A Dual Paradigm of Machine Learning and Deep Learning for Early Identification of Autism Spectrum Disorder in Children Using Facial Image Dataset

Autism Spectrum Disorder (ASD) is a complex, lifelong developmental condition that influences how individuals perceive, process, and engage with the world around them. Early childhood diagnosis of ASD is essential for appropriate treatment and independent adult function. Most autistic children are not diagnosed until they are at least three years old. Unusual facial characteristics induced by developmental disorders are strongly associated with underlying neurological difficulties. Researchers propose facial images as a viable tool for diagnosing autism since autistic children frequently start developing distinguishable features from the toddler period. Our prior research focused on feature extraction from children’s facial images to detect autism spectrum disorder with an accuracy of 88.33%. Based on that research, we conducted extensive work and employed various architectural models to suspect early autism. We used the same dataset as in that research work but partitioned the data differently. For early identification of Autism in children, this study used five pre-trained CNN models (MobileNetV3Large, DenseNet121, EfficientNetB2, Xception, and VGG19) as feature extractors, followed by logistic regression with AdaBoost, support vector machine with AdaBoost, and a deep neural network (DNN) model as binary classifiers. Moreover, ANOVA analysis was employed as a significant feature selection tool. The test data classification accuracy for DNN, AdaBoost logistic regression, and AdaBoost SVM reached 90.625%, 89.21%, and 88.92%, respectively.