International Journal of Engineering Trends and Technology Volume 70 Issue 4, 343-359, April 2022

ISSN: 2231 – 5381 / https://doi.org/10.14445/22315381/IJETT-V70I4P230 © 2022 Seventh Sense Research Group®

Original Article

Effective Autism Spectrum Disorder Prediction to

Improve the Clinical Traits using Machine Learning
Techniques
Dr. R. Surendiran1, Dr. M. Thangamani2, C. Narmatha3, M. Iswarya4
1
School of Information Science, Annai College of Arts and Science, Kumbakonam, India, ORCID: 0000-0003-1596-7874
2
Department of Information Technology, Kongu Engineering College, Erode, India. ORCID: 0000-0001-8864-3315
3,4
Department of Computer Technology, Kongu Engineering College, Erode, India.

1Corresponding author: [email protected]

Received: 17 November 2021 Revised: 03 March 2022 Accepted: 10 April 2022 Published: 30 April 2022

Abstract - Autism spectrum disorder (ASD) is a neuro- persons. Still, individuals with ASD could converse,
developmental complaint that influences an individual’s interrelate, function, perform and hear in the approach part
communication, announcement, and knowledge talents. from mainly individual people. The knowledge, thoughts,
Analysis of Autism can be completed at any age-group level. and difficulty-resolving facility of persons with ASD can
Autism patients look at diverse kinds of disputes learning provide a choice from talented to more confronted. An
disabilities, and complexity with meditation. Mental health analysis of ASD contains a lot of provisions that applied to
problems, motor difficulties, and sensory problems are some be analyzed alone as autistic disorder syndrome. These
of the problems faced by Autism patients. Earlier diagnosis situations in named ASD.
and proper medication at the early stage are essential to
control ASD. The ASD prediction framework is built to ASD is a state connected to mind growth that forces how
support a behavioral aspect-based analysis model without a person recognizes and meet peoples, obtaining difficulty in
any device in this research. The ASD prediction process is public contact and dealings. The disorder also contains
focused on the childhood and adolescent analysis model genetic factors, environmental aspects, biological factors,
utilized in the system. The behavioral parameters are and inadequate and recurring behavior patterns. In the United
collected with the support of the Autism Query collections. States, more children's affected by Autism. Each person with
The decision tree (DT) and Support Vector Machine (SVM) Autism has different issues, like anxiety, seizures, and
techniques, K-Nearest Neighbors (KNN), and Artificial depression, as shown in Figure 1. Autism can be obtained at
Neural Network (ANN) are applied for the ASD prediction two or three and diagnosed as early as before 18 months.
process. The Correlated Feature selection based Random Early detection gives a good impact on the life of persons.
Forest (CFS-RT) algorithm is applied for the ASD prediction Persons with ASD may perform work, interconnect,
process, giving an accuracy of 93.03%, and ANN produces interrelate, and acquire behaviors that are changed from
97.68% and outperformance other methods. person to person. The skills of a person with ASD can differ
significantly. The person with ASD are issues with the social
Keywords - Autism Spectrum Disorder, Decision Tree, announcement and limited behaviors. The children have
Machine Learning, Data Mining, Support Vector Machine. varying habits of culture and give less attention. Diagnosing
ASD is also a problem because there is no proper medical
1. Introduction test or blood test to identify the disorder. Doctors see the
Autism spectrum disorder (ASD) is a progressive child's behavior and can start a diagnosis to increase the
disability that can benefit important basis community, quality of life.
contact, and behavioral issues. There are frequent tells how a
community with ASD looks that sets them together from new

This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)

 Dr. R. Surendiran et al. / IJETT, 70(4), 343-359, 2022

Oxidative stress Biologic factor

Anxiety

ASD Genetic Factors

Environment Factors

Seizures
Depression Immune system function

Fig. 1 Outline the Factors for obtaining ASD

ASD disturbs how an individual reacts, study, and states themselves. For parents of a child with ASD, significantly which
level the child has can help prepare them for issues their child might face in daily life. Figure 2 shows the level of ASD. The
first level needs help for every work, and the second level indicates the need for sustainable help. Three-level indicate the need
for more sustainable help and very problem to changing focus and actions.

ASD stage 1 ASD stage 2 ASD stage 3

Fig. 2 Type of ASD

1.1 Supervised Machine Learning

Machine Learning (ML) is a division of Artificial Intelligence (AI) that studies and determines useful models in data to
formulate diagnoses in the medical field. It is put up on mathematical opinions of likelihood and data and information. This
supervised learning is an ML approach that studies by plotting the input features and output mutable for prediction. It is shown
in Figure 3 and learning through training and test data set.

INPUT Generate OUTPUT

Feature Set (w) Classifier Label (z)
Model

Fig. 3 Classifier Model

As shown in Figure 1, a representative controlled learning difficulty includes attribute space W called features in the data
set, a label space Z, and a prediction space Z’. Classifier map each feature w to one of the already defined class labels z.
A training sample set is illustrated as Y = ((w1,z1),…,(wm,zm)) Є (W ×Z)m that consists of prearranged class labels, where
each wi Є Z. The output is a model Hs: W Z’ that learns from the sample set These approaches find the class of new
instances, Z'= Z = {±1} and input samples are a map with two labels as Z = {±1} = {-1, +1} in binary classification.

