A deep learning framework for identifying children with ADHD using an EEG-based brain network

Abstract The convolutional neural network (CNN) is a mainstream deep learning (DL) algorithm. However, the application of DL techniques in attention-deficit/hyperactivity disorder (ADHD) studies is still limited. Electroencephalography (EEG) is an informative neuroimaging tool. In this study, we propose a DL framework for the ADHD identification problem by combining an EEG-based brain network with the CNN. By reorganizing the order of the channels, we proposed a new form of the connectivity matrix to adapt the concept of the convolution operation of the CNN. The correlations between the deep features derived from the CNN models and 13 hand-crafted measures of the brain network were also analyzed. We collected EEG data from 50 children with ADHD (9 girls, mean age: 10.44 ± 0.75) and 51 handedness- and age-matched controls, and we used mutual information (MI) to quantify the synchronization between channels. We demonstrated the feasibility of the framework and discussed some critical concerns in the application of the framework. Some of the practical suggestions were also given based on the validation results. The proposed framework achieved a convincing performance with an accuracy of 94.67% on the test data. We also validated the validity of the form of the connectivity matrix, which enabled the models to achieve better performance. This finding suggests that the data representation in the DL framework is important. Seventeen deep features showed significant between-group differences, and had significant correlations with hand-crafted measures, thereby reflecting the amazing learning ability of the method for finding the deviations in the brain network of children with ADHD. The proposed framework is broadly applicable to the ADHD identification problem. Nevertheless, the validation of this methodology with a large and well-matched sample of children is needed in the future.