Participant Condition Aware Hybrid Deep Learning for Automated EEG Biomarker Discovery and Autism Associated Neural Pattern Identification

Abstract

Autism Spectrum Disorder (ASD) is a neurodevelopmental condition associated with differences in social communication, sensory processing, and behavioural characteristics. Electroencephalography (EEG) provides a non-invasive approach for investigating autism-associated neural activity. However, conventional EEG analysis methods often rely on manually engineered features and may fail to capture complex temporal dependencies, inter-channel relationships, and stimulus-dependent neural patterns. To overcome these limitations, this study develops a hybrid deep learning framework for automated EEG feature extraction and autism-associated biomarker identification using the EEG Latinus Model dataset, comprising 44 participants, including 10 individuals with ASD and 34 neurotypical controls. The proposed framework integrates the REVE EEG Foundation Model for pretrained representation learning, a Dilated Temporal Convolutional Network (D-TCN) for temporal feature extraction, and a Graph Attention Network (GAT) for inter-channel relationship modelling. Furthermore, the proposed Participant–Condition Interaction Residualization (PCIR) method analyses condition-dependent EEG feature changes across photographic face, Mooney face, schematic face, and non-face stimuli. The experimental evaluation demonstrates that the fused spatiotemporal representation achieves a between-group separation value of 0.73, compared with 0.42 for REVE alone. Candidate EEG features exhibit effect sizes ranging from 0.63 to 0.88, with adjusted p-values between 0.002 and 0.029. The proposed framework achieves 97.50% accuracy, 97.40% precision, 97.10% recall, a 97.20% Macro-F1 score, and a 98.10% ROC-AUC. These findings demonstrate the potential of integrated spatiotemporal learning and participant condition analysis for identifying autism-associated EEG patterns and candidate biomarkers.