344
 Dr. R. Surendiran et al. / IJETT, 70(4), 343-359, 2022

1.2 Classification Algorithms 2. Related Work

The ID3 algorithm is an approximation supervised Thangamani et al.[1-3] investigated different models for
learning algorithm through the sample for various labels. The disease prediction using ML approaches. Early detection and
decision tree root and internal nodes hold the feature test follow-up healings have the most important blow on the
uniqueness, whereas the leaf node includes the class label. autistic group. Dilantha et al. [4] constructed an EEG
The decision tree holds numeric and categorical variables approach control improved classifier approach for ASD. The
from the sample and produces a visual representation. Many present diagnostic follow-up is behavior needy and delays
open-source tools contain a build-in decision tree algorithm. the prediction at a premature age. This research used the
For example, WEKA produces good data visualization by feature extraction technique to extract the relevant attributes.
visualize the tree option. The authors [35] applied ML and data mining algorithms to
predict ASD in toddlers and children. The researchers [6]
ID3 identifies features that discriminate one label from developed the system with the help of ML techniques for
another. The features should be continuous or chosen from a autism grading and assessment for the diagnosis of autistic
group of identified values, and the feature should be well- children. It produces 98% of accuracy. Inon Wiratsin et al.
known in progress. At the same time, this algorithm is more [7] proposed a feature selection algorithm to detect the
responsive to features when a huge number of data. relevant attributes and suggest treatment for various age
Limitations in the ID3 are overcome through the C4.5 groups. Bekerom et al.[33] investigated the ML algorithm for
algorithm with the help of information gain measures. early detection of ASD to develop good mental health of the
child. Daniel Bone [9] planned to use the ML approach to get
Naïve Bayes is a supervised learning classifier for data a better result in Autism and diagnosis with multi-level
mining and the field of machine learning that solve the instrument fusion for human behavior analysis and clinical
classification issues by generating the Bayes rule. This progress. Lucia Billeci [10-11] deliberated an incorporated
classifier holds any volume of data and also holds missing eye way techniques with EEG for learning of replying and
values effectively. The target is to forecast the class l as the beginning joint consideration in ASD. The correlation
label, fi as features, and select a significance of l that exploit between eye and brain features is measured in [12-13]. The
P(l|F1, F2,…, Fn) given a sample with attributes (m1, m2,…, authors [14-17] illustrated the study of EEG signals for ASD
mn) is shown in equation (1) detection using ML approaches. William J. Bosl [18]
suggested a data drive method for EEG diagnostics for a
𝑃(𝑚 |𝑙 )𝑃(𝑙) timely finding of ASD. Zhong Zhao et al.[19 ] addressed
P(c|F) – (1)
𝑃(𝑚)
restricted features are used according to patient behavior
patterns, which is used in ML algorithms to detect the ASD
K-Nearest Neighbors is an instance-based ML method prediction and produce an accuracy of 88.7%. J. Baio et al.
for classifier, regression approach, and it’s lazy learning [20] developed a surveillance network for ASD prediction,
because there is no explicit training. This algorithm stores approximating the popularity of ASD and further unique
the training observations and evaluates against the new children and evaluating the system by skilled clinicians.
opinions with the help of a similarity assessment as
Euclidean distance according to real and binary values as Thabtah et al. [34] designed an ASD screening method
shown in equations (2) and (3)and positions the stored data at with ML variation, which perform classification task and
some point in the test instant to formulate calculations. This generate a predictive model to detect the ASD effectively.
algorithm also holds the missing values effectively. The authors [22] focused on ML algorithms with behavioral
features to identify the ASD at the earliest. N. Jassimet al.
Euclidean Distance: [23] quantitatively compressed discovery from fMRI
f(wi,zj) = √∑𝐷 2 learning of non-community sensory observation in autistic
𝑖𝑚=1(𝑤𝑖𝑚 − 𝑤𝑗𝑚 ) (2)
evaluated to direct representative contributors by the
behavior of a series of conservatively- threshold Activation
Hamming Distance: Likelihood Estimation meta-analyses. Suman Raj et al. [24]
f(wi,zj) = ∑𝐷
𝑖𝑚=1 𝐼𝐼(𝑤𝑖𝑚 ≠ 𝑤𝑗𝑚 )
2
(3) Suggested ML Algorithm predicts and diagnoses ASD
problems for adults, adolescents, and children. J. Charpentier
The sample data j fields to frame a j-dimensional vector, et al. [25] diagnosed ASD by the mutual presence of societal
w=(w1,w2,…,wj), and the training data is characterized by destructions and warning, recurring performance models.
F={(w1,z1),…(wj,zj)} where xi is the content of the input and Antoine Frigaux et al. [36] proposed a new technique to
yi is the named as the label. The wi is a vector that contains F detect autism spectrum disorder. Jennie Hayes et al. [27]
attributes, whereas xim represents the m-th attributes of w. reviewed all the approaches used and suggested clinical
That means the result is zi Є{1,..., C} guidelines estimate ASD in adults and children in the UK.
Mengyi Liao et al.[28] investigate the ASD by facial and
EEG data of children with the help of the ML approach and

