Detection of Autism Spectrum Disorder (ASD) Symptoms using LSTM
Model
PRASENJIT MUKHERJEE1,2, MANISH GODSE3, BAISAKHI CHAKRABORTY4
1Department of Technology,
Vodafone Intelligent Solutions,
Pune,
INDIA
2 Department of Computer Science,
Manipur International University,
Manipur,
INDIA
3 Department of IT,
Bizamica Software,
Pune,
INDIA
4 Department of Computer Science and Engineering,
National Institute of Technology,
Durgapur,
INDIA
Abstract: - Autistic children will often exhibit certain behaviors that are unique to them and that are not typical
of neurotypical children. Parents will become familiar with these patterns over time and will be able to use this
knowledge to answer questions about their child's behavior. Deep learning models are very useful to solve
critical problems in the healthcare domain. Detection of ASD at the early age of a child is a challenging task.
Recent research reveals that there is an increasing trend of ASD among children. Communication, eye contact,
social behavior, and education are very poor for those who suffer from ASD. The proposed research work has
been done to detect ASD symptoms in a child. Data has been collected from the various autism groups from
social sites and organizations that are working on special children. A Deep learning model like the Long-Short
Term Memory (LSTM) model has been used to detect the sentiment of parents’ dialog. LSTM is the most
popular deep learning model that can able to solve complex natural language problems. The proposed LSTM
model has been trained with prepared data and accuracy is 97% according to the prepared data.
Key-Words: - LSTM, Deep Learning, Autism Detection, Machine Learning, ASD dataset, BERT Cosine.
Received: March 22, 2023. Revised: November 19, 2023. Accepted: December 22, 2023. Published: February 20, 2024.
1 Introduction
Autism spectrum disorder (ASD), which is
identified as an imbalance in brain functioning, can
be described as a neurodevelopmental problem, [1].
Individuals with ASD face difficulties in verbal and
non-verbal communication as well as social
interactions. Because of these complications, their
interpersonal skills and quality of life are greatly
hampered. In the recent past, the World Health
Organization, in 2019, released a report related to
the prevalence of ASD and it has been estimated
that 1 in 160 children is affected by this, [2]. Due to
the absence of objective interpretative mechanisms
for ASD, clinical diagnosis of ASD faces notable
challenges, [3]. The clinical diagnosis of ASD used
by the presented methodology depends primarily on
behavioral assessment, yet the accuracy of the
diagnosis is undermined by the considerable
heterogeneity of ASD and the diverse range of
clinical symptoms present, [4]. Depicted by
variations in symptoms, ASD can be categorized
into three distinct subtypes. These subtypes
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2024.21.5
Prasenjit Mukherjee,
Manish Godse, Baisakhi Chakraborty
E-ISSN: 2224-2902
40
Volume 21, 2024
comprise ASD, Asperger's disorder (APD), and
pervasive developmental disorder not otherwise
specified (PDD-NOS).A specific profile of
characteristics is represented by each sub-type that
allows a more refined understanding of the diverse
presentation of ASD within the stretch as in, [5].
Over the last two decades, due to the considerable
advancements in neuroscience research, certain
biomarkers have been identified which is a step in
the right direction to the understanding of neural
mechanisms associated with ASD. In this endeavor
an important role has been played by functional
brain imaging modalities, allowing the
characterization of these biomarkers as described in,
[6]. Parallel to this trend, the utilization of artificial
intelligence (AI) models within the domain of
medical diagnosis has a strong exponential, with a
particular priority on psychiatric disorders.The
prospect of the application of AI technology is
crucial in augmenting diagnostic capabilities,
including those associated with psychiatric
conditions. It is the predictive power of AI through
which medical professionals are assured to gain
precious insights and can enhance their expertise to
diagnose accurately and understand complex
psychiatric disorders, [7]. ASD, which is a complex
disorder, has many factors that can contribute to the
development of it. Moreover, there is no
multifunctional approach to managing the
symptoms. There are many problems to fulfill the
individual's needs who suffer from ASD. A right
combination of interventions is needed which are
child specific to address the issue to identify the
type of ASD and the corresponding treatment. The
process, however, is long and complicated.
Moreover, there is a lack of awareness about ASD
among the masses and little educated people are not
free to discuss the issue as they feel threatened by
social stigma and discrimination. There is a lack of
funding and resources for research related to ASD
treatments and therapies. As a result, sufficient
caregiving is limited and expensive as in, [8]. Social
skills, communication, ability to regulate emotions,
and behavior are affected by ASD. The signs and
symptoms of ASD can be identified by early
screening, and then appropriate interventions are
addressed and correspondingly, the relevant
treatments are applied. The quality of life for people
with ASD can be improved by the early
interventions that help them reach their full
potential. The ASD diagnosis requires a
comprehensive assessment of the child's behavior,
communication, and developmental milestones and
the process is complex. Additionally, families in
low-income communities often lack access to those
services because the cost of diagnosis and treatment
can be extensively expensive as in, [9]. Early
intervention strategies can help early diagnosis, the
severity of the symptoms thus can be reduced and
the overall functioning of the individual can be
improved. The associated healthcare costs can also
be reduced by early diagnosis because any potential
problems can be identified early which leads to
more effective and less expensive treatments. Since
interventions are most fruitful when delivered early
in life, early detection of ASD can help minimize
the long-term economic implications of the
condition. Machine learning algorithms enable to
identify patterns by analyzing the large amounts of
data in a short period, which can help diagnose the
disorder accurately. The time and resources needed
to address the disorder can be helped by this, and
the accuracy of the diagnosis process can be reduced
as in, [10]. Computers can analyze data quickly,
accurately, and cheaply with the intervention of AI
and ML. Detection of signs of autism earlier than
ever before and minimizing the cost and time
involved with traditional screening tests are helped
by the application of AI and ML as in, [11].
Enormous research has evolved in numerous areas
of healthcare system development, including the
field of security, [12] and the improvement of
healthcare quality and services, [13]. The
incorporation of artificial intelligence (AI)
technology benefited healthcare providers by
significantly empowering paramedics to support
necessary preliminary treatments ignoring the late
arrival of medical specialists, and thus it provides
more cost-effective healthcare delivery at an initial
stage as in, [14]. Several noteworthy studies have
contributed to the progress of medical diagnostics.
The authors in, [15], [16] and [17], have made
significant contributions in this area, conducting
valuable research in medical diagnostic techniques.
Their findings have helped advance the field,
enabling more accurate and efficient diagnoses,
ultimately leading to improved patient care and
outcomes.
