Discrimination of neuropsychiatric disease using EEG and Neurophysiological Biomarker Toolbox (NBT) with Machine Learning

Electromyogram (EMG) contamination has been shown to affect electroencephalogram (EEG) signals.
Therefore, methods of isolating and removing EMG contamination are a focus of research. One of the most
common ways to eliminate this contamination is through independent component analysis (ICA). Also, surface
Laplacian (SL) has been proven to isolate the distant sources of EEG signals. The objective of this paper is to
demonstrate the effects of EMG contamination on EEG signals using the Neurophysiological Biomarker
Toolbox (NBT) and the impact of applying ICA, and ICA + SL on raw data. In this paper, the method for
preparing the data is ICA with an auto-pruned method and SL using a flexible spherical spline. Machine
learning was used to classify three neuropsychiatric diseases (anxiety, depression, and epilepsy) against control
subjects under the three types of data pre-processing and raw data + SL. The data has been split into one
second segments and classified according to features extracted from the NBT, which are the amplitude and the
normalised amplitude for all frequency bands. Principal component analysis (PCA) was used for reducing the
features, and 10-fold cross-validation and artificial neural networking were the methods that has been used for
the classification. The results show a high percentage of accuracy in ICA + SL in all frequency bands. However,
ICA in general has a percentage quite similar to the raw data, while SL, as well as ICA with a small percentage
improved more than ICA and raw data. Overall, the gamma band for both amplitude and normalised amplitude
in ICA + SL showed the best results, with accuracy over 87%, when comparing it with all disease
classifications. Both results indicate that ICA + SL eliminate and isolate EMG contamination. However, the
classification of ICA shows no significant change in the percentage of accuracy