345
 Dr. R. Surendiran et al. / IJETT, 70(4), 343-359, 2022

obtain 87.50%. This model used the naïve Bayes algorithm discover ASD with behavioral features from different age
and did not train more than one model. Chelsea et al. [29] levels.
reviewed the state of MLCand ASD with the help of some
questions and answered for quires. The authors use the ML Specific Objectives
model with MRI and fMRI attributes to know the neuronal • To design a system for Autism Spectrum Disorder
behavior changes in ASD to predict the disease earlier [30- (ASD) discovery with behavioral analysis.
31]. • To perform age group based analysis on the ASD
discovery process
From existing research, Autism screening tests are a • To design the Correlation-based Feature Selection
huge cost and not a time-saving one. The behavioral with Random Forest (CFS-RT) technique for the
differences are not measured in the device-dependent ASD ASD prediction process
prediction process. Technical people are required to handle • To eliminate device and technician dependency in
the data collection operations. Age group-based analysis is the ASD detection process
not supported in the ASD prediction process. • To minimize cost and time in the ASD prediction
operations
3. Problem Definition
The AI and ML techniques are used in autism detection. 4. Proposed System Methodology
The neuroimaging and kinematic data are used to discover 4.1 ASD Detection through Behavioral Features using
the Autism Spectrum ASD levels. The ASD classification is Decision Trees
carried out with electroencephalogram (EEG) signal Adult autism detection is designed to analyze the
dispensation and knowledge approaches. The EEG is a behavioral attributes of patients above 18 years. The
consistent biomarker for diagnosing ASD abnormalities. The behavioral features are gathered with the help of a
electrodes attached to the scalp are used to detention the EEG questionnaire. The Adult Questioner (AQ10) data set is get
signals. The mobile applications collect answers from the from the University of California, Irwin ML repository. Each
people for the Autism Questions. The MLg-based methods person is diagnosed with 10 questions. The ASD is identified
are applied to predict ASD at any age-group level. The using the responses given by the individual. The decision
Random Forest-CART (Classification and Regression Trees) tree classification is applied to discover the disease levels.
algorithm predicts ASD traits. The Random Forest-ID3 Figure 4 shows the proposed system architecture, and a
(Iterative Dichotomiser 3) algorithm is applied to perform the detailed view is shown in Fig. 5.
ASD prediction process. The Correlation Feature Selection
based Random Forest (CFS-RF) technique is built to
Adult ASD Data
Childhood ASD Data

Data Collection and

Preprocess

Autism Discovery Correlated Feature

Selection

Decision Tree Support Vector

Machine (SVM) Random Forest
Classification model
Classification model

Prediction Results

Fig. 4 System Architecture Diagram

346
 Dr. R. Surendiran et al. / IJETT, 70(4), 343-359, 2022

The children dataset and the adolescent dataset is stored in the data repository. It contains missing and incorrect values.
This can be preprocessed and get relevant data to build the ML classification models after preprocessing, splitting the operation
as testing and training to generate the classifier to detect the Autism of children and adolescents. The proposed system uses DT,
SVM, and Random Forest models to generate a good model to predict Autism in the early stage. In classification, 10-fold
cross-validation is applied to estimate the model.

ASD
Dataset

Preprocessing
(Filling Missing Values)

Splitting operation

Testing Data Training Data

Build Models (Autism detection)

Decision Tree SVM Random

Forests

Test classification model

Autism discovery

Fig. 5 Proposed Autism discovery model

347
 Dr. R. Surendiran et al. / IJETT, 70(4), 343-359, 2022

Algorithm steps to create the decision tree in R- print("Class probability rate analysis for ASD training data
Environment set")
For creating decision tree, first load from database usin R prop.table(table(TrainData$V21)) * 100
tool as follows print("Class probability rate analysis for ASD testing data
# Load the ASD Data from datasets set")
# Attribute Relationship File Format (ARFF) data prop.table(table(TestData$V21)) * 100
ASDData <- read.csv("ASDData.arff", header = FALSE, TrainData[["V21"]] <- factor(TrainData[["V21"]])
comment.char = "@") #Matrix construction with penalty
print("All ASD data") penalty.matrix <- matrix(c(0,1,10,0), byrow=TRUE,
print(paste("Number of Rows : ",nrow(ASDData))) nrow=2)
print(paste("Number of Columns : ",ncol(ASDData))) # Building the decision tree
print(" Attribute Details") #Classification Tree construction
print("1. Answer for Question - 1") tree <- rpart(TrainData$V21~., data=TrainData, parms =
print("2. Answer for Question - 2") list(loss = penalty.matrix), method = "class")
.................................................. # visualization Tree process
.................................................... rpart.plot(tree, nn=TRUE)
print("19. Used the screening app befosinre") # ASD prdiction using the tree model
print("20. Screening Method Type") TestPredict <- predict(object=tree,TestData,type="class")
("21. Screening Score") TestPredict
View the data by the command as summary(ASDData) # Performance of the aid
Missing value analyzed using the following coding # Build Confussion Matrix
ASDData[, 13][ASDData[, 13] == '?'] <- NA confusionMatrix(table(TestPredict, TestData$V21))
ASDData[, 20][ASDData[, 20] == '?'] <- NA print(paste("Time : ",Sys.time()))
missmap(ASDData) Options (warn=0)
Next step is data cleaning process to remove inconsistent
data Figure 6 shows the workflow of the decision tree to
# Data cleaning process discover Autism in children and adults data. The input
Assign alternate values to missing values dataset is given to the classifier for modeling. Split the
<- sum(is.na(ASDData[1])) dataset using information gain. The highest information gain
(paste("Missing values in Variable 1 : ",mcount)) gives the best solution. Its result is positive, and it predicts
mcount <- sum(is.na(ASDData[2])) that the patient may have ASD based on reading the given
................ questionnaires. Otherwise patient doesn’t suffer from any
----------------- autistic spectrum disorder. Figure 7 shows the flow of the
Then data optimization process by the following command SVM classifier for autism discovery. In this model, "C “ is
ASDData <- ASDData[c(1:12,14,32,21)] the number of misclassified samples. It contains a small
# Data Factorization - Categorical data coversion for class value and no weightage for misclassification. Compute the
attribute loss function using equation 4. Figure 8 shows the autism
# Data spliting for training and testing process detection by the Random Forest classifier of the machine
DataMat <- createDataPartition(y = ASDData$V21, p = 0.5, learning model. It combines all the results and aggregate
list = FALSE) results based on majority voting.
TrainData <- ASDData[DataMat, ]
TestData <- ASDData[-DataMat, ] L=1/2||W||^2+C (4)
# Dimensions of the split
print("Class probability rate analysis for ASD data set")
prop.table(table(ASDData$V21)) * 100