The proposed research work has stated the
detection of ASD symptoms from parents’
dialogues. As the first step, the sentiment analysis
has been described using parents’ dialogues with
autistic children. Data has been collected from the
organizations of ASD children and social sites. A
parent of an autistic child spends maximum time
with their autistic child and they are well aware of
the ASD symptoms of their baby. The proposed
collected data consisted of parents’ experiences and
thoughts about their ASD children. The dataset has
been prepared from the collected data. Each
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2024.21.5
Prasenjit Mukherjee,
Manish Godse, Baisakhi Chakraborty
E-ISSN: 2224-2902
41
Volume 21, 2024
sentence is labeled as true (1) if the sentence
contains many words that indicate ASD symptoms.
If the sentence does not contain any words that are
related to ASD symptoms, then it will be marked as
false (0). The first approach is the LSTM model
which is the most popular type of RNN model in
NLP. The LSTM model has been trained using the
proposed dataset for sentiment prediction which is
related to ASD symptoms. Section 2 discusses
Related works on ASD, Section 3 elaborates on the
system architecture for ASD detection, Section 4
focuses on results, and Section 5 points out the
limitations. Section 6 deals with conclusion of the
proposed work and section 7 winds up the paper by
discussing possible future work on ASD detection.
2 Related Works
Autism Spectrum Disorder (ASD) is a notable
health issue among children, and thus healthcare
researchers have come forward to devise methods of
detection of this disorder. Artificial Intelligence (AI)
has come up as a significant tool for studying and
treating ASD, and various studies have employed
AI-based methods to make people aware of this
disorder. In addition, other mental health issues
have been addressed by AI, and various important
works in this field have been incorporated in this
section on related research. People with ASD face
an obstacle related to social interactions and
communication, and more often they have difficulty
in understanding and replying to social indications.
Sensory sensitivities are also experienced by them
which can lead them to respond to certain sounds or
textures in an exaggerated manner. This disorder has
also other behaviors such as repetitive actions and
limited interests. Providing the necessary support
and interventions to the children is very important
which can help them handle their social and
communicational difficulties. ASD detection and
treatment face a lack of resources and knowledge.
For example, many parents are unaware of the
developmental milestones that can specify ASD.,
Children with ASD are not accurately and easily
diagnosed because of lack of enough diagnostic
tools. ASD is a complex and multifactorial disorder
comprising a variety of environmental and genetic
factors. The study is difficult because each case may
have a unique set of symptoms. Additionally, many
of the present diagnostic tests are high-priced and
time-consuming. To identify correlations between
certain traits and the presence of ASD and to
analyze data from numerous sources, the paper
examines the extent to which supervised machine
learning algorithms may be applied for ASD
detection. To identify patterns by this analysis is the
goal of the paper to be used as markers for early
diagnosis of ASD. Traditional machine learning
algorithms are worthwhile in identifying features
that can differentiate between individuals with and
without ASD. However, deep learning architectures
have the prospect of improving the classification
process accurately by leveraging larger datasets and
more complex feature sets as in, [18]. Autism
appears from a complex interaction of genetic and
environmental influences that disturb brain
development. Challenges in social interaction,
communication, and repetitive behaviors are
typified by the condition. It has been investigated
that the genesis of this syndrome is because of
genetic predispositions, environmental factors, and
lifestyle choices. However, the specific cause
remains difficult to find, with the current consensus
suggesting a compound, miscellaneous nature of
ASD. Access to appropriate care is disrupted by the
scarcity of skilled professionals and resources for
diagnosing and dealing with ASD. Proper detection
and classification of ASD is a complex process
which means establishing an accurate biomarker for
accurate detection of ASD is difficult. Authors, [18],
have stated that traditional machine learning models
like decision trees and support vector machines are
capable of identifying ASD according to symptoms.
However deep learning models are more capable of
identifying ASD from high-dimensional data
accurately. According to accuracy and efficacy,
deep learning models are better for the detection of
ASD symptoms as in, [18]. Difficulty in
understanding and responding to verbal and
nonverbal indications, trouble in social interactions,
and difficulty in expressing one's thoughts and
feelings characterize ASD. Sensory issues create
difficulties among individuals with ASD. They are
facing difficulties in communication, eye contact,
emotions, and behavior. The symptoms of ASD may
vary from one to one who is suffering from ASD. In
childhood time, symptom identification is difficult
but after childhood, the symptoms of ASD can be
detected more prominently. Authors, [19], have
analyzed the first dataset that is related to ASD and
detected symptoms of ASD. The second dataset
contains 965 instances with 16 attributes whereas
the third dataset contains 1019 instances with 13
attributes. The Second dataset is related to the
emotions and the third dataset is related to motor
skills as in, [19]. Authors, [19], have used
Convolutional Neural Networks (CNNs) for better
accuracy and perfection. CNN [19] can handle high-
dimensional data with 99.53%, 98.30%, and 96.88%
accuracies according to datasets as in, [19]. The
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2024.21.5
Prasenjit Mukherjee,
Manish Godse, Baisakhi Chakraborty
E-ISSN: 2224-2902
42
Volume 21, 2024
work in, [20], helps to recognize an ASD class
where data is collected from various sources,
including teachers, parents, and medical
professionals. It then utilizes algorithms to analyze
the data and provide an output on ASD diagnosis.
The dataset includes various attributes like age,
gender, nationality, and class, collected from the
UCI ML repository and Kaggle, the dimension was
reduced to reduce the number of variables and
standardization was applied. With the minimized
dataset, the DNNPC model can then be used to
properly classify the ASD as in, [20]. The number
of ASD patients is increasing where researchers
opine that environmental, genetic, and neurological
factors may be contributing to the increment in
cases. Additionally, the syndrome of ASD can differ
greatly between individuals, making its diagnosis
complicated. There is often a lack of scientific
affirmation and validity in the tests for ASD, which
means that they may not always properly reflect the
condition of an individual. Furthermore, since ASD
is a disorder in the spectrum, it is complicated to
evaluate properly someone's exact level of
functioning. Automated diagnosis approaches are
faster and more precise than traditional methods.
