Enhancing Autism Spectrum Disorder Recognition in EEG Data through Filtering-Driven CNN Approach

Wearable Electroencephalogram (EEG) devices hold immense potential for optimizing health monitoring processes. This paper introduces an innovative approach that seamlessly integrates deep learning techniques into EEG-based recognition, bridging the gap between traditional feature-based methods and Convolutional Neural Networks (CNNs). The methodology revolves around crafting CNNs to emulate the information extraction principles of feature-based techniques, resulting in a streamlined yet robust architecture tailored for EEG-based classification.To validate this approach, two comprehensive experiments were conducted, involving EEG data collection and subsequent evaluation within health monitoring scenarios. Specifically, the focus was on the diagnosis of Autism Spectrum Disorder (ASD), necessitating precise identification and characterization of EEG signals. Deep learning, embedded within a neural network framework, emerges as a powerful tool for the identification and classification of ASD.The proposed deep CNN structure employs separated temporal and spatial filters, selecting raw EEG signals from electrode pairs over the motor cortex region as hybrid samples without preprocessing or artificial feature extraction operations. This 5-layer CNN is augmented by a 4-layer max pooling mechanism to reduce dimensionality, and a fully-connected (FC) layer for classification. Dropout and batch normalization techniques are incorporated to mitigate overfitting risks.In the experimental phase, 20m EEG data from 5, 10, 20, and 50 subjects sourced from the Physionet database, utilizing a 14 Channel Emotiv EPOC EEG headset, were employed. States of focus, unfocused, and drowsed were considered. Results indicate a remarkable global averaged accuracy of 92.31% in group-level classification, an outstanding area under the receiver operating characteristic (ROC) curve at 0.981, and, notably, the electrode pair FC3-FC4 achieved the highest accuracy of 99.61% on a 10-subject dataset.This work contributes to the evolving landscape of EEG-based health monitoring, showcasing the potential of a deep CNN architecture in accurately classifying states related to ASD, and thereby advancing the precision and efficacy of neurodevelopmental disorder diagnoses.