Robust Autism Spectrum Disorder Classification via Optimized 15-Channel Feature Mapping Through Multi-Domain Features Analysis

The paper describes a powerful machine learning algorithm to classify the Autism Spectrum Disorder (ASD) correctly through electroencephalography (EEG) activities. Our solution combines an important channel choice according to the statistical spectral analysis with an exhaustive multi-domain feature extraction. EEG data of 40 children (20 ASD, 20 typically developing) underwent bandpass filtering (1-45 Hz) along with Fz re-referencing, which resulted in 14,349 epochs. A statistical analysis was used to reveal 11 significant channels with differing spectral power (p<0.05). We used relative spectral power in 5 frequency bands, Hjorth parameters, sample entropy, and statistical moments to create a 494-dimensional feature matrix. There were 4 machine learning classifiers that were tested on 5-fold cross-validation. XGBoost had better performance in terms of 96.11…