This can help families to get the proper treatment
and support in the minimum possible time. It
reduces financial load and enhances the quality of
life for those having ASD. To analyze the data from
various modalities, such as audio, video, and text,
and to recognize patterns that can help to point out
the characteristics connected with ASD, the DANN
model was designed as in, [21]. It has succeeded in
differentiating various types of ASD, like low-
functioning or high-functioning autism. The ABIDE
repository has been used to benchmark the model
and assess its activities against methods of standard
machine-learning models. The outcomes show that
the DANN model was capable of properly classify
ASD patients with 0.732 accuracy, which was
notably higher than the outcomes of other models
that were tested. This exhibits the potency of the
model in consolidating various scales of brain
functional connectives and private characteristic
data for ASD categorization. By using more
validation methods, the model was capable of
exhibiting a high level of activeness on unseen data,
pointing out that the model had learned a
nonexclusive depiction of the data and was robust to
the outliers. This advocates that the model could be
useful in healthcare applications as in, [21]. ASD is
characterized by the struggle in communication and
social reciprocation, as well as confinement and
repetitive behavior. People with ASD can have
obstacles with everyday activities and tasks and may
require appropriate support to gain their full
potential abilities. These tests are often labor
intensive and require the existence of trained
specialist(s) who through several screening sessions
decide to detect ASD and its type or intensity. This
a time-consuming process. Authors of, [22], have
used six private characteristics; age, sex,
handedness, and three individual measures of IQ
which identify the Personal Characteristic Data
(PCD) of an individual to construct a novel
predictive model of ASD detection that suggested
that PCD can supply valuable insights into the
detection of ASD. To enhance the understanding of
the biological basis of autism and to advance better
diagnostic and treatment methods are the goals of
this project. The models were able to detect
differences in brain functioning between ASD and
non-ASD large datasets were used. The data was
divided into various subsets, then the model was
trained on one subset and then examined on another
subset. This eliminates data overfitting. Therefore,
this study results in a mean AUC (SD) of 0.646
(0.005), followed by a k-nearest neighbor with a
mean AUC (SD) of 0.641 (0.004) ensuring the
efficiency of clinical ASD detection. Such models
could allow earlier identification, more effective
intrusions, and better personalized treatments for
ASD people as in, [22]. This is because ASD is a
complicated neurological disorder that can manifest
in a variety of symptoms and the syndrome may be
subtle and tough to distinguish from other mental
health issues. Additionally, diagnosis can be
complex by the fact that the disorder is extremely
individual-specific. Machine learning technology
analyzes large data sets that would otherwise be
more complicated to process and recognize patterns
that could be used to properly diagnose disorders of
mental health. ML could also be used to detect
crucial treatments for these disorders and enhance
better interventions. ASD-DiagNet utilizes deep
learning to recognize patterns in the fMRI data that
can be used to distinguish between those with ASD
and those without. The proper detection of ASD is
directly related to the symptoms that should be
identified correctly. Neurological and biological
factors are important parts of ASD symptoms.
Authors [23] have used the Auto-Encoder technique
to extract features from ASD related data. The
parameters of the machine learning model, [23]
have been optimized by the Single-layer perceptron
(SLP). These two techniques have been applied to
make the datasets in various shapes and sizes. The
machine learning models can be trained with good
accuracies and as a result, the machine learning
models can predict more accurately as in, [23]. This
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2024.21.5
Prasenjit Mukherjee,
Manish Godse, Baisakhi Chakraborty
E-ISSN: 2224-2902
43
Volume 21, 2024
is because the referred model uses a blend of
convolutional neural networks and transfer learning
techniques to acquire knowledge from a large
number of data samples and recognize patterns for
the detection of the disorder. Furthermore, the
model is maximized to run in a fraction of the time
in comparison to other methods, which makes it
much more practical and value-effective as in, [23].
To have better understanding of the differences in
hemodynamic fluctuations between ASD and
Typically developing (TD) children, a multilayer
artificial neural network has been used in [24]. To
notice the differences in hemodynamic reply to an
auditory oddball activity between the ASD and TD
groups, a study was conducted in this work to
identify any differences in neural activation patterns
between the two groups. To pull out features from
the raw data, (comma) CNNs are utilized, while to
capture the temporal dependencies between the
features GRUs are applied. By combining the two,
CGRNN can properly recognize patterns in the data
and then classify them successfully. This approach
enabled the authors of, [24], to get appropriate
needful features from the data and capture the
temporal dynamic characteristics of brain actuation.
As a large set of data gets trained, the probability of
over-fitting is reduced. The use of DL networks
results in a more accurate classification of outcomes
than traditional methods which depend on a single
layer of neurons. It achieves 85.0% sensitivity,
92.2% accuracy, and 99.4% specificity. The
multilayer neural network CGRNN can recognize
features that relate to ASD, even in a short period.
Supervised learning models can produce better
results in certain cases of ASD. The range of
accuracies of these supervised learning models may
be between 0.78 to 0.86 where the F1 score is
between 0.72 and 0.84. The authors, [25], have
stated that NBSVM, [25] is the best model
according to 10 train-test cycles. The support vector
machine can select nonlinear relationships between
the input variables and the output labels where
NBSVM, [25], will not be considered for
recognizing complex patterns. Authors, [25], have
stated that a deep learning model may be a good
choice for complex pattern recognition. Researchers
were capable of classifying brain activation patterns
of patients with ASD by using deep learning
algorithms, and then those patterns are used to
identify ASD patients in large datasets. It is possible
to detect distinct patterns in the brain that may be
connected to ASD by comparing the brain imaging
data of ASD patients with the control patients.
Paper, [26], discusses and elaborates on the
possibility of deep learning models being able to
accurately differentiate ASD from typically
developing (TD) controls, based on a comparison of
functional MRI (fMRI) brain scans. The model
identified certain regions of the brain that
contributed most to the differentiation, which is
presented in the results mentioned in, [26]. Image is
a crucial source in ASD detection where Content-
Based Image Retrieval (CBIR) can play a vital role
in detecting ASD. Authors of, [27], use CBIR to
extract color, shape, texture, and spatial layout from
an image for index representation which will be
helpful for data preparation. Authors, [28], have
implemented their task using K-means clustering
techniques that can be applied on the numerical
variables. Such kind of unsupervised machine
learning models can be implemented in the
healthcare domain where data scientists are using
data for the improvement of the services in the
healthcare domain as in, [29].
Table 1 describes a comparative analysis
between proposed LSTM models with similar kinds
of machine learning models that can diagnose
mental disorders. The proposed table contains
“Models”, “Description”, “Dataset”, “Accuracy”,
and “Remarks”.
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2024.21.5
Prasenjit Mukherjee,
Manish Godse, Baisakhi Chakraborty
E-ISSN: 2224-2902
44
Volume 21, 2024
Table 1. Comparative Analysis between the proposed LSTM model and similar machine learning models to
detect Mental Disorders
Sl.No.
ML Models
Description
Accuracy
Remarks
Machine Learning Models Related to Mental Disorder (Autism Spectrum Disorder)
1
Logistic
Regression,
SVM, Naïve
Bayes, KNN,
[19]
These Traditional Models
can detect ASD from ASD
screening data.
96.69%,
98.11%,
96.22%, and
95.75%
The dataset contains 20
attributes and preprocessing
techniques are used like
removing null values and
normalization tasks.
2
ANN, CNN,
[19]
These advanced models
are used to detect ASD
from ASD screening data.
97.64% and
99.53%
The dataset contains 20
attributes and preprocessing
techniques are used like
removing null values and
normalization tasks.
