Neurological state changes indicative of ADHD in children learned via EEG-based LSTM networks

Abstract

Attention-deficit/hyperactivity disorder (ADHD) is a neurodevelopmental disorder that pervasively interferes with the lives of individuals starting in childhood. Objective. To address the subjectivity of current diagnostic approaches, many studies have been dedicated to efforts to identify the differences between ADHD and neurotypical (NT) individuals using electroencephalography (EEG) and continuous performance tests (CPT). Approach. In this study, we proposed EEG-based long short-term memory (LSTM) networks that utilize deep learning techniques with learning the cognitive state transition to discriminate between ADHD and NT children via EEG signal processing. A total of 30 neurotypical children and 30 ADHD children participated in CPT tests while being monitored with EEG. Several architectures of deep and machine learning were applied to three EEG data segments including resting state, cognitive execution, and a period containing a fusion of those. Main results. The experimental results indicated that EEG-based LSTM networks produced the best performance with an average accuracy of 90.50 ± 0.81% in comparison with the deep neural networks, the convolutional neural networks, and the support vector machines with learning the cognitive state transition of EEG data. Novel observations of individual neural markers showed that the beta power activity of the O1 and O2 sites contributed the most to the classifications, subjects exhibited decreased beta power in the ADHD group, and had larger decreases during cognitive execution. Significance. These findings showed that the proposed EEG-based LSTM networks are capable of extracting the varied temporal characteristics of high-resolution electrophysiological signals to differentiate between ADHD and NT children, and brought a new insight to facilitate the diagnosis of ADHD. The registration numbers of the institutional review boards are 16MMHIS021 and EC1070401-F.