Early Detection of Alzheimer’s Disease Using Machine Learning Techniques

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 07 | July 2022 www.irjet.net p-ISSN: 2395-0072
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 2719
Early Detection of Alzheimer’s Disease Using Machine Learning
Techniques
Debabrata Sahoo1, Bijaya Kumar Ekka2
1Department of Electronics and Instrumentations Odisha University of Technology and Research Bhubaneswar,
Odisha, India
2Professor, Department of Electronics and Instrumentations Odisha University of Technology and Research
Bhubaneswar, Odisha, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract:
Alzheimer's disease (AD) is a fatal, degenerative
brain ailment that gradually decreases people of
their ability to think clearly and recall things. A
common cause of dementia is Alzheimer's disease.
Dementia is the term used to describe the loss of
cognitive functioning, which includes thinking,
recalling, and reasoning, as well as behavioural skills
to the point where it affects daily life. To identify
diseases and aid doctors in making observations-
based decisions, image processing is frequently
employed in the medical industry. The goal of the
study is to identify Alzheimer's disease as early as
possible so that patients can be treated before their
brains experience irreversible alterations. A
significant development in Machine Learning and
Deep Learning technology leads to accurately
classifying MRI-based images. But a high benchmark
is needed while classifying Medical related tasks. A
small mistake may lead to serious health complexion
over time or may lead to fatal. We proposed a
method to uses the brain's magnetic resonance
imaging (MRI) from the coronal, axial and sagittal
planes using a Deep Learning-based Image classifier
to emphasise the detection of the damaged brain
with an accuracy of 99.5 %. Trying to compare our
suggested model to many state-of-the-art models, it
has accomplished a high-level benchmark.
Keywords- Alzheimer’s disease, Mild cognitive
impairment, Image processing , Dementia .
1. INTRODUCTION
The brain is the main organ of the human body.It is vital
to treat brain illnesses since, in most situations, once
alterations take place, they cannot be reversed unless in
rare circumstances. The loss of cognitive and practical
reasoning is known as dementia. The leading cause of
dementia is Alzheimer's disease. Mid-60s is when
Alzheimer's initially manifests. More than 6.5 million
people are thought to have Alzheimer's disease.
Memory loss, language difficulties, and behavioral
modifications are some of the symptoms of Alzheimer's
disease. Word finding problems, visual problems,
decreased reasoning, and poor judgments are the
symptoms of the non-memory aspect. The biological
indications are blood, cerebrospinal fluid, and brain
imaging. Mild Alzheimer's, moderate Alzheimer's, and
severe Alzheimer's are the three stages of the disease.
The hereditary component of early-onset Alzheimer's
disease and the complicated chain of brain changes that
lead to late-onset Alzheimer's disease are the causes. The
capacity to detect Alzheimer's disease by studying
changes in the brain, body fluids, and lifestyle are the
other reasons, along with genetics, environment, lifestyle,
and health. The aberrant protein or chemical aggregates
(amyloid plaques), tangled fibre bundles (tau tangle), and
loss of connections between nerve cells in the brain are
all symptoms of Alzheimer's disease. A decade after the
beginning of the development of the disease, the
symptoms of Alzheimer's start to show. When a healthy
neuron stops functioning, the connection with the other
neurons is lost, and the cell eventually dies. The
accumulation of amyloid plaques and protein tau tangles
in the brain is what generates this. The hippocampus, a
crucial component in establishing memories, will be the
first brain region to be damaged. The affected areas of the
brain started to shrink as it spread to other regions
gradually, and by the time it reached its ultimate stage,
the entire brain had significantly shrunk in size. [1].
There are many techniques are available but MRI and CT-
Scans are more effective in diagnosis. While MRI is more
effective to capture low-level features of the brain so MRI
is very useful for training a deep neural network. We have
trained our proposed model with MRI datasets and
compared it with different Machine Learning and Deep
Learning based State-of-the-art Models.
2. LITERATURE SURVEY
MRI scans can be utilised in image processing to
determine the likelihood of AD early detection. Intensity
adjustment, K-means clustering, and the region-growing
method for extracting white matter and grey matter are
three image processing techniques utilised in MRIs. Using
the same approach, the brain's volume may be
determined. The axial (top view), coronal (back view),
and sagittal (side view) planes of a brain MRI are

