A Machine Learning Approach for the Classification of Sleep Disorders Using Time and Frequency Domain Features of a Single Channel EEG Signal

Sleep is crucial for both mental and physical health. Poor sleep quality can lead to numerous health issues. Accurate detection of different sleep disorders is crucial for early diagnosis, treatment, and prognosis. An automated sleep disorder-based detection scheme is proposed using single-channel EEG signals for six classes of sleep disorders: naming, Insomnia, Nocturnal frontal lobe epilepsy (NFLE), Narcolepsy, REM behavior disorder (RBD), Periodic leg movement (PLM) and Sleep Disorder Breathing (SDB) are featured. Time and frequency domain features are introduced for the classification of sleep disorders using machine learning classifiers such as Random Forest, multilayer perceptron (MLP) and so on. P-ANOVA analysis is carried out to demonstrate the discrimination of the features for various classes of disorders. An extensive evaluation of the CAP (Cyclic Alternating Pattern) sleep database was carried out in two different cases. In Case 1, binary classification between healthy subjects and a specific disorder is carried out. Multiclass classification is performed in Case 2. Each classification is done over different sleep stages. For Case 1, the accuracy rates are 97% for insomnia, 90% for NFLE, 99% for RBD, 94% for narcolepsy, 94% for PLM, and 98% for SDB, with the corresponding F1 scores being 92%, 94%, 88%, 87%, 88%, and 93%, respectively. On the other hand, 86% accuracy and 86% F1 score are obtained for multistage classification. The experimentation also indicates that, in general, the best classification performance can be achieved using the EEG signal corresponding to sleep stage 4,(S4).