A Review of Machine Learning Models to Detect Autism Spectrum
Disorders (ASD)
PRASENJIT MUKHERJEEa, SOURAV SADHUKHANb, MANISH GODSEc
aDept. of Technology
Vodafone Intelligent Solutions
Pune
INDIA
aDept. of Computer Science
Manipur International University
Manipur
INDIA
bDept. of Business Management
Pune Institute of Business Management
Pune
INDIA
cDept. of IT
Bizamica Software
Pune
INDIA
Abstract: - Autism Spectrum Disorder (ASD) is a neurodevelopmental condition that can manifest in a variety
of ways. One common characteristic is difficulty with communication, which may manifest as difficulty
understanding others or expressing oneself effectively. Social interaction can also be challenging, as individuals
with ASD may struggle to comprehend social cues or adapt to new situations. Many machine-learning models
have been developed or are in progress to detect ASD automatically. Three machine learning model-based
frameworks have been studied and elaborated on, each with a clear concept of the detection of ASD among
children and adults. This research paper has done a closer review of these frameworks and their datasets to
diagnose ASD automatically. In the first framework, deep learning models such as Xception, VGG19, and
NASNetMobile have been utilized for the detection of autism spectrum disorder (ASD). In addition, other
models such as XGBoost, Neural Network, and Random Forest have been employed in the second framework
to detect ASD from a clinical standard screening dataset for toddlers. Meanwhile, the third framework involves
traditional machine learning models that have been trained using the UCI dataset for ASD. The accuracy of
each model has been discussed and elaborated on.
Key-Words: - Deep Learning, Autism Spectrum Disorder, Machine Learning, ASD Detection, ML-based
Framework, Traditional Machine Learning
Received: June 23, 2023. Revised: August 11, 2023. Accepted: September 14, 2023. Published: October 5, 2023.
1 Introduction
Autistic children often have difficulty understanding
and responding to social cues, so they may not know
how to start and maintain conversations.
Additionally, they may have difficulty
understanding abstract concepts and may be more
comfortable with concrete concepts. They may also
have trouble interpreting sensory information, such
as touch or sound, which can lead to sensory
overload. Finally, autistic children may be obsessed
with certain topics or routines due to their difficulty
processing changes in their environment. These
difficulties have been attributed to the lack of
reliable and valid screening instruments, the wide
range of severity of ASD symptoms, and the overlap
of symptoms with other disabilities. Additionally,
early intervention can be expensive and may not
always be available, depending on the particular
situation of the family, as in [1]. ASD is a
neurodevelopmental problem of the brain that has a
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.21
Prasenjit Mukherjee, Sourav Sadhukhan, Manish Godse
E-ISSN: 2224-2872
177
Volume 22, 2023
wide range of symptoms and severity [2]. ASD has
been included in the International Statistical
Classification of Diseases and Related Health
Problems (ISCDRHP) under the category of mental
and behavioral disorders, as in [3]. The symptoms
may appear in the first year of a toddler with less
eye contact and poor responses, as in [4] [5] [6] [7].
People with autism may experience difficulties in
social communication, such as difficulty
understanding body language, facial expressions,
and the meaning of words. They may also struggle
with sensory processing, such as being over or
under-sensitive to certain sounds, textures, lights,
and tastes, as in [8]. It is characterized by
difficulties with social interaction and
communication, as well as restricted, repetitive
behaviors. It is typically diagnosed in early
childhood and can last throughout a person's
lifespan. Symptoms of autism spectrum disorder are
usually noticeable before the age of three and can
range from difficulty communicating and interacting
with others to repetitive behaviors and
hypersensitivity to certain stimuli. These symptoms
can vary greatly in severity and type between
individuals. Machine learning algorithms can be
used to analyze patterns in the behavior of children
with autism and detect any abnormalities that might
indicate the presence of autism. This can help
clinicians diagnose the condition earlier and begin
treatment sooner, which can improve the outcome
for the child. Autism Spectrum Disorder (ASD) is a
neurodevelopmental condition that affects social
communication, behavior, and sensory processing.
Early identification and intervention are crucial to
improving outcomes and quality of life for
individuals with ASD. Some of the core features of
ASD include difficulties with social interaction,
communication, and repetitive behaviors or
interests. These challenges can make it difficult for
individuals with ASD to form and maintain
relationships, understand social cues, and participate
in everyday activities. Although the exact causes of
ASD are not fully understood, research suggests that
a combination of genetic and environmental factors
may contribute to its development. While the
condition is more prevalent in males, it is important
to note that ASD affects individuals of all genders,
races, and ethnicities, as in [9]. Diagnosing ASD
can be a complex process that involves a thorough
evaluation of a person's behavior, communication,
and developmental history. However, access to a
timely and accurate diagnosis can be limited,
particularly for families in low-income
communities. This can lead to delays in accessing
appropriate services and support. Advances in
technology, such as machine learning algorithms,
have the potential to improve the accuracy and
speed of ASD diagnosis. By analyzing large
datasets, these algorithms can identify patterns and
features that are characteristic of the condition,
which may assist clinicians in making more accurate
and efficient diagnoses. While these tools are not
intended to replace clinical judgment, they may help
supplement traditional assessment methods and
increase access to diagnostic services for individuals
and families affected by ASD, as in [10].
