Autism Spectrum Disorder (ASD) Detection Using Multiple Deep Learning Models and Comparison of Model Performances

Abstract

Autism Spectrum Disorder (ASD) impairs speech, social interaction, and behavior. Machine learning is a field of artificial intelligence that focuses on creating algorithms that can learn patterns and make ASD classification based on input data. The results of using machine learning algorithms to categorize ASD have been inconsistent. More research is needed to improve the accuracy of the classification of ASD. To address this, deep learning techniques such as 1D CNN, DNN, LSTM etc. have been proposed as an alternative for the classification of ASD detection. The proposed techniques are evaluated on publicly available three different ASD datasets (children, Adults, and adolescents). The conventional method for diagnosing ASD involves clinical evaluations which are often time-consuming and subjective. With the rising prevalence of ASD, the demand for automated, scalable, and accurate detection systems has increased. Deep Learning (DL) models, including 1-Dimensional Convolutional Neural Network (1D CNN), Deep Neural Network (DNN), LSTM, BiLSTM, CNN-LSTM have shown remarkable results in this domain. This paper compares multiple deep learning techniques using publicly available ASD datasets for Children, Adolescents, and Adults. Experimental results demonstrate that the DNN achieves the highest accuracy of 93.65% for adolescents, 99.90% for adults and 99% for children, while 1D CNN, LSTM, BiLSTM and CNN-LSTM also offer competitive results.