Personalized Educational Framework for Autism Using EEG and ERP Biomarkers with Deep Learning Tools

Abstract

Children with Autism Spectrum Disorder (ASD) exhibit heterogeneous auditory processing profiles that are not accommodated by standardised acoustic environments. This thesis presents a two-stage framework for objective and personalized acoustic assessment using electroencephalography (EEG). The first stage employs a Convolutional Neural Network–Bidirectional Long Short-Term Memory (CNN-BiLSTM) classifier with trial-level voting to identify children with ASD from single-trial EEG. The second stage introduces the Neural Processing Efficiency Score (NPES), a composite index derived from four Event-Related Potential (ERP) biomarkers using within-subject normalisation, to recommend the optimal acoustic configuration for each child across five stimulus dimensions. The within-subject approach was adopted after group-level ERP comparisons yielded no significant differences between ASD and typically developing children. Evaluation on 22 ASD participants confirmed substantial heterogeneity in the recommended configurations, supporting the need for individualised rather than group-level acoustic assessment. The framework provides a methodological basis for neurophysiologically grounded acoustic personalisation in educational and therapeutic contexts.