Machine Learning Approaches for Electroencephalography-based Biomarker Discovery in Autism Spectrum Disorder

Abstract

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by challenges in social interaction, communication, and repetitive behaviors. Electroencephalography (EEG), with its high temporal resolution, offers valuable insights into the neural dynamics associated with ASD. This paper proposes a novel convolutional neural network architecture designed to enhance the accuracy of ASD screening by separately extracting spatio-temporal features from EEG signals. The network is structured into three main modules: preprocessing, feature extraction, and ASD screening. Initially, EEG signals are transformed into topological maps to serve as input for the CNN. The feature extraction module then processes these maps to independently capture spatial and temporal features, which are subsequently aggregated and used by the ASD screening network. Additionally, the model incorporates demographic information as auxiliary input to improve screening accuracy. We also conducted brain anatomy-driven experiments, filtering out specific brain regions from the input EEG signals to determine which parts are most highly correlated with ASD screening. The ablation study demonstrated the effectiveness of the proposed spatio-temporal approach, and additional t-SNE evaluations further validated its robustness. The proposed network achieved state-of-the-art performance, with an accuracy of 97.09% on a public dataset.