348
 Dr. R. Surendiran et al. / IJETT, 70(4), 343-359, 2022

Start

Present the user with the option of answering 10 questions based on the extracted
features to be used for Autism Detection

Read the input

Pass the input to the Decision Tree classifier

Split the dataset using the Information Gain modeling

technique

Yes
Predict that patient may Is the result classified as
have Autism positive for Autism?

Display another set of Questionnaire consisting of 18 questions to the user

Read the input

Pass the input to the classifier discussed

Split the dataset using Information Gain

No

Predict that patient may be

Is the result positive for
suffering from ASD
ASD detection?

The patient mostly does not suffer from any Autistic Spectrum Disorder

End

Fig. 6 Workflow of Decision Tree classification model

349
 Dr. R. Surendiran et al. / IJETT, 70(4), 343-359, 2022

4.2 Support Vector Machine TestData <- ASDData[-DataMat, ]

# Display the dimension of the TrainData
Algorithm step to building SVM in R environment dim(TrainData)
# Load the CARET library # Display the dimension of the TestData
library('caret') dim(TestData)
# Dispaly the structure of the dataset in the DataFrame # Data Factorization - Categorical data coversion for class
str(ASDData) attribute
# Dispaly the observations # Training process for SVM classification
View(ASDData) # trctrl <- trainControl(method = "repeatedcv", number = 10,
# Display the dimension of the data set repeats=3)
print("ASD data dimension : ") trctrl <- trainControl(method = "cv", number = 10)
dim(ASDData) SVM_Linear <- train(V21 ~., data = TrainData, method =
# Displays the Summary of the ASD data values "svmLinear", trControl = trctrl, preProcess = c("center",
summary(ASDData) "scale"), tuneLength = 10)
Data preprocessing: Replacing missing values with suitable TrainData[["V21"]] <- factor(TrainData[["V21"]])
values # Autism Spectrum Disorder (ASD) prediction process with
# Data splitting process trained model
DataMat <- createDataPartition(y = ASDData$V21, p = TestPredict <- predict(SVM_Linear, newdata = TestData)
0.02, list = FALSE) TestPredict
TrainData <- ASDData[DataMat, ] Accuracy analysis process

Autism Datasets

Describe the value of C (Number of misclassified samples)

Compute the loss function using L=1/2||w||^2+C

Classify based on the decision of the soft margin SVM

Autism detection (yes/No)

Fig. 7 Generating the Model using SVM classifier

4.3Autism Spectrum Disorder Prediction using Random Forest

The Autism Spectrum Disorder (ASD) prediction framework is built to support a behavioral aspect-based analysis model
without any device. The behavioral parameters are collected with the support of the Autism Query collections. The ASD
prediction process is focused on the age boundary-based analysis process. The childhood and Adolescents analysis model is
utilized in the system. Random Forests is an ensemble classification that creates multiple decision trees during the training
period and produces individual trees by means of mode or means finding. Random Forest of this machine learning algorithm
avoids the overfitting problem in the Decision Tree.

350
 Dr. R. Surendiran et al. / IJETT, 70(4), 343-359, 2022

Step to build Random Forest

The Random Forests relate to the bagging techniques. Suppose training set W = w1 and with reply Z = z1 and selects a sample
with replacement of the training sample.
1. Sample, with replacement, n training examples from W, Z,
2. Train a classification
Data splitting process
Categorial data conversion for class attributes
Autism Spectrum Disorder (ASD) prediction process with trained model
TestPredict <- predict(RFModel, newdata = TestData)
TestPredict
# Accuracy analysis process
# Build Confussion Matrix
confusionMatrix(table(TestPredict, TestData$V21))
print(paste("Time : ",Sys.time()))
options(warn=0)

Training set (Autism datasets)

Features: 21, Instance 1,100

Training
data 1 Training
data n

Training Training
data 2 data 3

1 1 0
1

Voting
[Averaging]

Predict

Fig. 8 Autism detection using Random Forest Machine learning Model

The ML algorithms are practiced to perform the ASD prediction process. The Correlated Feature selection (CFS) based
Random Forest algorithm is applied for the disease prediction process. The prediction scheme is associated with the Naïve
Bayes and Support Vector Machine (SVM) classifiers.

