Real-Time Classification for EEG Data in Children With ASD Using Deep Learning Techniques

Autism spectrum disorder (ASD) presents unique challenges in diagnosis and treatment, necessitating innovative approaches to understanding its underlying neurophysiological mechanisms. Real-time classification of electroencephalography (EEG) data in children with ASD faces significant challenges due to variability in EEG signals caused by individual differences in brain activity, age, and behavioral states, complicating robust algorithm development. This study develops and validates a deep learning-based framework for real-time EEG classification in children with ASD, aiming to enhance diagnostic accuracy and enable timely interventions. The dataset includes EEG recordings from 60 children (30 with ASD and 30 typically developing), representing diverse age groups and behavioral profiles to improve generalizability. Pre-processing removes noise and artifacts through segmentation, short-time Fourier transform (STFT), and independent component analysis (ICA). Grid search optimization (GSO) enhances model performance by systematically searching hyperparameter combinations to find the optimal configuration. A hybrid convolutional neural network (CNN)-long short-term memory (LSTM) framework is proposed, combining convolutional layers for spatial feature extraction with LSTM layers for temporal sequence modeling. This hybrid model is the primary proposed solution for real-time EEG classification due to its ability to capture both spatial and temporal features critical for interpreting sequential EEG data in children with ASD. The model achieves an accuracy of 87.5%, a precision of 85.0%, a recall of 90.0%, and an F1 score of 87.5% implemented using MATLAB software. In comparison, ResNet, a baseline deep CNN model, achieves slightly higher accuracy (89.1%) but lacks temporal modeling capabilities essential for sequential EEG interpretation. Despite ResNet's marginally higher accuracy, the hybrid CNN-LSTM is favored as the final model for its superior temporal modeling, critical in EEG analysis. Future work may include real-time feedback mechanisms, mobile application development, and longitudinal data expansion.