AI techniques can be used to analyze large amounts
of data from various sources, such as genetics,
medical records, and environmental factors. With
AI, patterns can be identified and used to develop
predictive models for ASD, which can help identify
individuals at risk for the disorder and provide early
interventions. The challenge arises because high-
dimensional data has a large number of features and
variables, which can make it difficult to identify
meaningful patterns in the data. Furthermore, the
sheer size of the data can make it difficult to process
and analyze. As a result, the analysis of high-
dimensional datasets requires specialized algorithms
that can accurately identify patterns in the data.
These algorithms must also be computationally
efficient enough to process large amounts of data in
a reasonable amount of time, as in [8].
The proposed research work is a review of AI
applications to detect autism spectrum disorder
among children and adults. Three frameworks that
contain the machine learning models have been
discussed with the dataset. The dataset plays an
important role because each model uses datasets to
train for predictions after getting new data. The
facial images of ASD-detected children and general
children have been taken as primary sources of data.
Deep learning models like Xception, VGG19, and
NASNetMobile have been applied to detect ASD, as
in [11]. The second framework uses models like
XGBoost, Neural Network, and Random Forest to
detect ASD from the clinical standard ASD
screening dataset of toddlers, as in [12]. The third
framework uses traditional machine learning models
that are trained with the UCI dataset of ASD, as in
[13]. These frameworks and some other similar
models have been elaborated on in Section III. The
entire study is given in Section III, where the results
of each model and observation have been discussed
in Section IV and the application of the proposed
study has been included in Section V. The proposed
study ends with a conclusion in Section VI.
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.21
Prasenjit Mukherjee, Sourav Sadhukhan, Manish Godse
E-ISSN: 2224-2872
178
Volume 22, 2023
2 Related Works
Autism Spectrum Disorder (ASD) is a significant
challenge for children's health today, and it has
become a key area of focus in healthcare research.
Many studies have explored the potential of
artificial intelligence (AI) to address this disorder
and other mental health-related issues. This section
highlights some notable AI-based research on
mental health that has been conducted.
By analyzing social media posts and biomedical
images, doctors can identify patterns in behavior
and physical symptoms that may be indicative of
ASD. This data can then be used to accurately
diagnose and treat the disorder. By recognizing
facial features associated with ASD, it is possible to
identify individuals with the disorder earlier in life.
This can lead to earlier diagnosis and intervention,
which can be beneficial for those affected by the
disorder. In addition, this system could also be used
to help identify individuals with ASD in social
media posts, which can help connect those with the
disorder with resources and support. Deep learning
techniques rely on accurately identifying key facial
features, such as eyes, nose, and mouth, and then
mapping those features to a template. This allows
the algorithm to recognize the face and identify the
landmarks associated with it. The exception model
achieved the highest accuracy result of 91%,
followed by VGG19 (80%) and NASNETMobile
(78%). The dataset used had a good variety of face
images of different backgrounds, angles, and
lighting conditions, which allowed the deep learning
models to accurately perceive and recognize
patterns and features of the faces. This enabled the
three models to detect a wide variety of faces, which
is why the exception model achieved the highest
accuracy result. The application is designed to
assess facial features from images of people's faces
and compare them to a database of images of people
with and without autism. The convolutional neural
network is trained to recognize the differences
between the two sets of images and categorize the
images accordingly. The Flask framework then
makes the application available online and allows
users to easily interact with the system, as in [11]. It
is also associated with difficulty processing sensory
information and difficulty with motor skills such as
handwriting or balancing. People with autism often
have difficulty understanding and responding to
social cues and may have difficulty forming
relationships with others. Because autism is a
spectrum disorder, it can manifest differently in
each individual. This means that the symptoms can
range from mild to severe, making it difficult to
distinguish between typical development and
autism. Furthermore, autism is often comorbid with
other mental health issues, which can make it even
more difficult to diagnose. Early screening and
treatment can help identify and address any
underlying health issues before they become severe.
This can help reduce the risk of long-term
symptoms as well as improve the overall quality of
life. The goal of this research is to develop an
automated pipeline that can quickly and accurately
identify the signs of autism in toddlers and to use
machine learning models to analyze the indicators
of autism and determine which are the most
significant for diagnosis. The dataset used for this
research was curated from the UC Irvine Autism
Spectrum Disorder dataset, which contains over
10,000 examples of autism-related features from
children aged 4-5. The neural network model was
designed to learn patterns from large datasets, while
the random forest model was designed to identify
relationships between variables. After they were
trained on the data, they were tested on a new
dataset to determine how accurately they could
identify the presence of autism. LightGBM is an
algorithm that measures the importance of each
feature in a dataset by assigning a score to each one.
We used this to identify which physical
characteristics had the highest scores, indicating that
they are most significant in giving rise to autism. To
arrive at this conclusion, the study used a
combination of genetic and physical features,
including facial features, to create a machine-
learning model to analyze the data. The model was
then tested and validated against a set of data
containing individuals with and without autism. The
results indicated that the model was highly accurate
at predicting the presence of autism, indicating the
importance of physical characteristics in identifying
autism. By catching signs of autism early, doctors
can intervene and help the patient learn coping skills
and manage the symptoms. This can help minimize
the impact of autism on their lives and increase their
quality of life, as in [12]. Background: Machine
learning algorithms, when applied to data collected
from patients with ASD, can help identify the
features of the disorder, such as social and
communication deficits, and thus enable more
accurate and efficient diagnosis. With the help of
machine learning, doctors will be able to better
identify, diagnose, and treat patients with ASD. This
is likely due to improved awareness and diagnosis
of ASD, as well as an increase in environmental
factors that can contribute to its development.
