An ML model for the detection of brain-related
diseases
S.Ravikanth
Senior Assistant Professor
Department of ECE
CVR College of Engineering
Email;s.ravikanth@cvr.ac.in
Abstract—The brain, as the most complex part of the body,
operates as the central control system, rendering it susceptible
to a broad spectrum of diseases. This continuous unveiling of
new challenges in diagnosis and detection establishes an
evolving landscape for brain-related disorders, posing an
ongoing research frontier. Early identification of these diseases
holds pivotal importance for effective treatment, and the
incorporation of artificial intelligence (AI) has markedly
revolutionized the field of neurology. This review extensively
explores recent advancements in machine learning and deep
learning techniques dedicated to detecting Alzheimer's disease
(AD), epilepsy, and Parkinson's disease. The overarching goal
of the review is to pinpoint the most accurate techniques for the
detection of various brain diseases, providing valuable
perspectives for future improvements in diagnostics.
Keywords—Alzheimer’s disease, deep learning, epilepsy
seizures, Parkinson’s disease, machine learning.
I. INTRODUCTION
The increasing prevalence of brain-related diseases, coupled
with the aging global population, has underscored the need
for robust and efficient diagnostic tools. Alzheimer's Disease,
characterized by progressive cognitive decline, Parkinson's
Disease, associated with motor impairments, and epileptic
seizures, causing unpredictable neurological events,
collectively represent a significant burden on healthcare
systems worldwide. Traditional diagnostic methods for these
disorders often rely on clinical assessments, which may be
subjective and prone to variability. The integration of
medical imaging, particularly Magnetic Resonance Imaging
(MRI) and Electroencephalography (EEG, provides valuable
insights into the structural and functional aspects of the brain.
However, the sheer volume and complexity of data generated
necessitate using advanced computational techniques.
Machine learning algorithms have demonstrated the ability to
analyze intricate patterns within medical imaging data,
enabling the development of accurate and reliable diagnostic
models. In this paper, we propose a comprehensive ML
model designed to detect and differentiate between
Alzheimer's Disease, Parkinson's Disease, and epileptic
seizures based on multi-modal neuroimaging data. In
subsequent sections, we delve into the methodology,
experimental setup, and results, presenting a thorough
evaluation of the proposed ML model's performance. The
ultimate goal of this research is to provide clinicians with a
reliable and efficient tool for the early detection and
monitoring of Alzheimer's Disease, Parkinson's Disease, and
epileptic seizures, thereby advancing the frontier of
personalized and precision medicine in neurology.
II. DATA COLLECTION
A. Alzheimer’s disease
The MRI data employed in this study is retrieved from the
Open Access Series of Imaging Studies (OASIS) portal,
accessible at https://www.oasis-brains.org. The meticulously
curated dataset comprises a diverse collection of images
representing various levels of dementia. Specifically, the
dataset encompasses 5,002 images corresponding to the mild
dementia category, 488 images representing the moderate
dementia category, 67,222 images associated with the non-
demented category, and an additional 13,725 images
belonging to the very mild dementia category. In aggregate,
this dataset encompasses a substantial collection of 86,437
MRI images, providing a comprehensive and varied resource
for the study of dementia across different severity levels.
B. Parkinson’s disease
The numerical data in this study is retrieved from the
repository of the Donald Bren School of Information and
Computer Sciences (ICS) portal accessible at
https://archive.ics.uci.edu...This dataset comprises
biomedical voice measurements obtained from 42 individuals
in the early stages of Parkinson's disease participating in a
six-month trial for remote symptom monitoring through a
telemonitoring device. The voice recordings, totaling 5,875,
were automatically collected in the participants' homes. The
table includes columns for subject number, age, gender, time
interval from baseline recruitment, motor UPDRS, total
UPDRS, and 16 biomedical voice measures. Each row
represents a distinct voice recording from the study
participants.
C. Epilepsy seizures
The numerical data in this study is retrieved from the
repository of the Donald Bren School of Information and
Computer Sciences (ICS) portal accessible at
https://archive.ics.uci.edu...The initial dataset, as described
in the reference, is organized into five folders, each
containing 100 files representing individual subjects. Each
file corresponds to a 23.6-second recording of brain activity,
sampled into 4097 data points. In total, there are 500
individuals, and each person has 4097 data points capturing
brain activity over 23.5 seconds. The 4097 data points were
rearranged and randomized into 23 chunks to enhance the
dataset. Each chunk now comprises 178 data points,
representing a 1-second interval, with each data point
reflecting the EEG recording at a specific moment.
Consequently, the dataset has been transformed into 11,500
pieces of information, combining the 23 chunks for each of
the 500 individuals.
 Since transformers don't have any inherent
understanding of spatial relationships, positional
III.METHODS information needs to be injected into the input data. This
is usually done by adding positional encodings to the
A. Vision Transformer input vectors. The positional encoding PE is defined as:
A CNN model named Vision Transformer is used for PE
building the model. Departing from conventional
( )
pos
(pos ,2 i)=sin 2i
Convolutional Neural Networks (CNNs), the Vision 10000 d
Transformer employs a transformer architecture with
self-attention mechanisms to capture global dependencies PE
( )
pos
(pos ,2 i+1 )=cos
within images. Rather than processing data in fixed-size 10000 d
2i

