Abstract
PURPOSES: There is a great need for mechanistically informed biomarkers to understand autism spectrum disorder (ASD) and guide treatment. Electroencephalography (EEG) is a non-invasive method for identifying objective biomarkers, but traditional trial-averaged metrics may mask neural variability, a meaningful feature of ASD reflecting sensory, attentional, and cognitive differences. METHODS: This study investigates whether across-trial EEG variability enhances ASD classification compared to conventional mean EEG features. We hypothesize that capturing dynamic within-subject neural variability improves classification accuracy and offers deeper insights into ASD-related neural disruptions. We analyzed EEG power spectral features in individuals with and without ASD, extracting across-trial variability in five frequency bands alongside traditional mean EEG power metrics. Using machine learning, we compared classification performance and identified the most predictive neural markers. RESULTS: Results show that across-trial EEG variability outperformed mean EEG metrics, achieving 70.7% classification accuracy. Variability in delta and gamma bands is critical for distinguishing ASD, with robust cross-validation results and significant correlations with behavioral scores, supporting the clinical relevance and generalizability of neural variability as an ASD biomarker. CONCLUSIONS: By incorporating neural variability into machine learning models, this study introduces a novel framework for improving biomarker-driven assessments. These findings highlight the potential for personalized tools that inform targeted interventions while offering insights into ASD neurophysiology. Future research should integrate longitudinal EEG analyses and multimodal neuroimaging to advance precision diagnostics in autism.