3
Deep Neural
Network
Prediction and
Classification
(DNNPC), [20]
This model is a deep
learning-based classifier
that detects ASD among
children.
92%
This model is trained in two
phases. First, this model is
trained with missing data and
then it is trained with
complete data in the second
phase.
4
multichannel
DANN, [21]
This model is integrated
with multiple layers of
neural networks, attention
mechanisms, and feature
fusion to detect ASD
automatically.
73.2%
These experiments have been
done on the dataset. The k-
fold cross-validation and
leave-one-site-out cross-
validation have been designed
to complete the experiments.
5
KNN, SVM,
Decision Tree,
Logistic
Regression,
Random Forest,
and Neural
Networks, [22]
These models have been
used for ASD diagnosis
where ABIDE data
repositories have been
utilized for dataset
preparation.
61.8%,
54.7%,
59.1%,
57.2%,
and 62%
Some fixed data points have
been taken from ABIDE data
repository to train all the
models.
Proposed LSTM Model in Mental Disorder (Autism Spectrum Disorder)
6
Proposed LSTM
Model
The LSTM model has
been to predict positive
ASD symptoms from
parents’ dialogue. This
proposed model has been
trained with the proposed
dataset which is prepared
from the collected parents’
dialogues.
97%
The data has been
accumulated in textual format.
The discourse from parents
discussing their experiences
and perspectives regarding
their children with autism is
highly valuable. A parent with
an autistic child serves as an
optimal resource for grasping
the patterns of ASD
symptoms.
3 Proposed System Architecture for
ASD Detection
3.1 Dataset
The dataset has been curated by collecting dialogues
from parents who have shared their experiences and
thoughts concerning their autistic children. These
dialogues were gathered from various social
networks and organizations dedicated to the therapy
of special children, specifically focusing on
communication, behavior, and speech enhancement.
An example of some parent dialogues can be found
in Table 2. These parent dialogues serve as critical
data, providing a wealth of potential ASD symptoms
that can be identified. This information is utilized to
create a comprehensive dataset for training and
testing the proposed machine learning models.
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2024.21.5
Prasenjit Mukherjee,
Manish Godse, Baisakhi Chakraborty
E-ISSN: 2224-2902
45
Volume 21, 2024
Table 2. Example of Parents’ Dialogues
Sl. No.
Parents’ Dialogues
1.
Does anyone have advice on how they bring their
children out into busy places my son is 14 months
he’s being tested for ASD and ADHD but he
breaks down when we’re outside, especially near
traffic he will cry to the point he is sick.
2.
Today I found out my son have Level 3 of autism.
I wasn’t surprised to hear he is autistic because he
shows all the signs . Me and his dad moved to AZ
from IL and I really haven’t had much luck
meeting some ppl who has kid(s) that are autistic
and some ppl don’t be understanding. Looking for
more mommy friends or family
3.
Help! Almost three year old is biting and picking
her nails. One of her fingers is bleeding and she
won't stop it. Screams at me if I try to distract or
anything. She's biting them and shouting and
crying that it hurts and then straight back into her
mouth. Nothing is working she only started this
yesterday. Please help
4.
Hello, I am new in this group and I have a query
about speech therapist in Oldham, England. My
baby is just 2 years and he is not able to speak
properly. Please help me.
5.
My 3 yo is toilet trained (thankfully), however we
have noticed when he is in a socially awkward
situation he has been have little accidents and
today full release of his bladder. Is this common
in Autism/Sensory Processing Disorder? Thank
you.
The dataset has been compiled from the text
provided in Table 2. Every sentence has been
examined to determine whether it indicates a
symptom of ASD. As ASD does not exhibit a fixed
set of symptoms for identification, increasing the
number of dialogues from parents who have
children with autism could lead to a broader range
of symptom identification. Additionally, it offers the
advantage of providing a more robust training
dataset for machine learning models, potentially
enhancing their accuracy. A selection of data from
the suggested dataset is presented in Table 3.
Table 3. Example data in the proposed dataset
Sl. No.
Comments
Sentiment
1.
At 3 he had 0 words and now he
consistently says about 50-75
words.
1
2.
So my son turned 18 in January
& at the end of the school yr.
0
3.
I don’t think I can handle him
because he is very aggressive.
1
4.
when I call him not much eye
contact and also he’s not
talking.
1
5.
I have a doctor appointment
coming up next week.
0
The structure of the dataset used in the proposed
study is outlined in Table 3, where the first column
is the Serial Number, the second column contains
Comments, and the third column shows Sentiment.
The dataset was developed using text from the
dialogues of parents. Each sentence from these
parent dialogues was considered and analyzed to
determine if it signifies a symptom of ASD. If it was
a symptom, it was labeled as 1 (true), and if not, it
was labeled as 0 (false). As per Table 3, the
comments with serial numbers 1,3, and 4 represent
true ASD symptoms, while those with serial
numbers 2 and 5 are not indicative of ASD
symptoms. This ASD symptom-focused dataset has
now been prepared for LSTM training.
Table 4. List of Labels with ASD Problems
Sl. No.
Label
ASD Problems
1.
1
Problem of Speech
2.
2
Problem of Sensory
3.
3
Problem of Behaviour
4.
4
Problem of Special Education
5.
5
Problem of Social Interaction
6.
6
Problem of Eye Contact
7.
7
Problem of Cognitive Behaviour
8.
8
Problem of Hyper Active
9.
9
Problem of Child Psychological
10.
10
Problem of Attention
Table 4 illustrates the relationship between
various ASD problems and their corresponding
labels. Specifically, Label 1 is associated with the
"Speech Problem," while Label 2 and Label 3
correspond to the "Sensory" and "Behavior"
problems, respectively. Additional ASD problems
are also represented by labels listed in Table 4. After
the prediction of sentiment, which is related to ASD
symptoms, the proposed system will use Table 4 to
identify the problem according to the labels which
are associated. The cosine similarity model will use
each positive sentence to compare with each
sentence inside the dataset that has been given in
Table 5. Table 5 contains multiple positive
sentences with their respective labels. Each label in
Table 5 signifies an ASD problem based on the
associations outlined in Table 4. The cosine
similarity model, as explained in the Proposed
System Flow section, performs a similarity check
between the predicted positive sentences and the
dataset sentences to identify the sentence with the
highest similarity. The system selects the label
associated with this highly similar sentence.
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2024.21.5
Prasenjit Mukherjee,
Manish Godse, Baisakhi Chakraborty
E-ISSN: 2224-2902
46
Volume 21, 2024
Table 5. Dataset for Cosine Similarity check
Sl. No.
Positive Sentences
Label
1.
Hello, I need some advice on potty
training for my 4 years old girl. She
is non responsive.
7
2.
My 7y old just decided she hates
school and does not want to go.
