Autism Spectrum Disorder (ASD) is a neurodevelopmental condition that afects language, communication, cognitive, and social skills. Early detection of ASD in children is crucial for efective intervention, and machine learning techniques have emerged as promising tools to improve the accuracy and eficiency of detection. This paper presents a range of Machine Learning approaches that have been applied to identify individuals with ASD, with a particular focus on children, using images as input data. The results of these studies demonstrate the potential for Machine Learning to aid in the early detection and diagnosis of ASD in children, which can lead to better outcomes for individuals with this condition.
that may have similar symptoms to ASD. The incidence mon methods for autism detection include cognitive tests,
of ASD traits attributed to genetic factors is estimated questionnaires based on autism symptoms, video
recordaround 81% of the population, moreover environmental ing analysis, social interaction analysis, and brain
scanfactors have been associated in literature with about 14% ning. Cognitive tests can be used to evaluate memory,
to 22% of the risk of ASD[
It’s important to note that the diagnostic process for behavior patterns, and can be used to identify potential
ASD can be complex and may require multiple evalua- indicators of autism in young children. Social
interactions over time. Additionally, not all individuals with tion analysis involves observing the interactions between
ASD may be diagnosed in early childhood, and some may young children and their caregivers or peers and can
proreceive a diagnosis later in life. Autism, also known as vide information on the child’s social communication
Autism Spectrum Disorder (ASD), is a neurodevelopmen- abilities and potential indicators of autism. Finally, brain
tal disorder that afects how people interact with others, scanning techniques such as functional magnetic
resocommunicate, learn, and behave. nance imaging (fMRI) and electroencephalography (EEG)
Autism is a complex and lifelong condition that is typ- can provide insights into the neural mechanisms
underlyically diagnosed in early childhood. Individuals with ing autism and may be used to identify early indicators of
autism may have dificulty with verbal and nonverbal the disorder. There are also methods based on machine
communication, such as making eye contact, using facial learning (ML) that can analyze large amounts of data
expressions, and understanding body language. They and identify patterns that may be dificult to detect using
may also struggle to initiate and maintain social relation- traditional methods [
Natural Language Processing (NLP), a field of study that
focuses on the interactions between computers and
human languages. NLP has been used to analyze language 3. Implementation
patterns in individuals with autism, including their use
of pronouns, repetition, and lexical diversity; Transfer In this research project, we have employed diferent
maLearning; a technique that involves training a machine chine learning models to classify subjects either as having
learning model on a large dataset and then transferring ASD (autism spectrum disorder) or not. The performance
the learned features to a smaller dataset. Transfer learn- of each classifier was then evaluated to identify the best
ing has been used to develop models for early detection of model.
autism using electroencephalography (EEG) data.
Overall, machine learning approaches hold great promise for 3.1. Dataset
improving the understanding, diagnosis, and treatment The Dataset [25] used in this study included 3,014 facial
of autism spectrum disorder. However, further research is images of children with autism and typically developing
needed to ensure that these techniques are reliable, valid, children, sourced from the publicly accessible Kaggle
platand scalable in real-world settings. Several analyzes of form. Half of the images were of children with autism,
ML methods for disease detection are recently proposed. while the other half were of non-autistic children. The
In 2015 machine learning algorithms were used [
Regarding the choice of characteristic features that should be mentioned Vaishali et al. [22] discuss the use 3.2. Machine Learning methods of machine learning-based behavioral analytics for de- We used a variety of classification algorithms from Scikittecting the risk of autism. The study experiments with learn, a popular Python Machine Learning library. The a dataset of 21 features obtained from the UCI machine algorithms used include Decision Tree, Random Forest, learning repository using a swarm intelligence-based bi- Support Vector Machines (SVM), and k-Nearest Neighnary firefly feature selection wrapper. The experiment bors (k-NN). ifnds that 10 features are suficient to distinguish between ASD and non-ASD patients and achieve an average accuracy range of 92.12%-97.95% with the optimum feature 3.2.1. Decision Tree subset, which is comparable to the accuracy produced by Decision Trees are a type of non-parametric supervised the entire ASD diagnosis dataset. learning method that is used for classification and regres