Additionally, advances in technology and medical
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.21
Prasenjit Mukherjee, Sourav Sadhukhan, Manish Godse
E-ISSN: 2224-2872
179
Volume 22, 2023
care have made it easier to identify the signs of ASD
and diagnose it in a timely manner. Early diagnosis
of ASD can have a major impact on the quality of
life of individuals with ASD, as early interventions
can be more effective and provide better outcomes.
This study seeks to provide a simple and accurate
way to classify ASD data, which can help with early
diagnosis. By randomly splitting the data and
running the experiments multiple times, we were
able to identify the best method for each dataset.
This allowed us to compare the performance of the
different methods and determine which one was the
most effective for each dataset. The accuracy of
SVM and RF was compared to the other models and
shown to be the highest. Additionally, the results
indicated that SVM was better at generalizing and
was more efficient in terms of training time, while
RF was better at handling imbalanced data. This is
likely because SVM is a discriminative classifier
that tries to classify the data points by finding the
optimal hyperplane that separates the two classes,
while RF is an ensemble method that uses a
collection of decision trees to achieve better
performance than a single decision tree.
Additionally, the RF method uses randomization to
create diversity among its decision trees, which
allows it to better handle imbalanced data. Random
Forest (RF) is an ensemble machine-learning
method that uses multiple decision trees to make
predictions. Because it combines multiple models
and considers the relationships between variables, it
has been shown to outperform other machine
learning methods when it comes to diagnosing ASD,
as in [13]. Early detection and intervention are
critical for helping children with autism get the most
out of therapy and other interventions. If screening
methods are easily implemented, it will allow for
early detection, enabling families to get their
children the help that they need as soon as possible.
It is believed that ASD is the result of a combination
of genetic, environmental, and biological factors.
Research suggests that there may be distinct
differences in the brain structure and function of
individuals with ASD, which may explain their
different behaviors and abilities. The logistic
regression model is used because of its ability to
accurately predict binary outcomes, such as whether
a child has autism or not. The algorithm will be used
to quickly process large amounts of data and make
accurate predictions based on the data in the dataset.
With machine learning, doctors can detect the
disorder more quickly and accurately by using
algorithms that look for patterns in the data. This
can help them identify the disorder earlier and
provide the necessary care to the toddler in a timely
manner, improving their quality of life. These
challenges include a lack of reliable data sets and
data infrastructure, limited access to skilled
personnel, and a lack of understanding of the legal
and ethical implications of AI-powered applications,
as in [14]. This is due to increased awareness of the
condition and improved diagnostic tools, as well as
a greater understanding of the condition and its
effects on individuals' lives. More research is being
done on the topic, leading to improved treatments
and therapies. Some people with ASD may have
difficulty with communication and forming
relationships, while others may have only mild
symptoms. Additionally, some people may have
associated medical issues, such as seizures or sleep
disturbances. Other common symptoms seen in
those with autism include limited or inappropriate
social interactions, difficulty with communication,
restricted and repetitive behaviors, and sensory
sensitivities. Diagnosis of autism can be done at any
age through observation of these behaviors, physical
examinations, cognitive testing, and genetic testing.
This is to allow for a more accurate diagnosis of
ASD as well as to enable early intervention to
ensure that the symptoms do not worsen. This is
done by using ML algorithms to analyze data such
as patient records, behavior, and medical history to
identify patterns that could indicate the presence of
ASD. LR and SVM are two popular machine
learning (ML) algorithms that can be used to
classify data. The performance measure helps to
compare the accuracy of the predictions made by the
model with each algorithm. This can help users
determine which algorithm provides more accurate
results in a shorter amount of time, which can help
them determine if they are suffering from ASD or
not, as in [15]. ASD can cause difficulties in
communication, social skills, and repetitive
behaviors. It is believed to be caused by a
combination of genetic and environmental factors
and can affect people in different ways. Early
intervention is key to reducing the effects of ASD,
as it can help children learn the skills they need to
better manage their symptoms and lead more
independent lives. It also provides an opportunity
for parents and caregivers to better understand their
children and find ways to cope with and manage the
disorder. This makes it difficult for physicians to
accurately identify ASD symptoms and recognize
them as being uniquely associated with ASD. As a
result, the diagnosis is often delayed or missed
entirely. Deep learning algorithms are able to
uncover complex patterns in large amounts of data
that may be too subtle or too complicated for a
human expert to detect. By utilizing these
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.21
Prasenjit Mukherjee, Sourav Sadhukhan, Manish Godse
E-ISSN: 2224-2872
180
Volume 22, 2023
algorithms, medical experts can be provided with
more accurate and timely diagnoses of ASD, which
can help improve treatment and outcomes for those
affected by the disorder. With a larger dataset,
machine learning algorithms can be used to develop
better models for diagnosing ASD. The algorithms
can also take into account subtle or nuanced
symptoms that may not be easily detected by
medical professionals, which can lead to more
accurate diagnoses. The hybrid approach is
beneficial because it combines the power of deep
learning to extract complex patterns from data with
the interpretability of XAI to explain why certain
features are more important than others in predicting
ASD. This helps to reduce the bias in the predictions
and makes the results more trustworthy. The
proposed framework combines data from both
parents and clinicians to create a more
comprehensive picture of a child's development.