4
3.
My baby is a girl and she is just 3
years old. But I am not hearing any
sound from her. Maybe she is
nonverbal. Please help me.
1
4.
he's on the move always and always
into something.
8
5.
He can concentrate on TV but
cannot even have a little attention.
10
The LSTM model has been described with the
proposed algorithm in the next sections where this
dataset has been utilized to train this model.
3.2 The LSTM Model
The Long Short-Term Memory (LSTM) is an
advanced model of Recurrent Neural Network
(RNN) which is a deep learning model. The output
of the previous step is an input of the current step in
RNN where RNN is not able to store data for
prediction in a long-term basis. The prediction result
is more accurate on current data on RNN. This is the
main disadvantage of the simple RNN model. This
problem has been solved by the LSTM model which
is itself a RNN type model. The LSTM can store
data on a long-term basis. LSTM has been widely
used in classification problems due to its feedback
connectivity. The LSTM can handle not only single
data points, but it can also handle complete data
streams. According to Figure 1, a neural network
and multiple memory blocks are the main structure
of the LSTM model. Four units are there in the
LSTM model where the cell is the first unit, the
input Gate and output gate are as second and third
units, and the last one is the forget gate. The flow of
information process inside a cell is managed by the
three gates. The main work of this cell is to
remember values in long time intervals. The cell
inside the LSTM model stores information and it
acts as a memory where other gates manipulate
memory. The input gate takes responsibility to use
input value for changing the memory and here
sigmoid function is used to allow either 0 or 1. The
tanh function is used to assign weights to the data
for determining their importance according to the
score of -1 to 1. The mathematical equation of the
input gate in the n LSTM model has been given
below:
Inputt = Sigmoid ( Winput.[ht – 1, Xt] + binput)
Tc = tanh(WT.[ht -1, Xt] + bT)
Fig. 1: LSTM Architectural Diagram
The forget gate is used to remove information
from the cell using the sigmoid function. For each
number in cell state Tc -1, This gate looks into the
preceding state (ht-1) where the input is Xt and
generates a number between 0 and 1. The blocks
input and cell are used to identify the output where
the sigmoid function is used to allow 0 or 1 and tanh
function is used for the determination of values (0
and 1). After this, tanh function is used to assign
weight to the provided values on a scale of -1 to 1
and multiply it with a sigmoid value. The
mathematical equation of this gate has been given
below:
outputt = sigmoid ( Woutput [ht -1, Xt] + boutput)
ht = outputt * tanh (Tc)
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2024.21.5
Prasenjit Mukherjee,
Manish Godse, Baisakhi Chakraborty
E-ISSN: 2224-2902
47
Volume 21, 2024
Fig. 2: Architectural Diagram of Proposed System
3.3 The Proposed ASD Detection System
3.3.1 Sentiment Analysis using LSTM Model
According to Figure 2, the proposed system will
read the dataset. Then each word from the sentences
of the dataset will be transformed into lowercase
words if a word contains any uppercase alphabet.
Now special characters have been removed from
each sentence. After completion of this step, the
features have been generated using the tokenizer in
a sequence as input with labeled output data to the
proposed LSTM model. Before inputting prepared
data into the model, the prepared data is split into
the training data and testing data. The LSTM model
will be created, and training data will be sent to train
the model. After completion of this step, the LSTM
model can understand the pattern for the sentiment
analysis of Parents’ dialogues. In the final step, the
trained model will be tested using the testing data.
After completing the testing phase, the proposed
LSTM model is ready to predict the sentiment as 0
or 1 on new data. Any parent can send their
experience of their child as a text into the proposed
system as input, the positive or negative sentences
will be detected according to the ASD symptoms.
Pseudo Code:
Step 1: Read the Dataset in variable data.
Step 2: Pre-processing task on data
// data[‘Comments’] stores sentences and
data[‘Sentiment’] stores binary value 0 and 1.
data = data[['Comments','Sentiment']]
// At first covert all the sentences in lower case and
remove special characters from the sentences.
data['Comments'] =
data['Comments'].apply(lambda x: x.lower())
data['Comments'] =
data['Comments'].apply((lambda x: re.sub('[^a-zA-
z0-9\s]','',x)))
Step 3: Feature Generation from text.
// Maximum feature generation has been initialize
max_fatures = 2000
// Tokenize each sentences from the
data[‘Comments’] column
tokeniz = Tokenizer(num_words=max_fatures,
split=' ')
tokeniz.fit_on_texts(data['Comments'].values)
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2024.21.5
Prasenjit Mukherjee,
Manish Godse, Baisakhi Chakraborty
E-ISSN: 2224-2902
48
Volume 21, 2024
// Sequence generation for training and testing data
X =
tokeniz.texts_to_sequences(data['Comments'].values
)
X = pad_sequences(X)
Step 4: LSTM Deep Learning Model creation
embedim = 128
lstmout = 196
modl = Sequential()
modl.add(Embedding(max_fatures,
embedim,input_length = X.shape[1]))
modl.add(SpatialDropout1D(0.4))
modl.add(LSTM(lstm_out, dropout=0.2,
recurrent_dropout=0.2))
modl.add(Dense(2,activation='softmax'))
modl.compile(loss = 'categorical_crossentropy',
optimizer='adam',metrics = ['accuracy'])
Step 5: Training and Testing Data Creation
batch_size = 32
modl.fit(X_train, Y_train, epochs = 7,
batchsize=batch_size, verbose = 2)
valsize = 50
X_valid = X_test[-valsize:]
Y_valid = Y_test[-valsize:]
X_test = X_test[:-valsize]
Y_test = Y_test[:-valsize]
Step 6: Model Evaluation
scores,accuracy = modl.evaluate(X_test, Y_test,
verbose = 2, batch_size = batchsize)
print("Scores= %.2f" % (scores))
print("Accuracy=: %.2f" % (accuracy))
Step 7: Prediction using proposed LSTM model on
validation data
Pos_Cnt, Neg_Cnt, Pos_Correct, Neg_Correct = 0,
0, 0, 0
for x in range(len(X_validate)):
result =
modl.predict(X_validate[x].reshape(1,X_test.shape[
1]),batch_size=1,verbose = 2)[0]
rst1.append(np.argmax(Y_validate[x]))
rst2.append(np.argmax(result))
if np.argmax(result) == np.argmax(Y_valid[x]):
if np.argmax(Y_valid[x]) == 0:
Neg_Correct = Neg_Correct +1
else:
Pos_Correct = Pos_Correct + 1
if np.argmax(Y_valid[x]) == 0:
Neg_Cnt = Neg_Cnt +1
else:
Pos_Cnt = Pos_Cnt +1
3.3.2 The BERT Cosine Similarity Model
The proposed system adopts a two-step approach for
ASD symptom identification. Initially, it filters out
negative sentences from the input text, focusing
solely on positive sentences. These positive
sentences are then utilized as input for the BERT
Cosine Similarity Model. In the second step, the
BERT Cosine Similarity Model processes each
positive sentence from the ASD symptoms dataset
(Table 5) and computes the cosine similarity score
between the input sentence and each sentence in the
dataset. By comparing the input sentence with the
dataset sentences, the model identifies the sentence
with the highest cosine similarity score. The system
subsequently selects the label associated with this
highly similar sentence. According to Table 4,
which establishes the correspondence between ASD
problems and labels, this label indicates a specific
ASD problem. The cosine similarity model
independently applies this process to each input
sentence, allowing the system to identify ASD
problems based on the highest similarity scores and
their corresponding labels. The algorithm
summarized below outlines the steps involved:
1. A positive sentence will be selected from the
input text.