grids, the ViT divides images into manageable patches,

embedding them linearly and enabling the model to
comprehend relationships across the entire input space.
This departure from traditional CNN structures eliminates
the reliance on convolutional layers and proves highly
scalable, particularly with larger image resolutions. The
Vision Transformer has exhibited exceptional
performance in various computer vision tasks, surpassing
conventional architectures and establishing its potential to
redefine visual recognition and understanding.
B. Support vector Regression
For Parkinson’s disease and epilepsy seizures, we used
the Support Vector Regression method. Support Vector
Regression (SVR) is a powerful machine learning
algorithm designed for regression tasks, offering robust
capabilities in modelling complex relationships between
variables. Based on the principles of Support Vector
Machines (SVM), SVR excels in predicting continuous
outcomes by identifying a hyperplane that best captures
the trend of the data. Unlike traditional regression
models, SVR is effective in handling non-linear patterns
through the use of kernel functions that transform the
input space. The algorithm aims to minimize the error
between predicted and actual values while allowing for a
controlled margin of tolerance. SVR's strength lies in its
ability to adapt to diverse data distributions and its
resistance to overfitting, making it particularly suitable
for datasets with intricate structures.
MultiHead (Q, K, V) = Concat (head 1 ,..... , head h)
Wo,
IV.WORKING
A. Vision Transformer
Where pos is the position and i is the dimension index,
and d is the dimensionality of the input vectors.
3. Concatenate positional encoding:
concatenate the positional encoding to the input vectors:
^x i=[ x 'i , PE i ] ,
d +d pos
here ^x i ∈ R is the extended input vector and d pos is
the dimensionality of the positional encoding.
4. Linear Projection for Token embedding:
Linearly project the extended input vectors into the token
space:
z i=W z ^xi +b z ,
Where W z is the learnable weight and b z is a learnable
bias.
5. Transformer encoder blocks:
Apply a series of transformer encoder blocks to the token
embeddings z i . Each block consists of a multi-head self-
attention mechanism and a feedforward neural network.
Let h be the number of attention heads, d model be the
model dimension, and d ff be the dimension of the
feedforward layer.
head i = Attention (QW Qi , KW Ki, VW Vi ),

(√ )
A Vision Transformer (ViT) is a type of neural network Q KT
architecture that applies the transformer model, originally
designed for natural language processing, to image data. d model
The ViT model can be represented mathematically as Attention (Q, K, V) = SoftMax V,
h
follows:
1. Input representation: FFN ( x ¿ = ReLU ( x W ff ,1 +b ff ,1 ) W ff ,2+b ff ,2 ,

Let X be the input image, which is divided into non- Where W Qi , W Ki, W Vi , W o, W ff ,1, W ff ,2, b ff ,1 and
overlapping patches x i. Each patch is then linearly b ff ,2 are learnable parameters.
' d
embedded into a vector x i ∈ R using a learnable linear 6. Output Projection:
projection.
Linearly project the transformer output into the final
X ={x 1 , x 2 , … , x N } output space:
'
x i=W 0 x i+ b0 , Y = W y z N +b y,