4.4 Other Machine Learning Models

K- Nearest Neighbor (KNN) model used for regression and classification issues. The 'K' denotes the number of kernel
theme that is to be nominated to minimize the error. KNN is constructed on the knowledge of similarity, which can relate to
Euclidean distance, nearness, or closeness. The artificial Neural Network (ANN) model is a neural network that has a joining
with numerous neurons. Each neuron cell has a group of input values and associated weights shown in Figures 9 and 10.

351
Dr. R. Surendiran et al. / IJETT, 70(4), 343-359, 2022

Fig. 9 ANN with associate weight

Fig. 10 Architecture of ANN

352
 Dr. R. Surendiran et al. / IJETT, 70(4), 343-359, 2022

5. Experiment Result and Discussion

5.1 Environmental setup
The autism prediction models are implemented and verified using the R language. The prediction process is evaluated with
four factors. They are False Positive Rate (FPR), False Negative Rate (FNR), prediction accuracy, and prediction time
measures. The FPR rate and FNR rate parameters are used to estimate the fault level of the ASD detection process. The FPR
and the FNR are the findings of error criteria. The error stage should be reduced to increase the detection accuracy levels. The
prediction accuracy is applied to estimate the ASD discovery result accuracy levels. The processing time is calculated for the
prediction process.

5.2 Data set description

The ASD detection process is implemented to detect the autism stages in childhood, adolescence, and adulthood. The
AG10 dataset is taken from the University of California machine learning repository. The children's data set has 292 instances.
The adolescent data set has 104 instances. The adult data set is collected for 704 patients. The entire data set has 1100 instances
with 21 categorical, continuous, and binary attributes. The ML-based ASD forecast process uses the Decision Tree
Classification (DTC), SVM, and CFS-RF techniques. Table 1 and Table 2 show the ASD data set representation.
Table 1. Data set representation
S.No Name of the dataset Instances in dataset
1 Adult-ASD screening dataset 704
2 Children-ASD screening dataset 292
3 Adolescent-ASD screening dataset 104

Table 2. Features in ASD screening data

Features Features explanation
1 Age of the patient
2 Sex
3 Nationality
4 By birth patient suffered from jaundice
5 Patient family numbers anyone affected by progress disorders
6 Who is fulfilling the experiment
7 Country
8 Screening application used by the user before or not?
9 Test type
10-19 Based on the screening method, answer the questions
20 Screening Score

5.3 Measures used for evaluation

The verified criteria are used to assess the worth of a classifier. For verifying a classifier by contingency matrix, precision
and recall, Accuracy, ROC, and time analysis. The contingency matrix contains True Positive (tp), False Positive (fp), False
Positive (fp), and False Negative (fn)

• tp – Quantity of optimistic instances in the trial or experimental dataset that are categorized as positive
• fp – Quantity of undesirable instances in the trial or experimental dataset that are categorized as positive
• tn – Quantity of undesirable instances in the trial or experimental dataset that are categorized as negative
• fn – Quantity of optimistic instances in the trial or experimental dataset that are categorized as negative

Accuracy of the classifier: Accuracy is repeatedly used result metric for prediction. Obtaining more accuracy in the model in
the ML approach is the highest aspect of education (5).
𝑓𝑝+𝑓𝑛
Classification Error = (5)
𝑡𝑝+𝑡𝑛+𝑓𝑝+𝑓𝑛

353
 Dr. R. Surendiran et al. / IJETT, 70(4), 343-359, 2022

5.4 Performance Analysis

5.4.1 False Positive Rate (FPR) Analysis
The relation knows the FPR of falsely assigned positive ASD results. The FPR evaluated between the DTC, SVM, and
CFS-RF techniques are shown in Table 3 and Figure 11. The SVM-based technique decreases the FPR rate by 20% more than
the DTC technique. The CFS-RF) decreases the FPR rate by 35% more than the Support Vector Machine (SVM) technique.

Table 3. FPR Analysis between Decision Tree Classifier (DTC), SVM, and CFS-RF Techniques
Records DTC (%) SVM (%) CFS-RF (%)
220 9.48 7.92 4.74
440 8.88 7.64 4.44
660 8.36 7.14 4.18
880 7.64 6.56 3.82
1100 6.92 6.28 3.46

False Positive Rate Analysis between Decision Tree Classifier (DTC), SVM and CFS-RF
Techniques

9.48
10 7.92
False Positive Rate ( %)

8.88 7.64
8.36 7.14
8 4.74
4.44 7.64 6.56
4.18 6.28
6 3.82 6.92
3.46
DTC
4
SVM
2
CFS-RF
0
220
440
660
880
1100
Records
Fig. 11 FPR Analysis between Decision Tree Classifier (DTC), SVM, and CFS-RF Techniques