This data can then be used to make more accurate
predictions about which children are likely to have
ASD traits, allowing clinicians to provide earlier
interventions and support, as in [16]. People with
ASD typically have difficulty with social
interaction, communication, and understanding
language. They may also show restricted or
repetitive behaviors, such as having difficulty
transitioning from one activity to another. Children
with autism can struggle with social interactions,
body language, and understanding facial
expressions. Early diagnosis can equip families with
the necessary resources and interventions to help
their child reach their full potential. With the
prevalence of ASD increasing in recent years, it is
becoming increasingly difficult for medical
professionals to diagnose the condition in children
without the help of automated methods. Automated
methods can quickly and accurately detect signs of
ASD in children, allowing medical professionals to
make more informed decisions about diagnosis and
treatment. We selected the AutoML method because
it has the ability to automate the process of building,
optimizing, and selecting the best-performing model
with minimal manual effort. In addition, AutoML
can also be used to identify important features in the
dataset, which can then be further used to improve
the accuracy of the machine-learning models. This
is due to the fact that AutoML automates the
process of selecting optimized feature combinations
and hyperparameters, allowing us to quickly
identify the optimal settings for our model. The
combination of these techniques allowed us to
achieve the highest accuracy with minimal effort, as
in [17]. ASD is caused by a combination of genetic
and environmental factors, including gene mutations
and exposure to toxins. People with ASD may also
have trouble forming social relationships, have
difficulty with communication and language, and
struggle with sensory sensitivity. MRI imaging
modalities have the capability to detect subtle brain
abnormalities that are associated with ASD, such as
changes in the brain’s structure, connectivity, and
even chemistry. This makes it an invaluable tool for
diagnosing and monitoring ASD. fMRI uses
magnetic fields and radio waves to measure blood
flow in the brain and identify any abnormalities or
discrepancies in brain activity. sMRI uses high-
resolution images to map the structure of the brain
and detect any abnormalities in the brain's anatomy.
These two modalities work together to help
clinicians diagnose ASD with greater precision.
These systems use AI to analyze brain images, such
as MRI and fMRI scans, to assess an individual's
brain structure and connectivity. The AI algorithms
can detect subtle differences in brain structures,
which can be used to diagnose ASD more accurately
and quickly by specialists. ML algorithms are used
to analyze the image data, identify the relevant
features, and detect any abnormalities that could be
indicative of ASD. DL applications are used to
further analyze the data and identify patterns that
may be indicative of ASD. This allows for more
accurate and reliable diagnoses. Deep learning (DL)
techniques employ large datasets of MRI images
and AI algorithms to create models that can detect
patterns in the images that are associated with ASD.
These models can then be used to automate the
diagnosis of ASD and provide more accurate and
timely results. We compare the accuracy and
training times of ML and DL models to show that
DL models can learn faster and achieve higher
accuracy. We also discuss the importance of feature
selection and data pre-processing in improving the
accuracy of the models. Finally, we suggest the
potential of combining AI techniques with MRI
neuroimaging to detect ASDs, as in [18]. It is
usually diagnosed during early childhood, and
symptoms can range from mild to severe. Common
characteristics of ASD are difficulty with social
interactions, difficulty with verbal and nonverbal
communication, difficulty with sensory integration,
and an overall difficulty in adapting to change. As a
result, many healthcare providers are looking for
more cost-effective ways to diagnose ASD, such as
through the use of screening tools that can help
identify the presence of ASD symptoms in a shorter
amount of time. Additionally, research has found
that early detection and intervention of ASD can
have a significant impact on the child's
development, so it is important to identify the
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.21
Prasenjit Mukherjee, Sourav Sadhukhan, Manish Godse
E-ISSN: 2224-2872
181
Volume 22, 2023
disorder as quickly as possible. These methods are
designed to provide a clearer picture of what is
happening in the person's life, allowing for a more
accurate diagnosis. The AQ and M-CHAT use
standardized questions about social interaction,
communication, and behavior to assess the
individual's level of autism spectrum disorder. The
user must be knowledgeable about the various items
that need to be screened and be able to identify any
discrepancies that could lead to inaccurate results.
The screening items must be designed in such a way
that they allow for accurate and efficient screening.
ML algorithms can process large amounts of data
quickly and efficiently. By taking advantage of such
algorithms, we can greatly reduce the time needed
to detect patterns, uncover trends, and identify
anomalies in the data. These patterns and trends can
then be used to make more accurate diagnoses,
leading to improved accuracy and efficiency in the
diagnostic process. RML is based on a combination
of rule-based and ML techniques, which allows it to
detect patterns in data that traditional ML
techniques cannot. Furthermore, it provides users
with interpretable rules that can be used to gain a
better understanding of the data as well as identify
potential areas for further research. This is likely
due to RML's ability to learn from the data and
identify patterns in the data that are not visible to
traditional ML methods. Additionally, RML's ability
to handle complex data and its ability to adjust to
new data as it comes in make it a powerful tool for
classification, as in [19].
3 Machine Learning Models in ASD
Detection
Today, artificial intelligence has established its
presence in all sectors, including healthcare. Autism
Spectrum Disorder (ASD) detection is a difficult
challenge in the healthcare domain. Early detection
of ASD is needed to start treatment to reduce all the
symptoms of ASD. ASD is not curable, but it is
possible to manage its symptoms. Parents have a
crucial role in detecting ASD at the early age of a
baby. In the detection of ASD, many types of
research have been done or are in progress to use
machine learning models and various kinds of
datasets. In this section, a discussion has been done
on the detection of ASD using machine-learning
models. The discussion has progressed according to
ASD detection cases. Each case has been described
with the proper dataset, machine learning models,
and model performance. The discussion about each
framework has been given in the next section. Our
primary aim is to understand the architecture of each
system where simulation and numerical stability
have been normalized, as in [11], [12], and [13].