W 0 is a learnable weight matrix and Where W y is a learnable weight matrix and b y is a

Where
learnable bias.
b 0 is learnable bias .
2. Positional encoding:
 The model takes an input image, processes it through 1
linear projections, adds positional encodings, applies b=
transformer encoder blocks, and outputs the final result. ¿ S∨¿ ∑ ( y i−⟨ w , x i ⟩ ) ¿ ,
i ∈S
B. Support Vector Regression Where S is the set of support vectors.
Support Vector Regression (SVR) is a machine learning
algorithm used for regression tasks. The key idea behind This formulation captures the essence of linear Support
SVR is to find a hyperplane that best represents the data Vector Regression. In practice, kernels can be employed to
while minimizing the error within a specified margin. handle non-linear relationships by mapping the input features
into a higher-dimensional space.
1. Data Representation:
Assume we have a dataset with N samples:
V. RESULTS
{(x ¿ ¿ 1 , y 1 ),( x ¿ ¿ 2 , y 2) , … ,(x ¿ ¿ N , y N )}¿ ¿ ¿ A. Alzheimer’s disease
Here, the MRI images are uploaded in 128 X 128 images
Where x iis the input feature vector and y iis the into the machine learning model. The model processes the
corresponding target value. image and gives the output results. The machine learning
model gives you the classification of dementia depending
2. Linear SVR model: upon the input MRI data.
The linear SVR model can be represented as:
f ( x )= ⟨ w , x ⟩ + b ,
Where w is the weight vector, x is the input feature
vector, ⟨ . ,. ⟩ denotes the dot product and b is the bias
term.
3. Objective Function (Cost Function):
The goal of SVR is to find a hyperplane that
minimizes the error within a specified margin (ε). The
objective function for linear SVR is given by:
N
1
min w ,b ‖w‖ +C ∑ (max ⁡(0 ,|f ( x i )− y i|−ε )),
2

2 i=1

Where ‖w‖ is the L2 norm of the weight vector, C

is the regularization parameter and ε is the margin.
Figure 1
4. Support Vectors:
In figure 1, the first image is that of mild dementia, the
Support vectors are the data points that lie on the margin second image is that of moderate dementia, the third image is
or violate the margin (have non-zero errors). They play a of no dementia and the final image is of very mild dementia.
crucial role in defining the hyperplane. The decision Alzheimer’s disease is categorized towards the second case
function is influenced only by the support vectors. i.e. moderate dementia.
5. Dual Problem and Lagrange Multipliers:
B. Parkinson’s disease
The optimization problem is often converted into its dual Here, using the data available from the dataset, we plotted
form. Introducing Lagrange multipliers (α i) for the a graph which is between the actual readings of UPDRS
inequality constraints, the dual problem is: versus the predicted ratings of the UPDRS. Here UPDRS
N N
stands for ‘Universal Parkinsons Disease Rating System’.
1 This method is used to quantify the values related to the
max α ∑ α i− ∑ α α y y ⟨x ,x ⟩ diagnosis of the Parkinson’s disease.
i=1 2 i , j=1 i j i j i j ,
N

Subject to 0
≤ αi ≤ C
and ∑ αi y j=0.
i=1

6. Dual Problem and Lagrange Multipliers:

The weight vector w can be expressed as a linear
combination of the support vectors:
N
w=∑ α i y i x i.
i=1

The bias term b is determined using the support vectors

and the margin:

Figure 2
Here in figure 2, we have actual UPDRS in the x-axis and
predicted UPDRS in the y-axis. There is a positive
correlation between the predicted and actual total UPDRS
values. This means that as the predicted values increase, the
actual values also tend to increase.
C. Epilepsy Seizures:
Here, we got a confusion matrix from the above given
data from the obtained graph.
Figure 3
Figure 3 is a graph with the count on the y-axis and the
predicted label on the x-axis.
Here, there are parameters calculated such as True
Positives (TP) to find correctly predicted seizures, True
Negatives (TN) to find correctly non-seizures, False
Positives (FP) to find Non seizures incorrectly predicted as
seizures, and False Negatives (FN) to find seizures
incorrectly predicted as non-seizures respectively.
Here we get a matrix of 5 rows and 180 columns from the
extensive dataset which is reduced using the SVR method,
and we get the confusion matrix as follows:
[ 8313 109 ]
confusion ¿
Here, we got True Positives as 83, True Negatives as 10,
False Positives as 13, and False Negatives as 9.
After substituting all the values, we get the accuracy of the
system using the formula:
Accuracy = (TP + TN) / (TP + TN + FP + FN). Which is
equal to 81%. We also find out the values of Specificity and
Sensitivity which are obtained by calculating the ratios of
true positives and true negatives to the total number of
instances in each class.
Specificity (For the first class) = TP / (TP + TN) = 89.2%
Sensitivity (For the second class) = FP / (FP + FN) =40.9%
The F1 score which is used for evaluation of the performance
of the model is approximately 80%.
CONCLUSION
In this study, we presented a comprehensive
exploration into the development and evaluation of a
machine learning model designed for the detection of
brain-related diseases. Leveraging advanced
techniques in feature extraction, model training, and
validation, our proposed model has demonstrated
promising results in accurately classifying and
diagnosing such diseases. However there is a scope of
improvement as these models have a large scope of
scalability and need for extensive research and
development.
REFERENCES
[1] C. S. Eke, E. Jammeh, X. Li, C. Carroll, S. Pearson and E. Ifeachor,
"Early Detection of Alzheimer's Disease with Blood Plasma Proteins
Using Support Vector Machines," in IEEE Journal of Biomedical and
Health Informatics, vol. 25, no. 1, pp. 218-226, Jan. 2021, doi:
10.1109/JBHI.2020.2984355.
[2] P. Khan et al., "Machine Learning and Deep Learning Approaches for
Brain Disease Diagnosis: Principles and Recent Advances," in IEEE
Access, vol. 9, pp. 37622-37655, 2021, doi:
10.1109/ACCESS.2021.3062484.
[3] A. AlMohimeed et al., "Explainable Artificial Intelligence of Multi-
Level Stacking Ensemble for Detection of Alzheimer’s Disease Based
on Particle Swarm Optimization and the Sub-Scores of Cognitive
Biomarkers," in IEEE Access, vol. 11, pp. 123173-123193, 2023, doi:
10.1109/ACCESS.2023.3328331.
[4] A. Mitoubsi et al., "Evaluating the Fitness-to-Drive Using Evoked
Visual Responses in Alzheimer’s Disease," 2021 43rd Annual
International Conference of the IEEE Engineering in Medicine &
Biology Society (EMBC), Mexico, 2021, pp. 2382-2385, doi:
10.1109/EMBC46164.2021.9630321.
[5] G. G.S and N. S, "A Comprehensive Review on Early Diagnosis of
Alzheimer’s Disease Detection," 2023 International Conference on
Research Methodologies in Knowledge Management, Artificial
Intelligence and Telecommunication Engineering (RMKMATE),
Chennai, India, 2023, pp. 1-6, doi:
10.1109/RMKMATE59243.2023.10369111.
[6] M. Shetty, Deekshitha, M. Bhat and M. Devadiga, "Detection of
Alzheimer's Disease Using Machine Learning," 2022 International
Conference on Artificial Intelligence and Data Engineering (AIDE),
Karkala, India, 2022, pp. 117-120, doi:
10.1109/AIDE57180.2022.10060433.
[7] A. R. Zouaoui, H. Bentahar, M. Djeriuoi, Z. Mokadem and H. E.
Bentahar, "Machine Learning and Deep Learning Techniques for
Alzheimer's Disease Prediction Using CSF and Plasma Biomarkers,"
2023 International Conference on Networking and Advanced Systems
(ICNAS), Algiers, Algeria, 2023, pp. 1-6, doi:
10.1109/ICNAS59892.2023.10330506.
[8] V. Srilakshmi, A. Anumolu, M. A. Safali and V. S. Parvathi, "A
Hybrid-Layered Framework for Detection and Diagnosis of
Alzheimer’s Disease (AD) from Fundus Images," 2023 Third
International Conference on Artificial Intelligence and Smart Energy
(ICAIS), Coimbatore, India, 2023, pp. 1639-1643, doi:
10.1109/ICAIS56108.2023.10073930.