2. In the next step, the BERT cosine similarity
model will be applied.
3. The cosine similarity will be calculated between
the input sentence and each positive sentence
from the dataset of ASD symptoms.
4. In this step, identify the sentence according to
the highest cosine similarity score. After
identifying the positive sentence, retrieve the
corresponding label from Table 4 which is
associated with the ASD problem.
By leveraging the BERT Cosine Similarity Model
and utilizing the labels from Table 4, the proposed
system effectively identifies ASD problems by
analyzing the similarity between input sentences and
the ASD symptoms dataset.
The algorithm in Python-pseudo code has been
given below.
The Algorithm in Python pseudo-code:
from sentence_transformers import
SentenceTransformer, util
import pandas as pd
import pandasql as ps
// Dataframe df to be initialized by the dataset
df =
pd.read_csv(r"ASD_Symptoms.csv",encoding='Lati
n-1')
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2024.21.5
Prasenjit Mukherjee,
Manish Godse, Baisakhi Chakraborty
E-ISSN: 2224-2902
49
Volume 21, 2024
// List Array declare to store ‘Comments’,
‘Sentiment value’, and ‘Cosine Score value
Comments1=[]
Sentiment1=[]
cosine_value1=[]
// Function for bracket remove from Cosine value in
Python
def StringWithoutBrackets(lists):
return str(lists).replace('[','').replace(']','')
// Calculation with BERT Cosine function
def BERTCosine(strs1):
for indx in df.index:
#print(df['Comments'][indx],
df['Sentiment'][indx])
sentence = [df['Comments'][indx], strs1]
modl = SentenceTransformer('sentence-
transformers/all-MiniLM-L6-v2')
# Embeddings to be computed for both lists
embeddings1= modl.encode(sentence[0],
convert_to_tensor=True)
embeddings2 = model.encode(sentence[1],
convert_to_tensor=True)
comments1.append(df['Comments'][indx])
sentiment1.append(df['Sentiment'][indx])
scored=util.pytorch_cos_sim(embeddings1,
embeddings2)
cosine_value.append(StringWithoutBrackets(scored.
tolist()))
dfd=pd.DataFrame(
{'Comment': comments,
'Sentiment': sentiment,
'CosineScores': cosine_value
})
//Dataframe to CSV which contains Cosine Scores
of sentences with corresponding label values.
dfd.to_csv('ASD_Cosine_Data1.csv”)
dfd['CosineScores']=dfc['CosineScores'].astype('flo
at64')
i = dfd['CosineScores'].idxmax()
return dfd['Sentiment'][i]
strs1=pd.read_csv("ASD_Cosine_Data1.csv")
for sts in strs1['Comment']:
rslt=BERT_Cosine(sts)
print(sts,"=",rslt)
The result of this proposed algorithm has been
discussed in the Result and Discussion section.
4 Results and Discussion
4.1 Result and Discussion of Proposed LSTM
Model
The proposed LSTM model can handle sentiment
analysis on the new data. Parents who are unaware
of the ASD symptoms of their child will get the
benefit of detecting ASD symptoms as early as
possible without going through a long process of
ASD detection. The evaluation results of this
proposed LSTM model have been discussed here
one by one.
Fig. 3: Classification Report of Proposed LSTM
Model
Many matrices are there to evaluate the performance
of a machine learning model where F1 score is one
of them. The precision and recall values are used to
calculate the score of F1. F1 score indicates a
positive prediction ability of the machine learning
model. The equation is given below:
F1 = (2 x (Precision x Recall)) / (Precision + Recall)
Precision can be calculated by the true positive (TP)
and false positive (FP) whereas Recall can be
calculated by the true positive (TP) and false
negative. The equations are given below:
Precision = TP ÷ (TP + FP)
Recall = TP ÷ (TP + FN)
The average F1 score for multiple classes can be
calculated by two methods. Macro-Averaging
method and weighted method are used to calculate
average F1 scores. On the other hand, weighted
averaging takes the relative proportions of each
class in the dataset, where the weight of each class
is governed by its support, which is the number of
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2024.21.5
Prasenjit Mukherjee,
Manish Godse, Baisakhi Chakraborty
E-ISSN: 2224-2902
50
Volume 21, 2024
true occurrences of the class in the dataset. Scores
for those classes are indeed lower if the support
values of these certain classes are very imbalanced.
In such cases, it may be necessary to adjust the
model or the dataset to address the imbalance and
improve the overall performance.
The classification report of this model has been
given in Figure 3 where the precision and recall
value of class 0 and 1 is 0.96, 0.97, 097, and 0.97.
The F1 score is 0.96 and 0.97 of class 0 and 1 with
support values 1738 and 2001. The accuracy of the
proposed LSTM model according to the F1 score is
0.97 with a support value of 3739 which indicates a
strong predictive model. The precision and recall
values of the macro average and weighted average
are 0.97, 0.97, 0.97, and 0.97. The F1 score of the
macro average and weighted average are 0.97 and
0.97 with support values 3739 and 3739.
Fig. 4: Confusion Matrix of Proposed LSTM Model
According to Figure 4, The Y-axis has the
Actual 0s and Actual 1s whereas the X-axis has
predicted 0s and predicted 1s. 50 sentences have
been used to represent this confusion matrix using
the proposed LSTM model. 23 sentences have been
predicted by the proposed LSTM as 0 where these
are actual 0s. No sentences have been predicted as 0
but they are 1. The proposed LSTM model has
predicted 26 sentences as 1 which is 1 and 1
sentence has been predicted as 1 which is 0.
According to the confusion matrix, it is a clear
observation that the proposed LSTM can detect
sentences as ASD symptoms from the parent’s
dialogue.
