EEG based classification of children with learning disabilities using shallow and deep neural network. (April 2023)

• Record Type: Journal Article
• Authors: Guhan Seshadri, N.P.; Agrawal, Sneha; Kumar Singh, Bikesh; Geethanjali, B.; Mahesh, V.; Pachori, Ram Bilas
• Is Part Of: Biomedical signal processing and control. Volume 82(2023)
• Publication Date: 2023-04
• Subjects: Learning disability -- Digital wavelet transform -- EEG -- Shallow network -- Deep neural network; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.104553
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Rest EEG data was acquired from children with learning disability (LD). Shallow and deep networks were compared with different machine learning classifiers for LD classification. Shallow neural network offered high performance on LD classification. The proposed framework achieved highest classification accuracy of 97.5% using shallow neural network. Abstract: Learning disability (LD), a neurodevelopmental disorder that has severely impacted the lives of many children all over the world. LD refers to significant deficiency in children's reading, writing, spelling, and ability to solve mathematical task despite having normal intelligence. This paper proposes a framework for early detection and classification of LD with non-LD children from rest electroencephalogram (EEG) signals using shallow and deep neural network. Twenty children with LD and twenty non-LD children (aged 8–16 years) participated in this study. Preprocessing the raw EEG signal, segmentation and extraction of various features from the alpha, beta, delta, and theta bands obtained using digital wavelet transform (DWT). Filter based feature selection method were employed for the selection of most relevant features that reduces the computation burden on models. Afterwards, these ranked accumulated features were evaluated separately by machine learning (ML) classifiers and neural network (shallow and deep) models to investigate the performance. The performance of the ML classifiers and one-hidden layer shallow neural network and 3-hidden layer deep neural network were compared. Experimental results showed that the most relevant features computed by ReliefF algorithm along with the shallow neural network based classifier attained the highest average and maximum classification accuracy of 95.8 % and 97.5 % respectively, which is greatest among the existing literatures. The efficient and automatic LD classification from EEG signal could aid in the development of computer-aided diagnosis systems for early detection.