Neurophysiological EEG Characterization of Autism Spectrum Disorder Using DWT‐Based Frequency Analysis With Selective Electrodes and Brain Segmentation: An Explainable AI‐Driven Approach

Abstract

Autism Spectrum Disorder (ASD) diagnosis benefits from the technical analysis of neural oscillations. The objective identification of Autism Spectrum Disorder (ASD) is advanced through a dedicated engineering framework that analyzes neural oscillations via electroencephalogram (EEG) signals. This methodology synthesizes statistical signal processing, frequency‐domain transformation, and computational intelligence. Multi‐channel EEG data from 15 electrodes were first organized into seven functional brain regions. A Discrete Wavelet Transform (db4) was then applied for multiresolution decomposition, extracting and normalizing the spectral power of delta, theta, alpha, beta, and gamma frequency bands. This spectral characterization was augmented with statistical descriptors to quantify signal dynamics and complexity, creating a robust multi‐domain feature set. For classification, the framework implements a comprehensive comparative analysis across multiple learning paradigms. The engineered feature space was processed by nine classical machine learning algorithms, eight deep neural network architectures, and four meta‐learning strategies. Empirical validation revealed that Logistic Regression achieved 88.57% accuracy with an AUC of 0.77. Further Extra Trees achieving 86.42% accuracy. Long Short‐Term Memory (LSTM) networks demonstrated strong sequential modeling capability, attaining a competitive accuracy of approximately 87%. While meta‐learning preserved diagnostic sensitivity, it did not surpass the leading standalone models. To ensure clinical interpretability, Explainability was employed, identifying delta and theta band power as the most critical discriminative features. Consequently, this study delivers a technically rigorous, transparent, and automated pipeline for electrophysiological signal characterization, establishing a preliminary computational foundation for neurophysiological differentiation in pediatric ASD. All findings are preliminary and exploratory, pending validation in larger independent cohorts.