Another study by Thabtah et al. [23] proposed a new sion. The goal is to build a model that predicts the value machine learning technique called Rules-Machine Learn- of a target variable using simple decision rules derived ing (RML) that ofers a knowledge base of rules to under- from data features. A tree is an example of a piecewise stand the underlying reasons behind the classification of constant approximation. A decision tree is a type of maautism spectrum disorder (ASD) traits. The RML model chine learning algorithm that is used for classification and regression analysis. It is a graphical representation regression problem, the algorithm predicts a continuous of all possible solutions to a decision based on certain output variable based on a set of input features. Random conditions or input variables. In other words, a decision Forests are commonly used in a variety of applications, tree is a tree-like structure that represents a set of deci- such as image classification, customer churn prediction, sions and their possible consequences. In a decision tree, and financial forecasting. They are particularly useful each node represents a decision point based on a certain when dealing with large and complex datasets, where feature or attribute. The branches that emanate from the traditional statistical methods may not be suficient. node represent the possible outcomes or choices that can be made based on the decision point. At each subsequent 3.2.3. Support Vector Machine (SVM) node, the algorithm evaluates the available options and makes a decision based on the best available information. SVM is an algorithm for classification tasks that works For example, suppose we are trying to predict whether by finding the optimal hyperplane that separates the data a person will purchase a product based on their age and points of diferent classes. In image classification, SVM income. The decision tree algorithm would first split the can be used to separate the diferent classes of images data based on age into two groups: those under a certain based on their feature values. A Support Vector Machine age and those over it. The algorithm would then evaluate (SVM) is a type of machine learning algorithm used for the income of each group and determine which group is classification and regression analysis. SVMs are based on more likely to purchase the product based on that crite- the concept of finding a hyperplane that best separates rion. The algorithm would then split that group further data into diferent classes. In other words, an SVM tries based on other available features or attributes and con- to find the best possible boundary that separates difertinue to evaluate and split the data until a final decision ent classes of data, such that the distance between the is reached. Decision trees can be useful for a variety of boundary and the closest data points (known as support machine learning tasks, such as prediction, classification, vectors) is maximized. The SVM algorithm works by first and feature selection. They are often used in decision mapping the input data to a high-dimensional feature support systems, customer segmentation, and marketing space, where it becomes easier to separate the diferent analytics. classes. The algorithm then finds the hyperplane that maximizes the margin between the support vectors of 3.2.2. Random Forest Classifier the diferent classes. The margin is the distance between the hyperplane and the closest data points, and the support vectors are the data points that are closest to the hyperplane. SVMs can be used for both linear and nonlinear classification and regression problems. In linear SVMs, the data can be separated using a straight line or a hyperplane, while in nonlinear SVMs, the data can be separated using more complex curves or surfaces. SVMs are popular in machine learning because they have a strong theoretical foundation and have been shown to perform well on a variety of datasets. They are particularly useful when dealing with datasets that have a large number of features, as SVMs can handle high-dimensional data with relative ease. SVMs have been used in a wide range of applications, including text classification, image classification, and bioinformatics.
