Impact of Sliding Window Overlap Ratio on EEG-Based ASD Diagnosis Using Brain Hemisphere Energy and Machine Learning

Abstract

Featured Application: This research serves as a valuable resource for future designers and developers seeking to integrate and highlight this system’s potential as an online decision-support tool to assist healthcare professionals in diagnosing autism. Early screening is a key component of intensive intervention therapy and rehabilitation for children with autism spectrum disorder (ASD). Electroencephalogram (EEG) signals provide real-time, high-sensitivity monitoring of pathological activities in children with ASD. This study used a dataset of 52 samples with 19-channel sleep data and proposed a new approach to diagnosing ASD based on energy differences between the left and right hemispheres of the brain. The preprocessing stage included decimation, band-pass filtering to remove unwanted frequencies, artifact subspace reconstruction to eliminate artifacts, and amplitude normalization to preserve the relative relationships between the signal features, which were crucial for the subsequent analysis and classification. Five band-pass filters were applied to decomposing the EEG signals. For each decomposed band, the signals from each brain hemisphere were analyzed separately by calculating the peak sensor’s envelope and obtaining the mean envelope for each hemisphere, resulting in two mean signals (right and left). Features were extracted using a sliding window approach applied to the mean signals of each hemisphere, with varying overlap ratios (12.5% to 87.5%, in 12.5% steps). The maximum, mean, and minimum energy values were used individually as features. Three types of SVM kernels—linear (L), the radial basis function (RBF), and quadratic—were employed for classification. The proposed method achieved the highest accuracy, sensitivity, and F1-score of 91.7%, 91.4%, and 91.6%, respectively, in the Theta band using SVM-L with the maximum energy features and the maximum overlap ratio.