AI-Assisted Analysis of EEG for Diagnosis of Autism Spectrum Disorder: Pilot Study

This paper addresses the intricate challenge of classifying electroencephalogram (EEG) signals to distinguish between individuals with Autism Spectrum Disorder (ASD) and Normal controls. Autism Spectrum Disorder is a complex neurodevelopmental condition exhibiting a diverse array of symptoms across social communication, behavior, and cognition. The EEGNetv4 deep learning model is employed, trained on data from a limited cohort of 9 subjects. The study utilizes a Leave-One-Subject-Out cross-validation methodology to assess the model's generalization across a small and unbalanced dataset. Following the cross-validation process, the model exhibits varying accuracies across folds, emphasizing the heterogeneous nature of ASD. Notably, the highest cross-validation accuracy is achieved in fold #3, reaching 83.53%, showcasing the model's proficiency in certain subject-specific contexts. Subsequently, the model trained on fold #3 is selected for further testing, demonstrating a test set accuracy of 74.49% on EEG data from two additional subjects, one representing a Normal control and the other diagnosed with ASD. This research underscores the inherent challenges posed by the small and unbalanced dataset, highlighting the necessity for cautious interpretation of results. Moreover, the absence of EEG manual preprocessing with labeling by experts regarding EEG background activity and the presence of significant artifacts that were not eliminated by prior events rejection underscores the potential impact of signal quality on classification outcomes. Despite these challenges, the study provides insights into the feasibility of utilizing EEGNetv4 for ASD classification and calls for future research to address dataset limitations and enhance model robustness.