2022 5th International Conference on Contemporary Computing and Informatics (IC3I)

DL-ASD: A Deep Learning Approach for Autism

Spectrum Disorder
2022 5th International Conference on Contemporary Computing and Informatics (IC3I) | 979-8-3503-9826-7/22/$31.00 ©2022 IEEE | DOI: 10.1109/IC3I56241.2022.10072429

Ruchi Mittal Varun Malik* Ajay Rana

Chitkara University Institute of Chitkara University Institute of Engineering and Technology
Engineering and Technology Engineering and Technology Amity University
Chitkara University Chitkara University Uttar Pardesh, India
Punjab, India Punjab, India [email protected]
[email protected] [email protected]

Abstract—Identifying a person's feelings and sentiments is known creating mathematical models that can predict the class to
as emotion recognition and analysis. The emotion analysis approach which that point belongs, given a data point [8]. Sparse data
correctly recognizes normal people's facial emotions in the first may be used by machine learning classifiers to properly
attempt. Children with Autism Spectrum Disorder (ASD) who have represent complicated and high-dimensional input variables. As
trouble talking or expressing themselves can struggle emotionally to
they are computationally efficient and work well with small
understand. To predict ASD and No ASD in children aged 1-10 using
dynamic analysis, this work presents a robust deep learning model sample sizes, the standard supervised machine learning
with multi-label categorization. We proposed a DL-ASD framework methods were utilised to detect aberrant motion and postural
for identifying autism spectrum disorder. The proposed model has aspects [9]. Discriminant analysis, K-nearest neighbour, naive
used the Kaggle dataset as an image dataset. The datasets are trained Bayes, decision trees, support vector machines, and random
with an Improved Convolutional Neural Network (I-CNN), and the forests are some examples of well-known methods. No method
images are used to classify individuals as having autism spectrum is ideal for all datasets because of differences in characteristics,
disorder or not having ASD. Feature-based calculations of internal sample sizes, and data formats [10]. Researchers typically
and exterior distances are used to identify the emotion. Optimization conduct multiple experiments to determine the most efficient
procedures such as dropout, batch normalization, and parameter
supervised machine-learning algorithm for classification or
update are used to optimize the Improved Convolutional Neural
Network's (I-CNN) processing of the returning facial landmarks. prediction [11].
The proposed method correctly predicts six emotions in addition to In face detection, a rectangle is drawn around recognizable
four general emotions. According to the experimental results, the facial features [12]. This strategy employs the haar cascade
classification accuracy of the approach proposed in this study can classifier in conjunction with the viola jones face detection
reach 98%. technique. One definition of face recognition is "the ability to
Keywords—Autism Spectrum Disorder, Deep learning, DL-ASD, identify an individual in a photograph [13]. The frontal face
improved CNNKeywords— E-commerce, E-commerce products, prediction method uses a face's landmarks throughout face
IMEP, Social Media, IM, UCR recognition. The Dlib module of OpenCV includes 68 landmark
I. INTRODUCTION predictors to help with face recognition, feature extraction, and
emotion detection. "multi-label classification" describes
Characteristics of social communication in youngsters with
assigning labels to data or objects according to multiple sets of
ASD are more difficult to identify than in other conditions [1].
classification results [14]. Each label represents a unique set of
Because there is no accurate ASD diagnosis technique, these
dependent attributes and many distinct classes. Recent studies
features are crucial in diagnosing autism spectrum disorder.
have shown that a child's facial expressions, traits, and features
Emotions, arousal, and action units on the face are also used to
can be used to predict autism spectrum disorder and gain insight
diagnose ASD [2].
into the child's behaviour. Artificial intelligence, machine
In 2017, the face emotion image datasets EmotioNet and
learning, and deep learning assist autistic studies [15].
AffectNet were released [3][4]. CNN can learn and predict
Examined the condition and behaviour of ASD children using
emotions using these massive datasets. ASD lacks large-scale
facial landmarks, body position, and auditory and biosignals.
datasets, making it difficult to train CNN specifically with ASD
The main contribution of this paper is the improved CNN
data [5]. As a result, the dataset used in many ASD studies
algorithm for training and classification.
comes from autistic centres and physicians and is also restricted
The rest of this paper is organized as follows. In Section 2,
to sponsored studies. Gerry spent a year collecting 1857 ASD
several writers discuss various ASD diagnostic techniques.
and 1850 Typical Development (TD) photos from various
Section 3 depicts the DL-ASD model. Section 4 deals with
online sources and uploading them to Kaggle.com. He is the
implementation, while Section 5 summarises the study's
dataset's creator [6].
findings. Section 6 finishes with a discussion of the findings
Computers may use machine learning, a form of artificial
and future directions.
intelligence, to automatically search vast databases for patterns
and generate conclusions based on such searches [7].
Classification is possible in supervised machine learning by

