Robust and Interpretable Deep Learning on EEG Spectrograms for Autism Spectrum Disorder Detection

Abstract

Autism spectrum disorder (ASD) manifests early alterations in neural dynamics that often precede observable behavioral symptoms, yet current diagnostic practices rely predominantly on subjective clinical assessments. In this study, we present a robust, image-based convolutional framework for automated ASD detection from resting-state electroencephalography (EEG), validated across two independent cohorts: the King Abdulaziz University ASD dataset and the Autism Centre of Excellence (ACE) dataset. Raw EEG recordings were transformed into time–frequency spectrograms via continuous wavelet transforms and then fed into three deep architectures—a custom 4-layer convolutional neural network (CNN), ResNet50, and EfficientNet—to learn discriminative features of ASD versus neurotypical patterns. To enhance clinical trust and interpretability, we integrated Smooth Grad-CAM++ to generate high-resolution activation maps that pinpoint critical spectral–temporal regions driving each classification. On the King Abdulaziz dataset, our custom 4-layer CNN achieved a test accuracy of 98.59%, with an average F1-score of 0.99, precision of 0.97, recall of 0.98, and specificity of 95.36%. On the ACE dataset, ResNet50 yielded a test accuracy of 95.23%, F1-score of 0.96, precision of 0.97, recall of 0.94, and specificity of 97.43%. Across both cohorts, all models consistently exceeded 95% accuracy and demonstrated balanced sensitivity and specificity, underscoring their generalizability. These results establish a high-performance, explainable computer-aided diagnostic system for early ASD detection using EEG spectrograms. The conjunction of deep feature learning and explainable AI visualization not only accelerates diagnostic workflows but also offers actionable insights into the neural substrates of ASD, paving the way for timely, objective interventions.



Received: 28 September 2025 | Revised: 16 March 2026 | Accepted: 27 April 2026



Conflicts of Interest

The authors declare that they have no conflicts of interest to this work.



Data Availability Statement

The Dataset I (KAU dataset) that supports the findings of this study is openly available at https://www.earticle.net/Article/A207042, Reference number [23]. The Dataset II (ACE dataset) that supports the findings of this study is openly available in the NIH National Database for Autism Research (NDA) at https://nda.nih.gov/edit_collection.html?id=2021.



Author Contribution Statement

Andrew Jeyabose: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Writing – original draft, Writing – review & editing, Visualization, Supervision. Arav Chadda: Methodology, Software, Writing – original draft, Writing – review & editing, Visualization. Venkatesh Bhandage: Conceptualization, Software, Formal analysis, Resources, Data curation, Visualization, Supervision.