024.2.20230103.
[13] S. Denissen et al., “Towards multimodal machine learning prediction of individual cognitive evolution in
multiple sclerosis,” J. Person. Med., vol. 11, no. 12, 2021, Art. no. 1349. doi: 10.3390/jpm11121349.
[14] R. Liu, R. Yan, and Z. Peng, “A review of medical artificial intelligence,” Glo. Heal. J., vol. 4, no. 2, pp.
42–45, 2020. doi: 10.1016/[Link].2020.04.002.
[15] A. Gautam et al., “Investigating gene expression profiles of whole blood and peripheral blood mononuclear
cells using multiple collection and processing methods,” PLoS One, vol. 14, no. 12, 2019, Art. no. e0225137.
doi: 10.1371/[Link].0225137.
[16] M. Navaderi et al., “Identification of multiple sclerosis key genetic factors through multi-staged data
mining,” Mult. Scle. Rel. Dis., vol. 39, 2020, Art. no. 101446. doi: 10.1016/[Link].2019.101446.
[17] C. Chio and D. Freeman, Machine Learning and Security: Protecting Systems with Data and Algorithms,
1st ed. Sebastopol, CA, USA: O’Reilly Media, Inc., 2018.