3.1 ASD Detection Using Facial Images
Machine learning models are used to detect ASD
among children. Machine learning models like
Xception, VGG19, and NASNETMobile are very
advanced image-based machine-learning models.
Fig. 1. Framework of ASD Detection using ML
3.1.1 Dataset
The dataset [11] has been prepared using images of
autistic children and general children. The facial
Images have been captured, which are the main
input for the machine-learning models. The dataset
has been prepared with 2940 facial images, of which
half are of autistic children and the remaining half
are of general children. The images have been
collected from the social autism groups on
Facebook, as in [11].
3.1.2 Framework
A clear framework to detect ASD has been given,
where each section is described in Fig. 1. According
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.21
Prasenjit Mukherjee, Sourav Sadhukhan, Manish Godse
E-ISSN: 2224-2872
182
Volume 22, 2023
to Fig. 1, the given framework [11] shows that data
will be read from the dataset and split into the train
and test parts in the first and second steps. The
model will be prepared to train with training data.
After the completion of training, the model will be
fine-tuned to reduce the error in the next step. After
reducing the error, the model will be fitted and
tested with test data for validation. After completion
of this step, the model will be ready to predict using
new data from the user side. The predicted result of
the test data will be utilized to calculate the
accuracy, precision, recall, and confusion matrix in
the final step. Three advanced machine-learning
models have been used to detect autism from facial
images, as in [11]. The input dataset contains the
image data for training and testing these models.
The result of these models has been discussed in
section IV.
3.2 ASD Detection of Toddlers Using
Machine Learning Models
This work [12] has been proposed to detect ASD in
toddler children. The age range of toddler children
is between 12 months and 3 years. This is a good
time to detect ASD among children because early
detection helps to start ASD therapies according to
the need. AI already accepts this challenge to find
out the solution to early detection of ASD among
children. Many models have been developed that are
useful in the detection of ASD. XGBoost, neural
networks, and random forest models have been used
to detect ASD among toddlers, as in [12].
3.2.1 Dataset
The dataset [12] has been collected from
Kaggle.com, which is an open-source repository of
machine learning. The autism dataset was prepared
by the University of California, Irvine. This dataset
contains the screening data for toddlers. The dataset
contains 1054 records with 18 variables that point to
different attributes. 10 variables are questions that
determine ASD among toddlers. These 10 variables
are questions related to autism. The questions are set
from A1 to A10. If the answer to Questions A1 to
A9 is “sometimes”, "rarely," or "never," then the
value will be assigned as 1, and 0 will be the
opposite of these answers. If the answer to question
A10 is "always,", "usually,", or "sometimes,", then
the value will be assigned as 1, and 0 will be the
opposite of the answer. The scores of these
questions and other attributes have been used to
train the models for the prediction of ASD, as in
[12].
Fig. 2. Framework of ASD detection among
Toddlers using ML
3.2.2 Framework
Fig. 2 shows the framework of this system [12],
which is equipped with machine learning models
like XGBoost, neural networks, and random forests.
The data will be read from the dataset. The
preprocessing task will be applied when it requires
some cleaning in the second step. According to the
third step, the data will be split into training and
testing parts. Now, each model, like XGBoost,
neural networks, and random forests, will use
training data to train and understand the pattern, as
in [12]. After the completion of training, each model
will be evaluated using testing data. In the end,
models are ready to predict results according to the
user’s input. The Random Forest model has been
used with pre-optimization and post-optimization.
The XGBoost model is an ensemble model that is
equipped with many weak models. XGBoost stands
for ‘Extreme Gradient Boosting’ and is the most
popular machine learning model that accepts large
datasets, and its overall performance is good and
stable. The neural network has been developed by
inspiring the human brain. Neural networks are used
to solve complex machine-learning problems
because of their ability to compute quickly and
generate responses quickly. The other model is the
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.21
Prasenjit Mukherjee, Sourav Sadhukhan, Manish Godse
E-ISSN: 2224-2872
183
Volume 22, 2023
Random Forest, which is the most popular model to
solve classification problems in machine learning.
The random forest model has been developed by the
decision trees. A decision tree is the key point of the
Random Forest algorithm. The results of these
models have been discussed in Section IV.
3.3 ASD Detection Using Traditional
Machine Learning Models
The Support Vector Machine (SVM), K Nearest
Neighbour (KNN), and Random Forest models are
used to detect Autism using the UCI dataset as in
[13].
3.3.1 Dataset
Three datasets [13] have been used to solve the
ASD detection problem. These three datasets have
been taken from the UCI database. The three
datasets are AQ-10-Adult for adults, AQ-10-
Adolescence for adolescents, and AQ-10-Child for
children. The data has been classified into train data
and test data with different values. The values will
be selected randomly. The score of each subset of
data is measured with average accuracy, average
sensitivity, average F-measure, and average AUC,
as in [13].
3.3.2 Framework
The framework [13] is equipped with three models:
SVM, KNN, and Random Forest. All these models
are best for the classification problem. The first
model is SVM, which creates the best line or
decision boundary to classify the n-dimensional
space for plotting new data points in the correct
category. SVM uses the vectors to create the
hyperplan. The optimum hyperplan segregates the
vectors that define the classes. The KNN is another
supervised machine learning algorithm that can be
used for classification or regression. The K is the
nearest neighbor that has been used by the KNN
algorithm. A majority vote for a particular class
determines that a new observation should be inside
it. Larger values of K refer to stable decision
boundaries for classification, whereas small values
of K refer to decision boundaries that are not better
than a larger K value. Random Forest is a popular
model in classification. This model contains a
number of decision trees according to the various
subsets of the dataset, and it will calculate the
average for prediction. The greater number of
decision trees refers to the higher accuracy that
prevents the overfitting problem, as in [13].
