An Explainable Machine Learning Network for Classification of Autism Spectrum Disorder Using Optimal Frequency Band Identification From Brain EEG

Autism Spectrum Disorder (ASD) is characterized by repetitive and restricted behavior, involving challenges for subjective detection methods. There is a need for an objective diagnostic technique that maximizes accuracy while utilizing a minimum of Electroencephalography (EEG) channels. This research aims to enhance ASD classification by utilizing spatial patterns in EEG signals. The primary objective is to identify the optimal frequency band for classification. This work analyzes EEG data from ten ASD children and ten Typically Developed (TD) children. A novel spatial filtering method is proposed using wavelet transform, filterbank, tikhonov regularization, and common spatial pattern algorithms to design the spatial filters for EEG signals from central lobe regions (C3, C4, and Cz). In order to determine optimal frequency bands, peak-to-peak amplitudes are extracted from these spatially filtered signals, and the most informative features are selected using correlation coefficients. The proposed method identifies the optimal frequency band in the alpha and beta ranges, specifically 12-16 Hz within 8-32 Hz, using Support Vector Machine (SVM) with a Radial Basis Function (RBF) kernel. Additional features, including fuzzy dispersion entropy, fuzzy entropy, fractional fuzzy dispersion entropy, Detrended Fluctuation Analysis (DFA) exponent, and multifractal DFA-related features, are incorporated for the optimal frequency band and features selected using Shapley additive explanations (SHAP). Results from the experiments demonstrate the efficiency of this approach with selected features, and the SVM with RBF achieved an accuracy of 93.59%, precision of 96.97%, recall of 94.12%, specificity of 92.19%, and F1-score of 95.52% on the optimal frequency band.