5.4.2 False Negative Rate (FNR) Analysis

The FNR is calculated as the proportion of falsely consigned negative ASD performance. The FNR analysis among the
Decision Tree Classifier (DTC), SVM, and CFS-RF techniques are shown in Table 4 and Figure 12. The SVM-based technique
decreases the FNR by 25% more than the DTC technique. The CFS-RF reduces the FNR by 30% more than the SVM
technique.
Table 4. False Negative Rate Analysis between Decision Tree Classifier (DTC), SVM, CFS-RF Techniques
Records DTC (%) SVM (%) CFS-RF (%)
220 10.04 7.56 5.02
440 9.32 7.34 4.66
660 8.61 7.04 4.32
880 7.96 6.72 3.98
1100 7.08 6.38 3.52

354
 Dr. R. Surendiran et al. / IJETT, 70(4), 343-359, 2022

False Negative Rate Analysis between Decision Tree Classifier (DTC), SVM and
CFS-RF Techniques

12
10.04
False Negative Rate (%)

10 7.56 9.32
7.34 8.61 7.04
5.02
8 4.66 7.96 6.72
4.32 6.38
3.98 7.08
6 DTC
3.52
4 SVM
CFS-RF
2

0
220
440
660
880
1100
Records
Fig. 12 False Negative Rate Analysis between Decision Tree Classifier (DTC), SVM, and CFS-RF Techniques

5.4.3 Prediction Accuracy Analysis

The prediction accuracy rate is evaluated by fetching the accuracy level of the ASD prediction process. The prediction
accuracy analysis between the Decision Tree Classifier (DTC), SVM, and CFS-RF techniques is shown in Figure 13 and Table
5. The SVM-based technique increases the prediction accuracy level by 5% more than the DTC technique. The CFS-RF
increases the prediction accuracy level by 10% more than the SVM technique. Figure 14 indicates the comparison of KNN and
ANN. The ANN produces good accuracy of 97.68%.

Table 5. Prediction Accuracy Analysis between Decision Tree Classifier (DTC), SVM, and CFS-RF Techniques
Records DTC (%) SVM (%) CFS-RF (%)
220 80.48 84.52 90.24
440 82.19 85.01 90.89
660 83.03 85.82 91.49
880 84.41 86.72 92.21
1100 86.02 87.34 93.03

355
 Dr. R. Surendiran et al. / IJETT, 70(4), 343-359, 2022

Prediction Accuracy Analysis between Decision Tree Classifier (DTC), SVM and
CFS-RF Techniques

90.24 90.89 91.49 92.21 93.03

95
Prediction Accuracy (%)

90 84.52 85.01 85.82 86.72 87.34

85 86.02
80.48 82.19 83.03 84.41 DTC
80 SVM

75 CFS-RF

70
220
440
660
880
1100
Records
Fig. 13 Prediction Accuracy Analysis between Decision Tree Classifier (DTC), SVM, and CFS-RF Techniques

Accuracy of KNN and ANN

100
97.68
98
95.76
96
94
94 93.14
91.6
92 90.5 90.98
90.03
90 88.59
88
88
86
84
82
220 440 660 880 1100
KNN (%) 88 88.59 90.03 93.14 95.76
ANN (%) 90.5 90.98 91.6 94 97.68
Axis Title

KNN (%) ANN (%)

Fig. 14 Prediction Accuracy Analysis between KNN and ANN

356
 Dr. R. Surendiran et al. / IJETT, 70(4), 343-359, 2022

Prediction Time Analysis

The prediction time analysis is applied to estimate the duration of the ASD prediction process. The prediction time
analysis between the DTC, SVM, and CFS-RF techniques is shown in Table 6 and Figure 15. The SVM-based technique
reduces the prediction time by 25% more than the DTC technique. The CFS-RF reduces the prediction time by 30% more than
the SVM technique.

Prediction Time Analysis between DTC, SVM and CFS-RF Techniques

12
11.5
Prediction Time (s)

9.9 10.4 10.8

10 9.2 8.1
7.6 8.5 8.8 9.1
5.2 5.6 5.9 6.2 6.7
8
DTC
6
SVM
4 CFS-RF

0
220
440
660
880
1100
Records

Fig. 15 Prediction Time Analysis between DTC, SVM, and CFS-RF Techniques

6. Conclusion and Future work

The Correlation Feature Selection based Random Forest model increases the accuracy level in the prediction process. The
behavioral and age group parameters are integrated into the prediction process. The data dimensionality is managed with
feature selection methods. The system solves the device dependency, technician requirements, and delay problems. The ASD
mechanism can be enhanced to analyze the adult autism levels. The ASD prediction system can be enhanced with the following
features.
• The system can be improved with automated medicine and a treatment suggestion model.
• The system can be improved to analyze the treatment and its impacts on regular patients' activities.
• The ASD prediction process can be upgraded by integrating behavioral features and EEG-based features analysis.
• The ASD prediction process can be implemented as a mobile-based application to conduct an automated self-
assessment test.

