Feature-Based EEG Classification: A Comparative Study of EMD and EWT with EEGNet and Shallow FBCSP ConvNet

Autism is a neurodevelopmental disorder characterized by difficulties in social interaction, communication, and repetitive behaviors. Recent research has shown that electroencephalogram (EEG) signals from individuals with autism have distinctive brain activity patterns and can be utilized for automatic classification purposes. This study aims to compare two signal feature extraction methods, namely Empirical Wavelet Transform (EWT) and Empirical Mode Decomposition (EMD), in the context of EEG-based autism classification. The preprocessed EEG signals are segmented into four-second windows with 50% overlap, and then classified using two deep learning architectures: EEGNet and Shallow FBCSPNet. The evaluation process is performed using a 5-fold cross-validation scheme to ensure the reliability of the results. Classification accuracy is calculated for each fold and statistically analyzed using a one-way ANOVA. The results show that the combination of EEGNet with EMD produces the highest average test accuracy of 97.99%, followed by Shallow FBCSPNet with EMD (96.80%). To control family-wise error across the two within-model comparisons (EWT vs. EMD within EEGNet and within Shallow FBCSPNet), a Holm–Bonferroni correction at α=0.05 was applied; accordingly, both models show significant differences between EWT and EMD (adjusted p<0.05). This finding confirms that the choice of feature extraction method significantly influences the classification accuracy and overall performance of EEG-based autism detection systems.