Fig. 3. Framework of ASD detection using
Traditional ML
According to Fig. 3, first, the data is read from the
UCI dataset [13] and split into training and testing
sets. Then, a model is prepared for training using the
training data. After the model is trained, it is fine-
tuned to reduce errors and increase accuracy. The
model is then fitted with the data and ready for
testing using the test data. Once testing is complete,
the model can be used to predict results based on the
new data. The machine learning models SVM,
KNN, and Random Forest will be trained using the
UCI data as in [13]. The accuracy of each model has
been discussed in Section IV.
4 Results and Discussion
The models have been discussed with a framework
in Section III. This section refers to the discussion
about the results of each model. The first framework
is ASD detection using facial images of autistic
children and general children, as in [11]. The second
framework is the ASD detection of toddlers using
screening data [12], and the third frame is the ASD
detection from the UCI dataset as in [13]. Each
framework contains machine learning models, and
these models have been used for prediction after
successful training and testing.
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.21
Prasenjit Mukherjee, Sourav Sadhukhan, Manish Godse
E-ISSN: 2224-2872
184
Volume 22, 2023
4.1 Result of the Models of ASD Detection
Using Facial Images
The first framework [11] is about ASD detection
using facial images. Three deep learning models—
Xception, VGG19, and NASNetMobile—have been
implemented to recognize ASD among children
using their facial images as data. The accuracy of
the Xception model has scored 91%, whereas the
specificity and sensitivity of this model are 94% and
88%, respectively. The VGG19 model has scored
80% accuracy, and its specificity and sensitivity
values are 83% and 78%, respectively. The
NASNetMobile model has 78% accuracy, 75%
specificity, and 82% sensitivity. Specificity, on the
other hand, measures the ability of a model to
correctly identify negative instances of a given
category. It is calculated as the number of true
negatives divided by the sum of true negatives and
false positives. Sensitivity is a metric that measures
the ability of a model to correctly identify positive
instances of a given category. It is calculated as the
number of true positives divided by the sum of true
positives and false negatives. The accuracy,
specificity, and sensitivity of each model have been
given in Table 1 as in [11].
Table 1. Accuracy, Specificity, and Sensitivity
of Each Model
Sl.
No.
Models
Specificit
y
Sensitivity
Accuracy
1
Xception
0.94
0.88
0.91
2
VGG19
0.83
0.78
0.80
3
NASNE
TMobile
0.75
0.82
0.78
4.2 Result of ASD Detection of Toddlers
Using Screening Data
The baseline XGBoost model [12] has performed
well to detect ASD among toddlers. The toddler’s
dataset contains 18 variables, where A1 to A10 are
questions that need answers to train the model. The
other machine learning models are also used to
detect ASD among toddlers. Neural networks and
Random Forest pre- and post-optimization are the
models that are used, and their performance has
been given in Table 2 as in [12].
Table 2. Performance Scores of ASD Detection
Models among Toddlers
Sl.
No.
Model
Precision
Recall
F1
Accuracy
1
Neural
Network
100%
100%
100%
100%
2
Random
Forest(Pre
-
Optimizati
on)
98.15%
98.10
%
98.09
%
98.10%
3
Random
Forest(Pos
t-
Optimizati
on)
100%
100%
100%
100%
4
XGBoost
97.04% -
Mean
Accuracy
and
1.78%
Standard
Deviation
The performance scores of each model can be seen
in Table 2, where the neural network model has
100% accuracy with 100% precision and a 100%
recall value. The Random Forest (post-optimization)
model has the same scores as the neural network
model. It has 100% scores in precision, recall, and
accuracy. The Random Forest (pre-optimization)
scored 98.15% in precision, 98.10% in the recall,
and 98.10% in accuracy, whereas XGBoost has a
97.04% accuracy score with a standard deviation
value of 1.78%, as in [12].
4.2 Result of ASD Detection Using
Traditional Machine Learning Models
The UCI dataset [13] has been taken as the main
data source to detect ASD among children. The
three most popular traditional machine learning
algorithms have been used for ASD prediction,
according to the new input. These models are KNN,
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.21
Prasenjit Mukherjee, Sourav Sadhukhan, Manish Godse
E-ISSN: 2224-2872
185
Volume 22, 2023
SVM, and Random Forest (RF). The performance of
each model has been given in Table 3 as in [13].
Table 3. Performance of Traditional Machine
Learning Models to Detect ASD among Children
Sl.
No
.
