Optimal EEG Data Segmentation in LSTM Networks for Learning Neural Dynamics of ADHD

Abstract

Attention deficit/hyperactivity disorder (ADHD) is a condition that generally affects the neurodevelopment of children. Both electroencephalography (EEG) and continuous performance tests (CPT) can be used not only to provide objective identification ADHD, but also to directly observe and quantify the performance of subject task performance. In this study, we propose an optimized segmentation for learning EEG time-series information with long short-term memory (LSTM) networks used to differentiate ADHD and neurotypical (NT) children. A total of 30 NT children and 30 children diagnosed with ADHD participated in CPT while simultaneously monitored with EEG. The experimental results show that, whether it is a single feature of beta power at a the O2 electrode location or all the features, the optimal data segment is the same, an EEG segment containing the data from 30 seconds of eyes-open resting and 30 seconds of CPT task. This produced improved performance for discriminating the differences in EEG between the two groups, thus assisting in the diagnosis of ADHD.