analysed quantitatively and clinically using MATLAB.
[2]. Using various image segmentation techniques,
image processing is the process of removing the Region
of Interest from the image. The K-means clustering
approach, region expanding, watershed, thresholding,
split and merge, and other techniques are used in
picture segmentation. The segmentation of X-rays of
radiographic welds with abnormalities including
porosity and the lack of fusion, inadequate penetration,
and wormhole detected is done using these
segmentation approaches. This technique is used to
identify the defective regions. In processing medical
computer vision, optical character recognition, and
imaging a radiograph for industry, they are thus
frequently implemented. [3]. One of the algorithms that
the popular clustering algorithm. This article discussion
of k-means algorithms that have been updated, such as
Applying the initial partial stretching improvement to
the picture to increase image quality. Individual cluster
is generated the cluster's first centre using and Using
subjective clustering, you can create. The means
technique is used to segment pictures using the
produced centre[4]. For AD detection, the deep learning
architecture is recommended in order to solve the
limitations of the machine learning algorithm
methodology . It can identify both instances of AD and
mild cosgnitive impairment. In order to identify the
prodromal stage of AD and MCI, it suggests a deep
learning architecture that makes use of stacked
autoencoder and softmax output layer. This
architecture may conduct detection utilising domain-
specific prior knowledge while examining several
training sample classes and less labelled training
samples [5]. One of the most fatal diseases is a brain
tumour. In the identification procedure, image
processing can be quite helpful.
2.1 Related works
The state-of-the-art for applying DL and ML algorithms
to diagnose dementia and Alzheimer's disease is
covered in this section. With the use of the pattern
similarity score, the study proposes new metrics for
diagnosing Alzheimer's disease. The conditional
probabilities predicted by logistic regression are used
by the authors to describe the metrics. Furthermore,
they investigate the effectiveness of anatomical and
cognitive impairment, which is utilised to produce the
output of the classifiers from various forms of data. To
diagnose Alzheimer's disease, the authors employ
online databases of MRI scan pictures and other
cognitive parameters, like RAVLT tests, MOCA and FDG
scores, etc[6]. In particular, methods for grouping
patients with Alzheimer's disease are created based on
logistic regression and SVM. A system based on speech
processing was provided by Ammar et al. [7] to identify
dementia. With verbal description and human
transcription of the speech data, the framework was
utilised to extract characteristics from people with
dementia and those without dementia. In order to train
ML classifiers, The speech and textual characteristics
were employed. Only 79 percent of the time did the
authors get it right. The authors of [8] provided another
intriguing piece of work in which they described a
detection technique based on brain MRI images based on
Eigenbrain. In their method, the model was trained using
SVM and particle swarm optimization. In identifying the
areas of the brain affected by Alzheimer's disease, their
plan produced good results. In a similar vein, writers in
[9] used MRI data to identify dementia and other features
using gradient boost and Artificial Neural Network (ANN)
models. Based on cognitive and linguistic aspects, the
authors [10] presented a hybrid multimodal approach.
The model was trained by the authors using ANN to
identify Alzheimer's disease and its severity. Currently,
deep learning based technology is being used in most
cases, which results in improvement of results. The
imbalance of classes is the most common problem in
these methods. Recently DNNs are replaced by CNNs for
better training time, GPU utilisation, and accuracy.
Existing models for classifications can be more complex
for employing MRI datasets.[11]
3. METHODOLOGY
We have proposed the following methodology to train the
model and compared the performance of the trained
model using the test dataset. The entire methodology is
divided into 2 parts. Part-I is the Deep learning-based
model whereas in the second part we have used Machine
Learning based models. As ML-based models are more
efficient and time for training and processing is very less
it is preferable for low-end devices. Deep Learning
models are bulky but more precise so it is preferable
instead of ML models.[12]
Fig -1: Methodology
4. DATASET
The data is collected from Kaggle, which is an open-
source platform for data scientists and machine learning