Models
Case
AUC
1
KNN
AQ-10-Adults for the
Case of Complete Data
0.94
2
SVM
AQ-10-Adults for the
Case of Complete Data
1.00
3
RF
AQ-10-Adults for the
Case of Complete Data
1.00
4
KNN
AQ-10-Adults for the
Case of Missing Data
0.93
5
SVM
AQ-10-Adults for the
Case of Missing Data
1.00
6
RF
AQ-10-Adults for the
Case of Missing Data
1.00
7
KNN
AQ-10- Adolescence for
the Case of Complete
Data
0.87
8
SVM
AQ-10- Adolescence for
the Case of Complete
Data
0.97
9
RF
AQ-10- Adolescence for
the Case of Complete
Data
1.00
10
KNN
AQ-10- Adolescence for
0.85
the Case of Missing Data
11
SVM
AQ-10- Adolescence for
the Case of Missing Data
0.98
12
RF
AQ-10- Adolescence for
the Case of Missing Data
1.00
13
KNN
AQ-10-Child for the Case
of Complete Data
0.85
14
SVM
AQ-10-Child for the Case
of Complete Data
0.89
15
RF
AQ-10-Child for the Case
of Complete Data
0.99
16
KNN
AQ-10-Child for the Case
of Missing Data
0.85
17
SVM
AQ-10-Child for the Case
of Missing Data
0.91
18
RF
AQ-10-Child for the Case
of Missing Data
1.00
Table 3, Table 3 shows the performance graph of
the KNN, SVM, and RF, which is based on the
AUC scores. Six cases have been classified as: 1.
AQ-10-Adults for the Case of Complete Data, 2.
AQ-10-Adults for the Case of Missing Data, 3. AQ-
10: Adolescence for the Case of Complete Data; 4.
AQ-10: Adolescence for the Case of Missing Data;
5. AQ-10: Child for the Case of Complete Data; and
6. AQ-10: Child for the Case of Missing Data. The
AUC score has been calculated by the true positive
rate and the false positive rate, as in [13]. The AUC
scores according to the first case of KNN, SVM,
and RF are 0.94, 1.00, and 1.00. In the second case,
the AUC scores of KNN, SVM, and RF are 0.93,
1.00, and 1.00. The AUC scores of KNN, SVM, and
RF in the 3rd case are 0.87, 0.97, and 1.00, whereas
the AUC scores are 0.85, 0.98, and 1.00 in the 4th
case. The AUC scores of KNN, SVM, and RF are
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.21
Prasenjit Mukherjee, Sourav Sadhukhan, Manish Godse
E-ISSN: 2224-2872
186
Volume 22, 2023
0.85, 0.89, and 0.99 in the 5th case, whereas 0.85,
0.91, and 1.00 in the 6th case, as in [13].
Each framework with machine learning models
has been discussed with figures and tables that
indicate the procedure for ASD detection among
children or adults. These models are good for
detecting ASD, but it is more important that models
detect the kinds of symptoms in ASD individuals.
Early detection of ASD among children is much
more important than ASD detection among adults.
You will get a good result if you start therapies on
early ASD-detected children. If ASD is detected
after a certain age, then it will be difficult to get any
good improvement results for ASD individuals. The
discussed frameworks will work on ASD detection,
but ASD symptoms are also needed for further
understanding of the need for therapies. Sometimes
fixed attributes are used in the dataset, or some
questions regarding ASD have been used in the
dataset with answers, but according to the ASD
problem, the best data source for ASD is parents
because an ASD individual spends most of the time
with the parent. Facial images are also used to detect
autism, but it is difficult to segregate ASD and
ADHD children through images. A child may have
an ASD problem or ADHD problem, as well as a
Global Development Delay (GDD) problem, which
can be there. Facial images cannot segregate this
problem correctly. If there are hundreds of ASD
children, and if we find patterns, then it is possible
to get a hundred different patterns from ASD
children. There are no fixed patterns to detect ASD
among children, but a few may be common. The
study of these three frameworks has given a clear
understanding of the strong role of AI in ASD
detection, where a hybrid approach can be executed.
5 Application of the Proposed Study
The application of the proposed study is to
understand the various techniques of ASD detection
among children. Today, autism is a major issue
among children, according to the World Health
Organization (WHO, https://www.who.int/news-
room/fact-sheets/detail/autism-spectrum-disorders).
Many applications have been developed using
machine learning and natural language processing to
detect autism in the early stages, but the research
method may not be cost-effective and the data
regarding this problem is not up to par. The
proposed study can be useful in identifying space
for further research, like parent-child dialogue with
an autistic child. The deep learning models can be
applied to understand the symptoms of autism at an
early age from the parent's dialogue.
6 Conclusion
Three frameworks have been discussed with
machine learning models to detect ASD among
children, toddlers, and adults. The first approach is
ASD detection using the facial images of ASD
children and general children. Advanced machine
learning models have been used to detect ASD.
These models are trained with facial images, and
their accuracy is very high for the detection of ASD.
The second framework is about ASD detection
among toddlers. This framework used some
screening data from toddlers to train the machine
learning models. ASD detection at an early age is a
good option to start therapies to reduce the
symptoms of ASD. Each model has been trained
with the ASD dataset, and performance scores are
high according to the predictions. The third
framework contains some traditional machine-
learning models that are popular machine-learning
models for classification problems. These models
are able to predict ASD among children,
adolescents, and adults with high accuracy after
training with complete and missing data. These
three kinds of frameworks have elaborated on AI
applications in the healthcare domain with strong
results. The deep learning models can be applied to
the parent-child dialogues of an autistic child. The
parent's dialogues are nothing but textual
information about their children, and this data can
be utilized for identifying the symptoms of autism.
After the detection of symptoms from the parents’
dialogues, the symptoms can be analyzed according
to the severity of autism, and this task will be a
future enhancement.
Acknowledgement:
The authors extend their appreciation to the
Manipur International University, Imphal, India for
supporting this research work on Autism.
References:
[1] Maria Lai, Jack Lee, Sally Chiu, Jessie Charm,
Wing Yee So, Fung Ping Yuen, Chloe Kwok,
Jasmine Tsoi, Yuqi Lin, Benny Zee, A
machine learning approach for retinal images
analysis as an objective screening method for
children with autism spectrum disorder,
EClinical Medicine, 2020, pp. 1-20.
