1D-convolutional neural network and fast Walsh–Hadamard transform approach for diagnosing autism spectrum disorder

• Published in: Neural computing & applications Vol. 37; no. 18; pp. 13039 - 13057
• Main Author: Göker, Hanife
• Format: Journal Article
• Language: English
• Published: London Springer London 01.06.2025; Springer Nature B.V
• Subjects: Artificial Intelligence; Artificial neural networks; Autism; Computational Biology/Bioinformatics; Computational Science and Engineering; Computer Science; Correlation coefficients; Data Mining and Knowledge Discovery; Electroencephalography; Feature extraction; Image Processing and Computer Vision; Machine learning; Neural networks; Original Article; Principal components analysis; Probability and Statistics in Computer Science; Deep learning; Signal processing; Electroencephalogram signals; Autism spectrum disorder; Fast Walsh–Hadamard transform
• ISSN: 0941-0643; 1433-3058
• DOI: 10.1007/s00521-025-11208-3

Autism spectrum disorder is a neuro-developmental disability that can lead to a variety of communication, social, and behavioral challenges. The disorder affects approximately 1 in 54 children. Traditional diagnostic methods for autism spectrum disorder often involve subjective observations and extensive testing, which can be costly, time-consuming, and prone to inaccuracies. Electroencephalogram (EEG) presents a promising alternative because of its non-invasive, affordability, and ability to provide rapid results. This study combined the fast Walsh–Hadamard transform (FWHT) with one-dimensional convolutional neural network (1D-CNN) to propose an EEG-based solution for diagnosing autism spectrum disorder. The dataset has included resting-state EEG signals recorded from 62 subjects, including 31 autism spectrum disorder patients and 31 healthy controls using 17 channels. Multiscale principal component analysis (multiscale PCA) was used to improve the data quality by de-noising the raw EEG signals. FWHT coefficients were used to extract different feature sets and their performances were compared. The performance results of the FWHT feature extraction method and the 1D-CNN algorithm were computed as 0.971 precision, 0.972 specificity, 0.988 sensitivity, 0.979 f1-measure, 0.960 kappa statistic, 0.960 Matthew’s correlation coefficient, and 98.01% accuracy. Compared to the previous methods, the EEG-based deep learning model had a higher and more promising performance.