engineers to compete and collaborate to enhance their
skills. The Data is hand collected from various websites
with each and every label verified. The dataset is consist
of 4 files for each class
1. MildDemented
2. VeryMildDemented
3. NonDemented
4. ModerateDemeneted
The dataset also contains a train-set for training and a
test-set for validation. It contains around 5000 images.
Fig -2: Images from each class.
Tensorflow dataset API:- Using the tf.data API, you can
create intricate input pipelines from straightforward,
reusable parts[13]. The pipeline for an image model, for
instance, may combine information from files in a
distributed file system, make random alterations to
every image, and combine a batch of randomly chosen
photos for training. Extracting symbols from raw text
input, transforming them to embedding IDs using a
lookup table, and batching together sequences of
various lengths may all be included in the pipeline for a
text model. It is possible to manage significant volumes
of data, read from many data formats, and carry out
intricate changes thanks to the tf.data API.
DeepLearning model is trained using tf.data API with a
batch size of 32 images.
5. ALGORITHM USED
5.1 Deep Learning Based Algorithm
The human visual brain served as the inspiration for the
CNN design that we employed for this investigation. The
input stream of information is received by the human
eye in its receptive field, which is comparable to how
the input is convolved during the convolution
procedure and uses its input to operate on the image to
create the feature map. Which inspires the
Convolutional operation. A CNN consists of several
maximum layers with ReLU activation functions
completely linked layers as well as layer pooling. all
inputs are gone through various processes to arrive at
the finished product in the design of a multi- or binary
classifier. the morphing operation is shared by several
neurons and connected through them. shift-invariance,
local connectivity, and hyper-parameters enhance the
network's strength. Sometimes CNN model from scratch
is not so useful in case of lack of data so pre-trained CNN
model architecture like VGG16, and MobileNet is used.
VGG16- One of the famous model architectures which
won ILSVR-2013 and outperformed GoogleLeNet[14]. It
has achieved remarkable accuracy of 92.7 % on 1000
class images of 14 million in size.
Fig -3: The architecture of the VGG-16 Model
It has two or three convolution layers, then one pooling
layer. The same is repeated over 5-6 times and finally,
some dense layer has been added. This Dense layer is
trainable whereas the convolutions layers are non-
trainable. Trainable dense layers are used as a finetuning
layer. The input layer consists of the size of images and
the output layer is a Softmax layer whose unit is decided
based on the number of classes.
5.1.1 MobileNet
It is a very lightweight computer vision model intended
for very low-end devices[15]. It uses the method of
depthwise separable convolution methodology and
significantly reduces the parameters as compared to the
normal CNN model but it achieves remarkable
performances.

Fig -4: Architecture of Mobilenet
Fig -5: Depthwise Convolution Layer (B)
Depthwise convolution contains less no. of parameters
as compared to normal convolutions.
The model is followed by multiple layers of depthwise
convolution and a softmax layer at the final output
layer.
5.1.2 CNN model
Convolutional neural networks model is very useful for
a large dataset [16]. Our proposed CNN consists of
convolution layers, Batch Normalization layers,
Maxpooling layer, Dropout layers, and a softmax layer
which is used as Output according to no. of classes. The
input layer consists of the size of the image.
Fig -6: Proposed architecture of Neural Network Model
There is 5 Convolution layers followed by a
batchNormalization and a maxpooling layer. After 2
Convolution layers, one dropout layer is used for
fastening the training process. The input layer consists of
the size of the image which is (176, 208, 3), As the dataset
consists of 4 classes.
Loss Functions:-
The categorical crossentropy loss function [17] computes
the following sum to determine the loss of an example:
Loss= ∑ ̂
̂ is the goal value that corresponds to the i-th scalar
value in the model output, output size is the total number
of scalar values in the model output, and so on.
How easily two discrete probability distributions may be
distinguished from one another is extremely well
measured by this loss. In this situation, the likelihood that
event i happens is denoted by ̂, and the total of all ̂ is
1, indicating that precisely one event might happen.
The negative sign makes sure that the loss decreases as
the distributions approach one another.
5.2 Machine Learning Model
5.2.1 Gaussian Naive Bayes
Using the Bayes theorem, the Naive Bayes classification
method was created [18]. When applying supervised
learning approaches, it is a straightforward but efficient
method for predictive modeling. The Naive Bayes
approach is simple to grasp. For incomplete or
unbalanced datasets, it offers better outcomes. The
machine learning classifier NaiveBayes uses the Bayes
Theorem. Given P(C), P(X), and P(X|C), one may apply the
Bayes theorem to calculate the posterior probability of
P(C|X).