Fig. 5: Loss Visualization of Proposed LSTM
Model
According to Figure 5, two loss learning curves
have been shown which are blue lines and yellow
lines. The Y-axis is a Loss against each epoch as the
X-axis. 10 epochs have been considered here where
it has been seen that each line goes downwards after
epochs. The blue line refers to the loss during the
training using training data whereas the yellow line
refers to the loss during testing using test data. Both
lines indicate that losses (error) are reduced after a
certain time interval means completing epochs.
Reduction of loss or error means the model is ready
with good accuracy. The accuracy curves are plotted
in Figure 6. According to Figure 6, The Y-axis
refers to the accuracy whereas X-axis refers to the
epoch. The blue line is an accuracy line of the
proposed LSTM model during the training using the
training data. The yellow line is an accuracy line of
the proposed LSTM model during testing using test
data. Both lines are increasing towards 1.0.
Fig. 6: Accuracy Visualization of the Proposed
LSTM Model
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2024.21.5
Prasenjit Mukherjee,
Manish Godse, Baisakhi Chakraborty
E-ISSN: 2224-2902
51
Volume 21, 2024
4.2 Result and Discussion of BERT Cosine-
Similarity Model
Fig. 7: Example Output of BERT Cosine Similarity
Model
The proposed system employs a cosine
similarity model to identify ASD symptoms from
positive ASD sentences, predicted by machine
learning algorithms. The system's output, as
illustrated in Figure 7, assigns labels to sentences
such as "she will screams at the top of his lungs and
throws objects at others" (labeled 3), "But at school,
he's playing alone with toys" (labeled 5), and "she is
nonverbal and low functioning" (labeled 1).
Referring to Table 4, we can associate label 3 with
Behaviour problems, label 5 with Social Interaction
problems, and label 1 with Speech problems. These
labels provide valuable insights into the specific
ASD problems related to the identified symptoms.
Proper therapies can be initiated based on the
specific problem identified if the ASD problems are
detected. Tailored interventions play a crucial role
in delivering targeted support to people with ASD.
Accurate identification of ASD problems by the
proposed system through the analysis of positive
sentences allows for a focused and personalized
procedure for therapeutic interventions. This
personalized procedure promises a positive impact
on ASD patients and enhances their quality of life.
5 Limitation
The overall system performance degrades with large
datasets as the LSTM model may on work optimally
for large datasets. LSTM models have shown
restrictions in working with complicated
computations, like aggregation-type natural
language response generation. For example, models
like LSTM may fail to perfectly calculate the sum of
multiple float values concurrently.
6 Conclusion
To detect ASD symptoms, the proposed system is
designed to analyze natural language text extracted
from parent dialogues. The sentiment (positive or
negative) expressed in sentences related to ASD
symptoms is determined by sentiment analysis
techniques. The proposed system uses the LSTM
Model on the supplied dataset. Only the positive
sentences are chosen for additional analysis using
the BERT cosine similarity model after performing
sentiment analysis. An ASD symptoms dataset is
leveraged by the system, where each sentence is
tagged with a value corresponding to a particular
ASD symptom. Through the computation of cosine
similarity between the input sentence and the
sentences associated with the ASD symptoms
dataset, the system decides the label that
demonstrated the highest score signifying that the
input sentence has similarity to the sentence
associated with a specific ASD problem. The system
aims to bridge the gap in ASD diagnosis and
intervention by leveraging the ability of text-based
analysis, supplying valuable insights and assistance
to needy individuals and communities. Moreover
Large Language Model (LLM) can be launched to
uncover the symptoms from the parents’
conversations and such kind of LLM-based systems
are perfect for the generation of outcomes.
7 Future Work
Training the BERT and ChatGPT models using the
supplied dataset can be an important step for the
further development of the output and accuracy of
the proposed system. These models have proven to
be successful in several natural language processing
activities and can leverage the dataset to make
predictions, potentially enhancing the system's
performance. BERT, which is a deep learning
model, excels in classification activities and can
come up with more accurate ASD identification. On
the other side, ChatGPT, which is a large language
model, gives appropriate prediction abilities and can
supply important insights. Some hybrid processes
can be applied to handle aggregation-type reply
generation by using these models. In summary,
future work based on training the BERT and
ChatGPT models using the supplied dataset for
ASD identification holds assurance for developing
the system further. However, to achieve optimal
performance, a proper study needs to be done to
consider the dataset size and select the proper
models for our future work.
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2024.21.5
Prasenjit Mukherjee,
Manish Godse, Baisakhi Chakraborty
E-ISSN: 2224-2902
52
Volume 21, 2024
Acknowledgement:
The authors extend their appreciation to the
Manipur International University, Imphal, India for
supporting this post-doctoral research work on
Autism.
References:
[1] A.M. Pagnozzi, E. Conti, S. Calderoni, J.
Fripp, S.E. Rose, A systematic review of
structural MRI biomarkers in autism spectrum
disorder: A machine learning perspective, Int.
J. Dev. Neurosci. 2018, pp. 68–82.
[2] Autism Spectrum Disorders. World Health
Organization, [Online].
https://www.who.int/news-room/fact-
sheets/detail/autism-spectrum-disorders
(Accessed Date: February 7, 2021).
[3] S.N. Hansen, D.E. Schendel, E.T. Parner,
Explaining the increase in the prevalence of
autism spectrum disorders: The proportion
attributable to changes in reporting practices,
JAMA Pediatr., 2015, pp. 56-62.
[4] A.N. Witwer, L. Lecavalier, Examining the
validity of autism spectrum disorder subtypes,
J. Autism Dev. Disord., 2008, pp. 1611-1624.
[5] M.O. Mazurek, F. Lu, H. Symecko, E. Butter,
N.M. Bing, R.J. Hundley, M. Poulsen, S.M.
Kanne, E.A. Macklin, B.L. Handen, A
prospective study of the concordance of
DSM-IV and DSM-5 diagnostic criteria for
autism spectrum disorder, J. Autism
Dev.Disord., Vol. 47, 2017, pp.2783–2794.
[6] E. Conti, J. Mitra, S. Calderoni, K. Pannek, K.
Shen, A. Pagnozzi, S. Rose, S. Mazzoti, D.
Scelfo, M. Tosetti, F. Muratori, G. Cioni, A.
Guzzetta, Network over-connectivity
differentiates autism spectrum disorder from
other developmental disorders in toddlers: A
diffusion MRI study, Hum. Brain Mapp., vol.
38, 2017, pp. 2333–2344.
[7] F. Thabtah, Machine learning in autistic
spectrum disorder behavioral research: A
review and ways forward, Inform. Health Soc.
Care, vol. 44, 2019, pp. 278–297.
[8] G. Sandhu, A. Kilburg, A. Martin, C. Pande,
H. F. Witschel, E. Laurenzi, E. Billing, A
Learning Tracker using Digital Biomarkers
for Autistic Preschoolers– Practice Track,
EPiC Series in Computing, 2022, vol. 84, pp.