357
 Dr. R. Surendiran et al. / IJETT, 70(4), 343-359, 2022

References
[1] M. Thangamani, R. Vijayalakshmi, M. Ganthimathi, Rajita, P. Malarkodi, S. Nallusamy, Efficient classification of heart disease using
K-Means clustering Algorithm, International Journal of Engineering Trends and Technology, 68(12) (2020) 48-53.
[2] N. Suresh Kumar, M. Thangamani, V. Sasikumar S. Nallusamy, An Improved Machine Learning Approach for Predicting Ischemic
Stroke, An Improved Machine Learning Approach for Predicting Ischemic Stroke, 68(1)(2021) 111-115.
[3] Thangamani M., Ganthimathi M., Sridhar S.R., Akila M., and Keerthana R. Engineering, Detecting Coronavirus Contact using Internet
of things., International Journal of Pervasive Computing and Communications, 16(5) (2020) 447-456. https://doi.org/10.1108/IJPCC-
07-2020-0074.
[4] Dilantha Haputhanthri, Gunavaran Brihadiswaran, Sahan Gunathilaka, Dulani Meedeniya, and Sampath Jayarathna, An EEG based
Channel Optimized Classification Approach for Autism Spectrum Disorder, Moratuwa Engineering Research Conference (Mercon),
(2019).
[5] Surendiran, R., and Alagarsamy,K., 2010. Skin Detection Based Cryptography in Steganography (SDBCS) .International Journal of
Computer Science and Information Technologies (IJCSIT), 1(4) (2010) 221-225.
[6] Supriya H S, Sumanth Alva R*, Suprith K P, Vikas Kumar L, and P Kyshan Neheeth, applying supervised learning technique to
diagnose autism spectrum disorder (asd), International Journal of Recent Scientific Research , 11(5b) (2020) 38439-38441.
DOI: http://dx.doi.org/10.24327/ijrsr.2020.1105.5312.
[7] Inon Wiratsin, L. Narupiyakul, Feature Selection Technique for Autism Spectrum Disorder, Proceedings of the 5th International
Conference on Control Engineering and Artificial Intelligence, DOI, 10.1145/3448218.3448241, (2021).
[8] Surendiran, R., 2017. Development of Multi-Criteria Recommender System, SSRG International Journal of Economics and
Management Studies (IJEMS) ISSN: 2393 - 9125, 4(1) (2017) 28-33.
[9] Daniel Bone, Somer Bishop, Matthew P. Black, and Shrikanth S. Narayanan, Use of machine learning to improve autism screening and
diagnostic instruments: effectiveness, efficiency, and multi-instrument fusion, Journal of Child Psychology and Psychiatry, (2016).
[10] Lucia Billeci, Antonio Narzisi, Alessandro Tonacci, Beatrice Sbriscia-Fioretti, Luca Serasini, Francesca Fulceri, Fabio Apicella,
Federico Sicca, Sara Calderoni & Filippo Muratori An integrated EEG and eye-tracking approach for the study of responding and
initiating joint attention in Autism Spectrum Disorders, Scientific Reports, 7 (2017) 1-13.
[11] Filippo Muratori, Lucia Billeci, Sara Calderoni, Maria Boncoddo, Caterina Lattarulo, Valeria Costanzo, Marco Turi, Costanza Colombi
and Antonio Narzisi, How Attention to Faces and Objects Changes Over Time in Toddlers with Autism Spectrum Disorders:
Preliminary Evidence from An Eye-Tracking Study, Brain science, MDPI, 9(344) (2019) 1-11.doi:10.3390/brainsci9120344.
[12] S. Jayarathna, Y. Jayawardana, M. Jaime, and S. Thapaliya, Electroencephalogram (EEG) for Delineating Objective Measure of Autism
Spectrum Disorder, in Computational Models for Biomedical Reasoning and Problem Solving, IGI Global, (2019) 34-65.
[13] S. Thapaliya, S. Jayarathna, and M. Jaime, Evaluating the EEG and Eye Movements for Autism Spectrum Disorder, in IEEE
International Conference on Big Data (Big Data), (2018) 2328-2336.
[14] Surendiran, R., Similarity Matrix Approach in Web Clustering .Journal of Applied Science and Computations (JASC),5(1) (2018) 267-
272.
[15] Narzisi, A. Posada, M. Barbieri, F. Chericoni, N. Ciuffolini, D. Pinzino, M. Romano, R. Scattoni, M.L.Tancredi, R. Calderoni, S.,
Prevalence of autism spectrum disorder in a large Italian catchment area: A school-based population study within the ASDEU project.
Epidemiol. Psychiatr. Sci. (2018).
[16] E. Grossi, C. Olivieri, and M. Buscema, Diagnosis of autism through EEG processed by advanced computational algorithms: A pilot
study, Comput. Methods Programs Biomed., 142 (2017) 73-79.
[17] W. J. Bosl, H. Tager-Flusberg, and C. A. Nelson, EEG Analytics for Early Detection of Autism Spectrum Disorder: A data-driven
approach, Scientific Reports, 8(1) (2018) 1-20.
[18] Zhong Zhao, Xiaobin Zhang, Wenzhou Li, Xinyao Hu, Xingda Qu, Xiaolan Cao, Yanru Liu, and Jianping Lu, Applying Machine
Learning to Identify Autism With Restricted Kinematic Features, IEEE Access, (2019).
[19] J. Baio et al., Prevalence of Autism Spectrum Disorder Among Children Aged 8 Years - Autism and Developmental Disabilities
Monitoring Network, 11 Sites, United States, 2014, MMWR. Surveillance Summaries, 67(6) (2018) 1-23.
[20] Surendiran, R., Rajan, K.P. and Sathish Kumar, M., Study on the Customer targeting using Association Rule Mining. International
Journal on Computer Science and Engineering, 2(7) (2010) 2483-2484.
[21] Charlotte Küpper, Sanna Stroth, Nicole Wolff, Florian Hauck, Tanja Schad-Hansjosten, Luise poustka, Veit Roessner, Katharina
Schultebraucks & Stefan Roepke, identifying predictive features of autism spectrum disorders in a clinical sample of adolescents and
[22] Adults using machine learning, Scientific report, nature research, 10(4805) (2020). doi.org/10.1038/s41598-020-61607-w.
[23] NaziaJassim, Simon Baron-Cohen, and John Suckling, Meta-analytic evidence of differential prefrontal and early sensory cortex activity
during non-social sensory perception in Autism, Neuroscience and Biobehavioural Reviews, Elsevier, 127 (2021) 146-157.
[24] Suman Raj and Sarfaraz Masood, Analysis and Detection of Autism Spectrum Disorder Using Machine Learning Techniques, Procedia
Computer Science, Elsevier, 167 (2020) 994-1004.
[25] Judith Charpentier, Marianne Latinus, Frederic Andersson, Agathe Saby, Jean-Philippe Cottier, Frederique Bonnet-Brilhault,
EmmanuelleHouy-Durand, and Marie Gomot, Brain Correlates of Emotional Prosodic Change Detection in Autism Spectrum Disorder,
Neuro Image: Clinical, Elsevier, 28 (2020) 102512.
[26] Surendiran, R. and Alagarsamy, K., PCA-based geometric modeling for automatic face detection. Int. J. Comput. Sci. Inform.
Technol, 1(4) (2010) 221-225.
[27] Jennie Hayes, Tamsin Ford, Hateem Rafeeque, Ginny Russell, Clinical practice guidelines for the Diagnosis of Autism Spectrum
Disorder in Adults and Children in The UK: a narrative review, BMC Psychiatry, 13(18) (2018) doi: 10.1186/s12888-018-1800-1.

