Classifying sleep-wake states of patients by training on single EEG or EOG channel data from normal subjects

Abstract

Sleep-wake patterns are disrupted in patients with sleep disorders, and automated identification of sleep and wake states will be extremely useful for the timely diagnosis of various sleep disorders. In this work, by utilizing only a single EOG channel, we are able to accurately classify the sleep-wake states of both normal and pathological subjects. Data from 27 patients and 20 normal subjects are employed from the DRMS-PAT and DRMS-SUB databases, respectively. We utilized ensemble empirical mode decomposition and Hilbert Huang transform to derive various statistical features from the intrinsic mode functions. We also used relative bandpowers and approximate entropy of the EEG/EOG signal as features to the classifier. To deal with the problem of class-imbalance, we considered random undersampling with boosting technique for the ensemble of trees as the classifier. Interestingly, using our chosen features, the model trained only on the data from normal subjects generalizes well to be able to classify patients' data with an average accuracy of 86.5% using EEG. EOG is able to achieve an overall accuracy of only 73.7% in this mode. However, in the crossvalidation experiments, a single EOG channel obtains marginally better (93.9 %) classification performance than a single EEG channel (93.6 %) on the patients' dataset. For normal subjects, 10-fold crossvalidation accuracy of 96.20% is achieved using the Cz-Al EEG channel, and 94.97 % using the EOG1 channel. Clinical relevance— Information about the sleep-wake pat-terns are extremely crucial for the detection of sleep disorders. An automated sleep-wake classification method can provide a faster and objective way of sleep scoring. It can assist the clinicians in the diagnosis of sleep disorders like dyssomnia or sleep apnea-hypopnea syndrome. Further it can help in the effective treatment and prognosis of comatose patients.