Preliminary Evaluation of a Machine-Learning Based Characterization of Electroencephalogram Patterns in Sleep-Related Pathologies

Electroencephalography (EEG) patterns typically change in association with different types of sleep-related pathologies, with respect to normal patterns observed in control groups. Several of these changes have been well described in the scientific literature, based mainly on frequency-domain analyses for each pathology. However, the case-by-case nature of these studies, together with the inherent complexity of EEG signals, represents a challenge in obtaining complete characterizations for other pathologies, and many studies thus focus on combining features from different types of signals. In this paper, we present a machine-learning approach for detecting EEG-only changes in different pathologies. We apply different classifiers to frequency-domain EEG features and compare these features over control and experimental groups. Our study cases initially focus on nocturnal frontal lobe epilepsy (NFLE), narcolepsy, bruxism, insomnia, rapid eye movement sleep behavior disorder, obstructive sleep-disordered breathing, and periodic limb movement disorder. However, in this preliminary study, we provide results regarding NFLE, for which we provide the initial tests. Our results suggest that the use of random forests (RFs) and support vector machines (SVMs) can potentially distinguish the control and experimental groups based solely on EEG features, although the accuracies are still limited. The use of convolutional neural networks, on the other hand, did not provide improved detection, possibly due to the limited number of training cases. In the next stages of the research, we will focus on applying feature selection methods, to try and improve the detection performance when using RFs and SVMs, for NFLE and for the other described pathologies.