1767
979-8-3503-9826-7/22/$31.00 ©2022 IEEE

Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on June 25,2024 at 07:20:00 UTC from IEEE Xplore. Restrictions apply.
2022 5th International Conference on Contemporary Computing and Informatics (IC3I)
II. BACKGROUND STUDY that make up a CNN are as follows: Input, Convolution,
C. J. Brown et al. [1] for neural networks to acquire Pooling, Fully Connected, and Classification Output.
knowledge of edge-specific weighting, the author devised a A layer of Convergence CNN relies on its foundational
unique element-wise layer that included connective and data- convolution layer. When you shift these layers, tiny filters
specific priors. Adding this layer to a deep network architecture appear across the entire display. The dot product between the
tailored to brain network data improved the accuracy of autism image and the filter is calculated to perform convolution. The
spectrum disorder (ASD) prediction on 1013 functional filter region is modified by applying the Dot Product of each
connectives from the autistic brain imaging data exchange image pixel with the corresponding filter. After that, we move
(ABIDE) dataset significantly. the filter onto the next available period. All of the pictures are
D. Berardini et al. [3] the author shows a vast video hidden as you stride. Equation 2 might also help you make
collection of children's social interaction, painstakingly sorted sense of the convolution procedure. Here is a mathematical
into clinically significant behaviour for ASD diagnosis. A explanation of the convolution operation between an image (xl)
machine learning framework has been given for data collection. and a filter vector:
This research studies the use of feature selection, class = ∗ + ------ (1)
rebalancing, and neural network classifiers for inferring autism It's responsible for down-sampling tasks. Some different
spectrum illness from behavioural data. Self-supervised pooling functions exist. Maximum pooling is the most popular
systems are being developed to recognise an adult's head and utility. A picture is given a 2*2 filter with a stride of 2, from
things of interest in image frames, boosting look-face, look- which the pooling procedure may be inferred. For each sub-
object, and behaviour detection. region, the maximum pooling filter provides the greatest
M. Presecan et al. [7] In this study, the author introduces possible value. As a result, a feature of size (4 * 4 * 1) gets
the Faster R-CNN model for object recognition. The Faster R- down-sampled to size (6 * 6 * 1) when the greatest pooling filter
CNN has been widely used despite the fact that faster R-CNN of size (6 * 6 * 1) is applied to it. The output of the pooling
models have been shown to be a high-performance, accurate, layer may be characterized as:
and quick solution for recognising many objects within an In a completely interconnected layer, all of the neurons
image. from the previous layer are linked to those of the layer below it.
O. Rudovic et al. [8] Children with Autism Spectrum = ∗ + ------- (2)
Conditions (ASCs) who participated in a single session of
Where u represents the activation function (a sigmoid), b2j
robot-assisted autism therapy were photographed for this
is the bias, and u is the input node, we weigh the ith input node
research. The author examined several deep-learning settings
and the jth hidden node, denoted by w2ij. For input at Layer 1,
designed for automated engagement evaluation.
we have zl1j.
Q. Mohi-ud-Din and A. K. Jayanthy [10] this author
Every neuron in a fully connected layer can access the data
employed a transfer learning mechanism to classify ASD from the layer below it. As a result, several variables are related
patients and controls based on EEG data. The findings indicated
to training (weight). However, most hidden neurons are
a good degree of accuracy.
activated to a small degree. So that one hidden node, in
S. Sadiq et al. [12] Autism was becoming more frequent in
particular, may facilitate deep learning, a low activation value
the United States, with a corresponding rise in its prevalence
of neurons is required. Activation of neurons may be regulated
among ethnic minorities and low-resource communities. It
by adding Sparsity.
wasn't until the average child was four years old that doctors Classification Loss is calculated in this layer as part of a
began diagnosing the condition. This study provided an
CNN training process. For reliable data prediction, CNN relies
independent machine learning-based approach for identifying
on the minimization of a cost function (existing), which serves
acoustic regimes statistically predicting scores on the Autism
as its objective function.
Spectrum Disorder Social Affect Scale.
Eq gives the existing cost function.
III. PROPOSED METHOD , = + ∑ ------ (3)
Autism spectrum disorder has been classified with an The cross-entropy loss is
%
improved CNN algorithm. The datasets are collected from the = − ∑&' " #$ ------ (4)
Kaggle dataset. yP represents the prediction, m represents the training data,
a) Dataset and it represents the desired outcome. The L2 regularisation
Datasets are downloaded from parameters b are defined by the cross-entropy loss function in
https://www.kaggle.com/datasets/cihan063/autism-image- Eq.
data. The dataset has a test, train, test, valid and consolidated Overfitting is a potential problem due to Improved CNN's
memory size of 240MB. many hidden parameters. As previously established, the
greatest number of dormant neurons leads to the weakest
b) Improved Convolution neural network (I-CNN) possible neural activity. A concealed node's average activation
Multiple varieties of deep neural networks (DNNs) are value (lTJ) should be close to zero.
employed in image processing, and CNNs are just one of them. With the old function used to calculate the divergence of
In terms of data processing and retrieval, it is distinct from activation in Eq. the activation in the hidden layer of a
conventional neural networks. The CNN architecture convolutional neural network (CNN) may be lowered. By
represented by Equation 1 is quite popular. The many layers
1768

Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on June 25,2024 at 07:20:00 UTC from IEEE Xplore. Restrictions apply.
2022 5th International Conference on Contemporary Computing and Informatics (IC3I)
deactivating unused training parameters, Sparsity practice
reduces the total number of them, which helps prevent over-
fitting in CNN.

c) Algorithm: Improved CNN algorithm

Input: Regulation parameter: sample image x 2 RD,
Learning Rate (bÞ;
Sparsity is a parameter (b1), Sparsity (l); Iteration number (N).
Output: Parameters of Weight and Bias wð1Þ ij ;wð2Þ ij ; bð1Þ
j ; bð2Þ j
Step 1: The weight and bias parameters should be initialized.
For n = 1 to N, do
Step 2: Using Eq, compute the reconstruction result.
Step 3: The modified cost function equation directs you to
perform the following actions.
Step 4: Weight and bias parameters are updated using a gradient
method
Step 5: Steps 2 and 4 should be repeated until n = N.
End

IV. RESULTS AND DISCUSSION Figure 2: Classified image data

We proposed that the DL-ASD framework be implemented
using python programming with 3.8 versions. The proposed The dataset has been classified using the I-CNN algorithm
framework has achieved 98% accuracy. with autistic and non_autistic images shown in figure 2.

Figure 1: Classified with autistic and non-autistic

The number of images counted with non-autistic and

autistic has displayed in figure 1. The label name is represented
on the X-axis, and the count is represented on the Y-axis. Figure 3: Training and testing values
The training and testing with 20 epochs are displayed in figure
3. The Loss Plot and Accuracy Plot are shown in figure 3.