358
 Dr. R. Surendiran et al. / IJETT, 70(4), 343-359, 2022

[28] Mengyi Liao, Hengyao Duan, and Guangshuai Wang, Application of Machine Learning Techniques to Detect the Children with Autism
Spectrum Disorder, Hindawi Journal of Healthcare Engineering, 2022 (2022) 1-10,
Article ID 9340027.https://doi.org/10.1155/2022/9340027.
[29] Chelsea M. Parlett-Pelleriti, Elizabeth Stevens, Dennis Dixon, Erik J. Linstead, Applications of Unsupervised Machine Learning in
Autism Spectrum Disorder Research: a Review, Journal of Autism and Developmental Disorders, (2022) 1-17.
https://doi.org/10.1007/s40489-021-00299.
[30] Cooper J. Mellema 1,2, Kevin P. Nguyen 1,2, Alex Treacher 1 & Albert Montillo, Reproducible neuroimaging features for diagnosis of
Autism Spectrum Disorder with Machine Learning, Scientific Reports, 0123456789 (2022) 1-13.https://doi.org/10.1038/s41598-022-
06459-2.
[31] Yinghan Zhu, Hironori Nakatani, Walid Yassin, Norihide Maikusa, Naohiro Okada, Akira Kunimatsu, Osamu Abe, Hitoshi Kuwabara,
Hidenori Yamasue, Kiyoto Kasai, Kazuo Okanoya, and Shinsuke Koike, Application of a Machine Learning Algorithm for Structural
Brain Images in Chronic Schizophrenia to Earlier Clinical Stages of Psychosis and Autism Spectrum Disorder: A Multiprotocol Imaging
Dataset Study, Schizophrenia Bulletin, 10 (2022) 1-12.
[32] N. Sciaraffa, G. Borghini, P. Aricò, Joint Analysis of Eye Blinks and Brain Activity to Investigate Attentional Demand during a Visual
Search Task, Brain Science, MDPI, 11(562) (20210) 1-20. DOI:10.3390/brainsci11050562.
[33] Bram van den Bekerom, Using Machine Learning for Detection of Autism Spectrum Disorder, IEEE, (2017).
[34] Fadi Thabtah, Autism Spectrum Disorder Screening: Machine Learning Adaptation and DSM-5 Fulfillment, ACM, (2017).
[35] Rasool Azeem Musa, Mehdi Ebadi Manaa2, and Ghassan Abdul-Majeed Predicting Autism Spectrum Disorder (ASD) for Toddlers and
Children Using Data Mining Techniques, Journal of Physics: Conference Series, IOP publishing,
doi:10.1088/1742-6596/1804/1/012089, 1804 (2021).
[36] Antoine Frigaux ,Joëlle Lighezzolo-Alnot , Jean-Claude Maleval and Renaud Evrard, Differential diagnosis on the Autism Spectrum:
Theorizing an Ordinary Autism, L’evolution ´ psychiatrique, Elsevier, 86(2021) e1–e24.

359