[2] C. S. Paula, S. H. Ribeiro, E. Fombonne, and
M. T. Mercadante, Brief report: prevalence of
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.21
Prasenjit Mukherjee, Sourav Sadhukhan, Manish Godse
E-ISSN: 2224-2872
187
Volume 22, 2023
pervasive developmental disorder in Brazil: a
pilot study, Journal of Autism and
Developmental Disorders, vol. 41, no. 12,
2011, pp. 1738–1742.
[3] L. C. Nunes, P. R. Pinheiro, M. C. D. Pinheiro
et al., A Hybrid Model to Guide the
Consultation of Children with Autism
Spectrum Disorder, A. Visvizi and M. D.
Lytras, Eds., Springer International
Publishing, View at: Google Scholar, 2019, pp.
419–431.
[4] Apa–American Psychiatric Association,
Diagnostic and statistical manual of mental
disorders (DSM –5), 2020,
https://www.psychiatry.org/psychiatrists/practi
ce/dsm.
[5] R. Carette, F. Cilia, G. Dequen, J. Bosche, J.-L.
Guerin, and L. Vandromme, Automatic autism
spectrum disorder detection thanks to eye-
tracking and neural network-based approach,
the International Conference on IoT
Technologies for Healthcare, Springer, Angers,
France, 2017, pp. 75–81.
[6] L. Kanner, Autistic disturbances of affective
contact, Nerv. Child, Vol. 2, 1943, pp. 217–
250.
[7] E. Fombonne, Epidemiology of pervasive
developmental disorders, Pediatric Research,
Vol. 65, no. 6, 2009, pp. 591–598.
[8] D. Aarthi, M. Udhayamoorthi, G. Lavanya,
Autism Spectrum Disorder Analysis using
Artificial Intelligence: A Survey, International
Journal of Advanced Research in Engineering
and Technology, Vol. 11(10), 2020, pp. 235-
240.
[9] N. Ajaypradeep, R. Sasikala, Child Behavioral
Analysis: Machine Learning based
Investigation for Autism Screening and Early
Diagnosis, International Journal of Early
Childhood Special Education, Vol. 13(2),
2021, pp. 1199-1208.
[10] 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.
[11] Fawaz Waselallah Alsaade and Mohammed
Saeed Alzahrani, Classification and Detection
of Autism Spectrum Disorder Based on Deep
Learning Algorithms, Computational
Intelligence and Neuroscience, 2022, pp. 1-10.
[12] Arjun Singh, Zoya Farooqui, Branden Sattler,
Unyime Usua, Michael Helde, Using Machine
Learning Optimization to Predict Autism in
Toddlers, 11th Annual International
Conference on Industrial Engineering and
Operations Management, Singapore, 2021, pp.
6920-6931.
[13] Uğur Erkan1, Dang N.H. Thanh, Autism
Spectrum Disorder Detection with Machine
Learning Methods, Current Psychiatry
Research and Reviews, Vol. 15(4), 2019.
[14] Dr. Sherif Kamel, Rehab Al-harbi, Newly
proposed technique for autism spectrum
disorder based machine learning, International
Journal of Computer Science & Information
Technology (IJCSIT), Vol. 13(2), 2021.
[15] Sriram Dhanyatha , A. Greeshma, Gouthami,
M. Yeshwanth, Y Shobha, Prediction of
Autism Spectrum Disorder based on Machine
Learning Approach, International Research
Journal of Engineering and Technology
(IRJET), Vol. 8(7), 2021, pp. 2907-2917.
[16] Anupam Garg, Anshu Parashar, Dipto Barman,
Sahil Jain, Divya Singhal, MehediMasud,
Mohamed Abouhawwash, Autism Spectrum
Disorder Prediction by an Explainable Deep
Learning Approach, Computers, Materials &
Continua, Vol. 71(1), 2022, pp. 1459-1471.
[17] Basma Ramdan Gamal Elshoky, Eman M. G.
Younis, Abdelmgeid Amin Ali, Osman Ali
Sadek Ibrahim, Comparing automated and non-
automated machine learning for autism
spectrum disorders classification using facial
images, ETRI Journal, 2021, pp. 613-623.
[18] P. Moridian1, N. Ghassemi, M. Jafari, S.
Salloum-Asfar, D. Sadeghi, M. Khodatars, A.
Shoeibi, A. Khosravi, S. H. Ling, A. Subasi, R.
Alizadehsani, J. M. Gorriz6, Sara A Abdulla,
U. Rajendra Acharya, Automatic Autism
Spectrum Disorder Detection Using Artificial
Intelligence Methods with MRI Neuroimaging:
A Review, Frontiers in Molecular
Neuroscience, Vol. 15, 2022, pp. 1-51.
[19] Fadi Thabtah, David Peebles, A New Machine
Learning Model based on Induction of Rules
for Autism Detection, Health Informatics
Journal, 2020, pp. 1-23.
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.21
Prasenjit Mukherjee, Sourav Sadhukhan, Manish Godse
E-ISSN: 2224-2872
188
Volume 22, 2023
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
The authors equally contributed in 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
that are relevant to the content of this article.
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 COMPUTERS
DOI: 10.37394/23205.2023.22.21
Prasenjit Mukherjee, Sourav Sadhukhan, Manish Godse
E-ISSN: 2224-2872
189
Volume 22, 2023