This is a learning method that can be used for classification as well as regression. It works in the context of image classification by constructing multiple decision trees, each of which employs a subset of the features and samples. The final prediction is then made by aggregating all of the trees’ predictions. A Random Forest is a machine learning algorithm that is based on decision trees. It is an ensemble learning method that combines multiple decision trees to make a more accurate prediction. The algorithm works by creating a large number of decision trees, each of which is trained on a diferent subset of the data and a random subset of the available features. The algorithm then combines the predictions of each tree to arrive at a final prediction. Random Forests are known for their high accuracy and robustness against overfitting, a common problem in decision trees. By using multiple decision trees and aggregating their predictions, a Random Forest is able to reduce the impact of individual trees that may be overfitting the data. Additionally, the random selection of features at each node helps to increase the diversity of the trees and reduce correlation between them, further reducing the risk of overfitting. The Random Forest algorithm can be used for both classification and regression problems. In a classification problem, the algorithm predicts the class or category of a sample based on a set of input features. In a 3.2.4. k-Nearest Neighbors K-Nearest Neighbor (kNN) is a supervised classification algorithm used for predicting the class of a new data point. The basic idea behind kNN is to find the k closest data points in the training set to the new data point and assign the most common class among these neighbors to the new one. k-Nearest Neighbors (k-NN) is a machine learning algorithm used for classification and regression analysis. The algorithm works by finding the k data points in the training set that are closest to a given data 3.2.5. Results Classifier Decision Trees Random Forests k-Nearest Neighbors Support Vector Machines
learning methods (Table 2). point in the test set, and then using the labels of these be extremely efective at image recognition tasks and to nearest neighbors to make a prediction for the test point. be able of learning complex features from images. We In k-NN classification, the algorithm predicts the class trained and evaluated the CNN on the same dataset that of a test sample by identifying the k nearest neighbors we used to train and evaluate the other Machine Learnin the training set and assigning the class that is most ing algorithms, and we compared its performance to the common among them to the test sample. The value of k is other methods. With the addition of a CNN, we can inusually chosen by the user and can be a hyperparameter vestigate the advantages and disadvantages of using deep tuned through cross-validation. In k-NN regression, the learning methods for autism detection. algorithm predicts the value of a continuous variable for a test sample by identifying the k nearest neighbors 3.3.1. Architecture in the training set and taking the average (or weighted average) of their values. The k-NN algorithm is simple and easy to understand, and can be applied to a variety of problems with diferent types of data. However, it can be computationally expensive, especially when the training set is large. Additionally, k-NN can be sensitive to the choice of distance metric used to measure the similarity between data points. k-NN has been used in a wide range of applications, including image recognition, recommendation systems, and anomaly detection. It is particularly useful when the underlying distribution of the data is not well known or when the decision boundary is complex and nonlinear.
Our Convolutional Neural Network (CNN) consists of an input layer with images of size 224x224 and 3 RGB channels. It has 7 convolutional layers with 64 filters of size 3x3 and ReLU activation function. Batch normalization layers follow each convolutional layer to normalize the output of feature maps. Max pooling layers of size 2x2 reduce the dimension of feature maps and preserve salient features. The final Max Pooling Layer is followed by a Flatten Layer that converts the feature map into a one-dimensional vector. The network has two fully connected layers with 128 units and ReLU activation function in the first layer, and a single unit with a sigmoid activation function in the output layer. A Dropout Layer with a dropout rate of 50% is added between the fully connected layers to prevent overfitting. The output layer has a single unit with a sigmoid activation function. The model is compiled with binary cross-entropy loss, Adam optimizer, and accuracy as the metric.
The model is trained for 1000 epochs with a batch size of 5, and its performance is evaluated on the validation set (x_test, y_test). Early stopping is implemented to prevent overfitting, where the model training is stopped if the validation loss does not improve for a certain number of epochs.
classifier outperforms the others in terms of performance. Below are the results obtained from CNN (Table 3). Notably, Random Forests have attained the highest F1score, accuracy, and precision. On the other hand, k- Method Training Accuracy Validation Accuracy Nearest Neighbors had the highest precision but the low- CNN 0.810 0.871 est recall. The fact that k-Nearest Neighbors is more sensitive to noise and outliers in the data may cause its Table 3 low recall. On the other hand, Support Vector Machines Results of Convolutional Neural Network had the highest recall but relatively lower precision. This suggests that SVMs are more efective at identifying pos- Here are reported also the loss and accuracy plots for itive and negative samples. Regarding precision and re- our model during training (Figure 1 & Figure 2). The loss call, Random Forests and Support Vector Machines have plot shows how the model’s training loss decreased over demonstrated balanced performance, indicating that they each epoch, while the accuracy plot shows how well the can handle unbalanced datasets. model performed on the training and validation data. 3.3. Convolutional Neural Network
In conclusion, the results produced by CNN outperform the efectiveness of the machine learning techniques previously investigated. This implies that deep learning methods could be a useful tool for identifying autism from EEG signals. This study paves the way for additional research in this area and shows the utility of deep learning techniques for this application. Deep learning models may be further developed and optimized in order to increase the precision and reliability of autism detection, which could ultimately result in an earlier diagnosis and course of treatment for autistic people.