Resting State EEG Classification of Children With ADHD

From resting-state Electroencephalography (EEG) signals, two Attention Deficit Hyperactivity Disorder (ADHD) detectors were created. One of the detectors contained extreme gradient boosting (XGB), which used polynomial combinations of delta and theta relative power band (ID-2), and the other detector was created using a visual representation of the alpha and beta relative power with residuals convolutional neural network (ID-44). A total of 46 experiments were tested where various unsupervised, supervised, and feature extraction techniques were explored. They included K-Means Clustering of brain regions, relative power and sample entropy, random forest, and convolutional neural networks. The main design principles were implementation simplicity and minimum signal preprocessing. This allowed us to test a wider range of statistical techniques as well as to facilitate reproducibility. On 86 test subjects, ID-44 and ID-2 reached precision scores of 90% and 86.3%, and a Fl-Score of 76% and 73.3% respectively. Guided by activation maps in ID-44, we observed significant differences in relative power alpha band mainly in the frontal-temporal lobe and spread around the scalp alpha and beta interaction. Using features important for ID-2, we observed for theta and delta square significant differences for specific clusters regions. Note: ADHD EEG classification is a tool and not a replacement of conventional assessment by psychiatric or neurological experts. Constant expert interpretation is always encouraged.