219-230.
[9] N V Ganapathi Raju, Karanam Madhavi, G
Sravan Kumar, G Vijendar Reddy,
Kunaparaju Latha, K Lakshmi Sushma,
Prognostication of Autism Spectrum Disorder
(ASD) using Supervised Machine Learning
Models, International Journal of Engineering
and Advanced Technology (IJEAT), vol. 8(4),
2019, pp. 1028-1032.
[10] S.R. Surya, Kalpana G., Autism Spectrum
Disorder Using KNN Algorithm, European
Journal of Molecular & Clinical Medicine,
vol. 07(09), 2020, pp. 1628-1637.
[11] G.K. Suhas, N. Naveen, M. Nagabanu, Edwin
R. Mario, R. Nithish Kumar, Premature
Identification of Autism Spectrum Disorder
using Machine Learning Techniques,
Advanced Innovations in Computer
Programming Languages, vol. 3(3), 2021, pp.
1-10.
[12] S. Purwanti, B. Nugraha, and M. Alaydrus,
Enhancing security on E-health private data
using SHA-512, International Conference on
Broadband Communication, Wireless Sensors
and Powering, BCWSP, vol. 2018-January,
2018, pp. 1–4.
[13] H. H. Purba, F. Debora, C. Jaqin, H.
Adiyatna, Service quality analysis: An
empirical study of customer satisfaction in
healthcare, Jurnal Teknologi dan Manajemen,
vol. 19, no. 1, 2021, pp. 33–38.
[14] Yaya Sudarya Triana, Mohd Azam Osman,
Adji Pratomo, Muhammad Fermi Pasha, Deris
Stiawan, Rahmat Budiarto, Neural network
models selection scheme for health mobile
app development, IAES International Journal
of Artificial Intelligence (IJ-AI), vol. 12(3),
pp. 2023, 1191-1203.
[15] S. Kaur, J. Singla, L. Nkenyereye, S. Jha, D.
Prashar, G. P. Joshi, S. El-Sappagh, M. S.
Islam, S. M. Riazul Islam, Medical diagnostic
systems using artificial intelligence (AI)
algorithms: Principles and perspectives, IEEE
Access, vol. 8, 2020, pp. 228049–228069.
[16] L. Jiang, Z. Wu, X. Xu, Y. Zhan, X. Jin, L.
Wang, Y. Qiu, Opportunities and challenges
of artificial intelligence in the medical field:
current application, emerging problems, and
problem-solving strategies, Journal of
International Medical Research, vol. 49, no.
3, 2021.
[17] W. Walter, C. Haferlach, N. Nadarajah, I.
Schmidts, C. Kühn, W. Kern, T. Haferlach,
How artificial intelligence might disrupt
diagnostics in hematology in the near future,
Oncogene, vol. 40, no. 25, 2021, pp. 4271–
4280.
[18] Shomona Gracia Jacoba, Majdi Mohammed
Bait Ali Sulaimanb, Bensujin Bennetc,
Algorithmic Approaches to Classify Autism
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2024.21.5
Prasenjit Mukherjee,
Manish Godse, Baisakhi Chakraborty
E-ISSN: 2224-2902
53
Volume 21, 2024
Spectrum Disorders: A Research Perspective,
The 5th International Conference on
Emerging Data and Industry 4.0, 2022, pp.
470-477.
[19] Suman Raja, Sarfaraz Masoodb, Analysis and
Detection of Autism Spectrum Disorder Using
Machine Learning Techniques, Conference on
Computational Intelligence and Data Science
(ICCIDS), 2020, pp. 994-1004.
[20] Ashima Sindhu Mohanty, Priyadarsan Parida,
Krishna Chandra Patra, ASD classification for
children using deep neural network, Global
Transitions Proceedings, 2021, pp.461-466.
[21] Ke Niu, Jiayang Guo, Yijie Pan, Xin Gao,
Xueping Peng, Ning Li, Hailong Li,
Multichannel Deep Attention Neural
Networks for the Classification of Autism
Spectrum Disorder Using Neuroimaging and
Personal Characteristic Data, Complexity, vol.
2020, pp. 1-9.
[22] Milan N. Parikh, Hailong Li, Lili He,
Enhancing Diagnosis of Autism With
Optimized Machine Learning Models and
Personal Characteristic Data, Frontiers in
Computational Neuroscience, vol. 13, article
9, 2019, pp. 1-5.
[23] Taban Eslami, Vahid Mirjalili, Alvis Fong,
Angela R. Laird, Fahad Saeed, ASD-DiagNet:
A Hybrid Learning Approach for Detection of
Autism Spectrum Disorder Using fMRI Data,
Frontiers in Neuroinformatics, vol. 13, article
70, 2019, pp. 1-11.
[24] Lingyu Xu1, Xiulin Geng, Xiaoyu He, Jun Li,
Jie Yu, Prediction in Autism by Deep
Learning Short-Time Spontaneous
Hemodynamic Fluctuations, Frontiers in
Neuroscience, vol. 13, article 1120, 2019, pp.
1-12.
[25] Scott H. Lee, Matthew J. Maenner, Charles
M. Heilig, A comparison of machine learning
algorithms for the surveillance of autism
spectrum disorder, PLOS ONE, 2019, pp. 1-
11,.
[26] A. S. Heinsfeld, A. R. Franco, R. C.
Craddock, A. Buchweitz, F. Meneguzzia,
Identification of Autism Spectrum Disorder
using Deep Learning and the ABIDE Dataset,
NeuroImage: Clinical, 2017, pp. 1-29.
[27] Vishal Padole, Image Classification by Using
Multiclass Support Vector Machines, WSEAS
Transactions on Computer Research, 2019,
pp. 1-8.
[28] Suboh Alkhushayni, Taeyoung Choi, Du’a
Alzaleq, Data Analysis Using Representation
Theory and Clustering Algorithms, WSEAS
Transactions on Computers, vol.19, 2020, pp.
310-320,
https://doi.org/10.37394/23205.2020.19.38.
[29] Meera Sharma, Sonok Mahapatra, Adeethyia
Shankar, Xiaodi Wang, Predicting the
Utilization of Mental Health Treatment with
Various Machine Learning Algorithms,
WSEAS Transactions on Computers, vol..19,
2020, pp. 281-295,
https://doi.org/10.37394/23205.2020.19.34.
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
The authors equally contributed to the present
research, at all stages from the formulation of the
problem to the final findings and solution.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
No funding was received for conducting this study.
Conflict of Interest
The authors have no conflicts of interest to declare.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2024.21.5
Prasenjit Mukherjee,
Manish Godse, Baisakhi Chakraborty
E-ISSN: 2224-2902
54
Volume 21, 2024