Abstract The description of autism spectrum disorder (ASD) includes typically the alterations in social skills and attitude as well as problems of communication. At present, the most promising approaches suggested for ASD automatic diagnosis are the machine learning (ML)‐based procedures that begin by the identification and extraction of the most discriminant features. One of the proposed discriminant features of ASD is electroencephalogram (EEG) entropy that reflects the level of complexity or randomness of brain dynamics. The two main techniques currently used to describe the entropy in ASD are multi‐scale entropy (MSE) and spectral entropy (SE). However, MSE is unable to track the exact frequency content on the different time scales that suffer—themselves—from mismatching. In addition, the outcomes of SE approaches in literature are contradictory and confusing about the altered band, direction of change as well as about the affected brain region. The common point in those SE approaches is the examination of one or more of the well‐known frequency bands: Delta, Theta, Alpha, Beta and low Gamma. Hence, the range of high Gamma ripples (>80 Hz) is barely examined in ASD literature due to the needed high carefulness in its technical acquisition, detection and processing as well as to the common acceptance of the idea that it is rather related to pathological seizure activity; which is a misconception recently refuted. Further, the study of automatic ASD diagnosis based on ripples is almost inexistent in literature. The present work suggests an accurate automatic technique to classify ASD and neuro‐typical EEG entropy based on the whole range of ripples (75–250 Hz): (1) EEG signals are collected (120 ASD and neuro‐typical children) by an efficient 64‐channel system highly suitable in noisy environment with movement artefacts. (2) All signal disturbances are automatically removed. (3) The power spectra of all EEG channels are computed. (4) Shannon entropy values are calculated for the range of ripples without filtering. (5) Scaled Conjugate Gradient‐based Neural Network (NN) classification, k‐fold cross validation and T‐test are applied to entropy values. (6) Statistical assessment is conducted. The best testing and the overall accuracies were found to be 98.9% and 95.0% respectively.

Abstract Autism spectrum disorder (ASD) is a neurological and developmental disorder that commences usually in the early years of age. It impacts the social interaction, communication and learning. It is believed that it is mainly caused by an abnormal connectivity between brain zones. The presented work applies an EEG‐based nonlinear method for the classification of ASD and neuro‐typical groups. The suggested procedure does not require pre‐assumptions and is completely data‐driven. Also, the main advantage is the accurate tracking of the pace of EEG activity without the effect of amplitude modulation on the spectral information. In addition, the tracking is conducted point‐by‐point to avoid shortcomings of inexact global features. First, for every (ASD or neuro‐typical) volunteer, the recorded EEG channels (64 channels) are decomposed by empirical mode decomposition (EMD) in order to get the underlying components (intrinsic mode functions—IMFs). Second, the direct quadrature (DQ) method is used to normalize the IMFs, and to dissociate between amplitude and frequency contents of the resulted components, as well as to help extract then the point‐by‐point spectral information from the analytic normalized IMFs by Hilbert transform. Third, the correlation coefficients between the instantaneous frequency vectors of the counterpart components will be computed over all channels (i.e., between components number ‘i’ of channels ‘x’ and ‘y’, 1 < i < number of components, 1 < x < 64, 1 < y < 64). Fourth, correlation coefficients array is constructed. Fifth, the dimension of the feature array is reduced without loss of significant information. Sixth, classification of reduced features is achieved via neural network. Finally, the statistical assessment of the classification outcome is conducted. The proposed method yields a test accuracy of 94.1%–100%.