P(X|C).
Therefore,
P(C|X) = (P(X|C) P(C) /P(X) 12]
P (C|X) = posterior probability of target class
P (X|C) = probability of predictor class
P(C) = probability of class C ( which is being true)
P(X) = prior probability of predictor class
5.2.2 Decision Tree Classifier:
A classification-focused supervised machine learning
algorithm is a decision tree classifier. Nodes and
internodes are used for classification. Instances are
categorized by root nodes according to their properties.
Additionally, these nodes represent classification while
these leaf nodes are made up of two or more branches.
[19] Using the most data acquired across all criteria, the
decision tree selects each node at each level.
5.2.3 Logistic Regression
It is a supervised learning method that utilises a
predetermined set of independent factors for
categorical dependent variables [20]. It explains the
relationship between independent and dependent
variables and is utilised for predictive analysis.
Classifying an input into groups is the outcome of
minimising the cost function. The cost function can be
written as:
= ∑ log +(1- )log(1- )]
Where
(x)=
5.2.4 Random Forest
During training, random forests (RF) build several
distinct decision trees. The average prediction for
regression or the median of the classes for classification
is created by combining the predictions from all
trees[21]. They are referred to as ensemble approaches
since they combine results into a final judgement.
6. RESULTS AND ANALYSIS
6.1 Precision
Precision is the ratio of correctly predicted
observations to all expected positive observations in
terms of positive observations.
Precision = TP/TP+FP
6.2 Recall
Recall is the percentage of accurately anticipated positive
observations to all of the actual class observations. The
formula for the following is TP/TP+FN
6.3 Accuracy
The easiest performance metric to understand is
accuracy, which is just the proportion of properly
predicted observations to all observations. The formula
for the following is
Accuracy =
6.4 F1 Score
The weighted average of Precision and Recall is the F1
Score. Therefore, both false positives and false negatives
are included while calculating this score. Although it is
true that F1 is often more advantageous than accuracy,
especially if you have an uneven class distribution, it is
not as intuitively easy to grasp as accuracy. When false
positives and false negatives cost about the same,
accuracy performs best. If there is a significant difference
in the costs of false positives and false negatives, it is
preferable to include both Precision and Recall.
F1 Score =
7. Evaluation Matrix
Model Accuracy Precision Recall F1 Score
CNN
Model
99.47 98.22 99.01 98.61
Mobilenet 92.24 91.11 90.89 90.50
VGG - 16 93.24 89.33 87.22 88.32
Logistic
Regressio
n
79.00 81.00 81.00 81.00
Gaussian
NB
52.00 57.00 55.00 55.00
Decision
Tree
58.00 67.00 67.01 67.00
Random
Forest
64.00 69.00 64.00 59.00
Table- 1: Evaluation matrix

The dataset consists of 4 classes (Mention the four
classes). Our proposed model has achieved remarkable
accuracy on test data which is 99.47%.
7.1 Confusion Matrix
A confusion matrix of dimension n x n connected to a
classifier, where n is the number of distinct classes, the
predicted and actual classification are displayed[22].
The elements of a confusion matrix include the
percentages of accurate negative forecasts, wrong
positive predictions, incorrect negative predictions, and
correct positive predictions are as follows: a, b, c, and d.
This matrix may be used to determine the prediction
accuracy and classification error as follows:
Accuracy =
error =
Confusion matrix for our proposed methodology is
Fig -7: Confusion Matrix of proposed model
Chart -1: Model accuracy vs epochs (Proposed Model)
Chart -2: Model loss vs epochs (Proposed model)
Chart -3: Model comparison
8. CONCLUSION AND FUTURE REFERENCES
Since there is currently no known treatment for
Alzheimer's, it is more crucial to lowering risk, give early
intervention, and precisely evaluate symptoms. As can be
seen from the literature review, numerous efforts have

