Detection of attention state in children with autism spectrum disorder based on neural network classification of electroencephalograms

Abstract

Autism spectrum disorder (ASD) is a neurological condition characterized by impairments in social interaction. This diagnosis carries economic and social implications due to its high prevalence and associated morbidity. Data from electroencephalogram (EEG) sensors is numerical and serves as the input for machine learning-based predictions. The input data in this research includes features extracted from EEG signals, such as theta/beta ratio, theta/alpha ratio, and other relative power metrics, which are closely linked to cognitive control and attentional dynamics. These data are organized into two balanced classes: “Attention” and “No Attention,” comprising a total of 33 936 samples. This paper proposes 12 weighted and weighted-average ensemble models to enhance the accuracy of predicting attentional cues in individuals with ASD. For ensembling three multilayer perceptron architectures were developed and trained using various optimizers. The accuracy of the employed ensemble model of three multilayer perceptrons reached 95.90 %. The findings of this research can contribute to the advancement of novel diagnostic approaches and educational initiatives and serve as a foundation for future research utilizing machine learning techniques and the creation of innovative technologies for attention monitoring and training.