1769

Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on June 25,2024 at 07:20:00 UTC from IEEE Xplore. Restrictions apply.
2022 5th International Conference on Contemporary Computing and Informatics (IC3I)
characteristics may be used to predict their behaviour and the
likelihood that they may acquire autism spectrum disorder. To
further improve the accuracy with the use of hybrid deep
learning algorithms.
REFERENCES
[1] C. J. Brown, J. Kawahara, and G. Hamarneh, “Connectome priors in deep
neural networks to predict autism,” in 2018 IEEE 15th International
Symposium on Biomedical Imaging (ISBI 2018), 2018.
[2] C. Wu et al., “Machine learning based autism spectrum disorder detection
from videos,” in 2020 IEEE International Conference on E-health
Networking, Application & Services (HEALTHCOM), 2021.
[3] D. Berardini, L. Migliorelli, S. Moccia, M. Naldini, G. D. Angelis, and E.
Frontoni, “Evaluating the autonomy of children with autism spectrum
disorder in washing hands: a deep-learning approach,” in 2020 IEEE
Symposium on Computers and Communications (ISCC), 2020.
[4] J. Han, G. Jiang, G. Ouyang, and X. Li, “A multimodal approach for
identifying autism spectrum disorders in children,” IEEE Trans. Neural Syst.
Figure 4: Confusion matrix Rehabil. Eng., vol. 30, pp. 2003–2011, 2022.
[5] M. Lai et al., “A machine learning approach for retinal images analysis as an
The confusion matrix of the proposed work has shown in objective screening method for children with autism spectrum disorder,”
EClinicalMedicine, vol. 28, no. 100588, p. 100588, 2020.
figure 4. The TP is 125, FP is 143, TN is 25, and FN is 7. [6] Y. Li, M. A. Mache, and T. A. Todd, “Automated identification of postural
control for children with autism spectrum disorder using a machine learning
approach,” J. Biomech., vol. 113, no. 110073, p. 110073, 2020.
[7] M. Presecan, F. Petric, and Z. Kovacic, “Object classification for child
behavior observation in the context of autism diagnostics using a deep
learning-based approach,” in 2018 26th International Conference on
Software, Telecommunications and Computer Networks (SoftCOM), 2018.
[8] O. Rudovic et al., “CultureNet: A deep learning approach for engagement
intensity estimation from face images of children with autism,” in 2018
IEEE/RSJ International Conference on Intelligent Robots and Systems
(IROS), 2018.
[9] P. K. C. Prasad, Y. Khare, K. Dadi, P. K. Vinod, and B. R. Surampudi, “Deep
learning approach for classification and interpretation of autism spectrum
disorder,” in 2022 International Joint Conference on Neural Networks
(IJCNN), 2022.
[10] Q. Mohi-ud-Din and A. K. Jayanthy, “Detection of Autism Spectrum
Disorder from EEG signals using pre-trained deep convolution neural
networks,” in 2021 Seventh International conference on Bio Signals, Images,
and Instrumentation (ICBSII), 2021.
[11] S. G. Karuppasamy, D. Muralitharan, S. Gowr, S. R. Arumugam, E. A. Devi,
Figure 5: Predictive result and K. Maharajan, “Prediction of autism spectrum disorder using convolution
neural network,” in 2022 6th International Conference on Trends in
The predictive analysis has shown in figure 5. The I-CNN Electronics and Informatics (ICOEI), 2022.
has predicted the exact result. [12] S. Sadiq, M. Castellanos, J. Moffitt, M.-L. Shyu, L. Perry, and D. Messinger,
“Deep learning based multimedia data mining for autism spectrum disorder
(ASD) diagnosis,” in 2019 International Conference on Data Mining
V. CONCLUSION Workshops (ICDMW), 2019.
Here, we provided the DL-ASD framework for detecting [13] V. Zope, T. Shetty, M. Dandekar, A. Devnani, and P. Meghrajani, “ML based
ASD. The first step in this approach is to identify problematic approaches for detection and development of autism spectrum disorder: A
review,” in 2022 International Conference on Sustainable Computing and
behaviours, and the second is to use these same statistical Data Communication Systems (ICSCDS), 2022.
behaviour characteristics to predict autism spectrum disorder [14] Y. Wang, J. Wang, F.-X. Wu, R. Hayrat, and J. Liu, “AIMAFE: Autism
(ASD). The deep-learning baselines were presented for spectrum disorder identification with multi-atlas deep feature representation
and ensemble learning,” J. Neurosci. Methods, vol. 343, no. 108840, p.
behaviour recognition: two different models, one trained solely
108840, 2020.
on the young subject's facial features and the other on raw video [15] L. Xu, Z. Sun, J. Xie, J. Yu, J. Li, and J. Wang, “Identification of autism
frames. Eventually, we want to develop self-supervised spectrum disorder based on short-term spontaneous hemodynamic
algorithms that recognize the faces of adults and the objects of fluctuations using deep learning in a multi-layer neural network,” Clin.
Neurophysiol., vol. 132, no. 2, pp. 457–468, 2021.
interest in photographs, improving look-face and look-object
[16] B. Mathew, R. Narayanan, and A. Mittal, “A review of the perception of
identification. An Improved CNN model is suggested for healthcare quality among patients and employees,” Indian j. public health res.
ASD/NoASD children's ASD/NoASD and facial expression dev., vol. 10, no. 9, p. 452, 2019.
prediction. The efficiency and dependability of the suggested [17] S. Shamas, S. N. Panda, and I. Sharma, “Review on lung nodule
model are enhanced. An example application would be segmentation-based lung cancer classification using machine learning
approaches,” in Artificial Intelligence on Medical Data, Singapore: Springer
analyzing facial photos of autistic children to deduce their Nature Singapore, 2023, pp. 277–286.
action units, arousal levels, and valence. Children's facial

1770

Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on June 25,2024 at 07:20:00 UTC from IEEE Xplore. Restrictions apply.