Neurodevelopmental Screening for Autism: EEG-Driven Deep Learning in Early Childhood

Abstract

Autism Spectrum Disorder (ASD) is a neurodevelopmental condition characterized by atypical patterns of communication, social interaction, and behavior. The prevalence of ASD has risen sharply in recent decades, with recent estimates indicating that approximately one in every thirty-six children in the United States is diagnosed. Early diagnosis is crucial because interventions initiated during early childhood-when neural plasticity is greatest-can substantially improve behavioral, cognitive, and linguistic outcomes. However, since current diagnostic practices largely rely on behavioral assessments, symptoms often remain undetected until around three years of age. This underscores the urgent need for objective biological markers capable of identifying ASD risk earlier, ideally during infancy or toddlerhood. Previous research has demonstrated that ASD is associated with atypical event-related potential (ERP) responses across both early and late stages of cognitive processing. Nevertheless, few studies have explored whether such ERP features can support reliable individual-level predictions using modern machine learning (ML) approaches, and even fewer have applied systematic validation procedures. Addressing this gap, the present study employs multiple ML frameworks-particularly a multilayer perceptron (MLP)-to classify ERP features derived from a visual oddball paradigm. Model performance was rigorously evaluated using cross-validation at the single-subject level. The MLP achieved an overall accuracy of 0.746, with a precision of 0.72, recall of 0.736, and an F1 score of 0.734, demonstrating that the model effectively distinguished ERP feature patterns between ASD and typically developing participants.