Classifying developmental delays with EEG: A comparative study of machine learning and deep learning approaches

• Combined language and neurophysiological assessments differentiated DDG from TG. • Traditional ML models with hand-crafted features were evaluated for DDG detection. • ML and DL methods, including a 1D CNN model, were compared for classification performance. • The proposed 1D CNN showed promising results in identifying developmental delays . Early detection of developmental delays is crucial for improving children’s cognitive, social, and emotional outcomes through timely interventions. This study explores the potential of machine learning (ML) and deep learning (DL) to classify Electroencephalography (EEG) data from an oddball task, distinguishing between children with and without developmental delays. Participants underwent language assessments and EEG recordings, with subsequent analysis using Event-Related Potentials (ERPs), Event-Related Spectral Perturbations (ERSPs), and functional connectivity to characterize group differences. Three methodologies were employed in this research to classify EEG data. Firstly, statistical features are extracted from the EEG data and various ML algorithms are applied for classification, with feature selection techniques utilized to identify the most relevant features and enhance classification accuracy. Secondly, brain dynamics is utilized to incorporate ERP, ERSP, and functional connectivity measures as features for developmental delay detection. Similar to the first approach, feature selection techniques are again employed to enhance classification accuracy. Lastly, DL approaches are explored by implementing multiple convolutional neural networks (CNNs), including a 2D CNN (EEGNet), various hybrid CNN architectures integrating LSTM, GRU, and attention mechanisms, and a novel 1D CNN with a standardized convolutional layer (SCL) for improved stability and training performance. The effectiveness of each approach in accurately classifying EEG data for developmental delay detection is independently analyzed. The results demonstrate that the proposed 1D convolutional neural network outperforms both EEGNet and the employed ML classifiers. This model achieves an impressive accuracy of 96.4% and an F1 score of 96.6%, underscoring its potential as a valuable tool for early and accurate developmental delay detection using EEG data.