been made to identify Alzheimer's Disease using
various machine learning algorithms and micro-
simulation techniques; however, it is still difficult to
identify pertinent characteristics that can Kavitha et al.
Early-Stage Alzheimer's Disease Prediction detect
Alzheimer's very early. In order to increase the
accuracy of detection approaches, future studies will
concentrate on the extraction and analysis of novel
features that are more likely to help in the identification
of Alzheimer's disease as well as on the removal of
redundant and unnecessary characteristics from
current features sets.
Using more precise data with features of age, gender
and previous record of the disease may boost the
accuracy to a greater extent in practical real-time
scenarios. More models can be trained to segment the
affected part is also useful.
REFERENCES
[1] Bhushan I, Kour M, Kour G, et al. Alzheimer’s
disease: Causes and treatment – A review. Ann
Biotechnol. 2018; 1(1): 1002.
[2] S. Padmanaban, K. Thiruvenkadam, P. T., M.
Thirumalaiselvi, και R. Kumar, ‘A Role of Medical
Imaging Techniques in Human Brain Tumour
Treatment’, τ. 8, pp. 565–568, 01 2020.
[3] A. Hosny, C. Parmar, J. Quackenbush, L. H. Schwartz,
και H. J. W. L. Aerts, ‘Artificial intelligence in radiology’,
Nat Rev Cancer, τ. 18, τχ. 8, pp. 500–510, Αυγούστου
2018.
[4] C. Kalyani, R. Kama, και G. Reddy, ‘A review on
optimised K-means and FCM clustering techniques for
biomedical image segmentation using level set
formulation’, Biomedical Research, τ. 29, 01 2018.
[5] D. Jha και G.-R. Kwon, ‘Alzheimer’s Disease Detection
Using Sparse Autoencoder, Scale Conjugate Gradient
and Softmax Output Layer with Fine Tuning’, τ. 7, pp.
13–17, 02 2017.
[6] P. Lodha, A. Talele, και K. Degaonkar, ‘Diagnosis of
Alzheimer’s Disease Using Machine Learning’, 08 2018,
pp. 1–4.
[7] R. Ammar και Y. Benayed, ‘Speech Processing for
Early Alzheimer Disease Diagnosis: Machine Learning
Based Approach’, 10 2018, pp. 1–8.
[8] D. R. Sarvamangala και R. V. Kulkarni, ‘Convolutional
neural networks in medical image understanding: a
survey’, Evolutionary Intelligence, τ. 15, τχ. 1, pp. 1–22,
Μαρτίου 2022.
[9]R. Rawat, M. Akram, Mithil, και S. Pradeep, Dementia
Detection Using Machine Learning by Stacking Models.
2020.
[10] K. Dashtipour κ.ά., ‘Detecting Alzheimer’s Disease
Using Machine Learning Methods’, 2022, pp. 89–100.
[11] Shikalgar, Arifa & Sonavane, Shefali.κ.ά., Hybrid Deep
Learning Approach for Classifying Alzheimer Disease
Based on Multimodal 2020
[12] F. Emmert-Streib, Z. Yang, H. Feng, S. Tripathi, και M.
Dehmer, ‘An Introductory Review of Deep Learning for
Prediction Models With Big Data’, Frontiers in Artificial
Intelligence, τ. 3, 2020.
[13]D. G. Murray, J. Šimša, A. Klimovic, και I. Indyk,
‘Tf.Data: A Machine Learning Data Processing
Framework’, Proc. VLDB Endow., τ. 14, τχ. 12, pp. 2945–
2958, Ιουλίου 2021.
[14] C. Szegedy κ.ά., ‘Going deeper with convolutions’, 06
2015, pp. 1–9.
[15] A. G. Howard et al., “MobileNets: Efficient
Convolutional Neural Networks for Mobile Vision
Applications.” 2017.
[16] R. Yamashita, M. Nishio, R. K. G. Do, και K. Togashi,
‘Convolutional neural networks: an overview and
application in radiology’, Insights into Imaging, τ. 9, τχ. 4,
pp. 611–629, Αυγούστου 2018.
[17] Rusiecki, Andrzej. ‘Trimmed categorical cross-
entropy for deep learning with label noise. Electronics
Letters’. 55. 10.1049/el.2018.7980. 2019
[18] Y. Huang και L. Li, ‘Naive Bayes classification
algorithm based on small sample set’, στο 2011 IEEE
International Conference on Cloud Computing and
Intelligence Systems, 2011, pp. 34–39.
[19] A. Navada, A. N. Ansari, S. Patil and B. A. Sonkamble,
"Overview of use of decision tree algorithms in machine
learning," 2011 IEEE Control and System Graduate
Research Colloquium, 2011, pp. 37-42, doi:
10.1109/ICSGRC.2011.5991826.
[20] J. J. DeStefano, "Logistic regression and the
Boltzmann machine," 1990 IJCNN International Joint
Conference on Neural Networks, 1990, pp. 199-204 vol.3,
doi: 10.1109/IJCNN.1990.137845.
[21] L. Monno, R. Bellotti, P. Calvini, R. Monge, G. B.
Frisoni and M. Pievani, "Hippocampal segmentation by
Random Forest classification," 2011 IEEE International
Symposium on Medical Measurements and Applications,
2011, pp. 536-539, doi: 10.1109/MeMeA.2011.5966763.

[22] T. C. W. Landgrebe and R. P. W. Duin, "Efficient
Multiclass ROC Approximation by Decomposition via
Confusion Matrix Perturbation Analysis," in IEEE
Transactions on Pattern Analysis and Machine
Intelligence, vol. 30, no. 5, pp. 810-822, May 2008, doi:
10.1109/TPAMI.2007.70740.

Early Detection of Alzheimer’s Disease Using Machine Learning Techniques

More Related Content

What's hot (20)

Similar to Early Detection of Alzheimer’s Disease Using Machine Learning Techniques (20)

More from IRJET Journal (20)

Recently uploaded (20)

Early Detection of Alzheimer’s Disease Using Machine Learning Techniques