The detection of cortical arousals from Electroenchelaphalagraphy (EEG) plays an important role in determining sleep quality of a person, as well is used in determining the severity of several sleep disorders, such as sleep apnea and periodic limb movements. Manually scoring arousal events in EEG signals is time consuming. Automating the scoring of cortical arousals would save time for sleep technicians and improve the performance of other automatic sleep analysis.
The current criteria to score cortical arousals, by the American Academy of Sleep Medicine (AASM), is to detect abrupt shifts in EEG frequency lasting at least 3 seconds preceded by at least 10 seconds of stable sleep. In this paper a feed-forward artificial neural network (ANN) was used to classify arousals using a single EEG lead. The signal power of the EEG lead at certain frequency bands on 5 second long segments of the C4M1 EEG lead was used as the input into the ANN. During training the data set was sub-sampled such that arousal periods and non-arousal periods had an equal representation.
The ANN was trained on 435483 segments then validated on 108871 unseen segments that were extracted from the data supplied by Physionet 2018. The performance was then evaluated using area under PR (precision recall) and ROC (receiver operating characteristic) curves. With the respective scores 0.58 and 0.53 on the unseen validation data.
Scoring arousals in EEG signals is a labour intensive task, where a subjective rule is applied to determine if a section of the EEG signal contains an arousal or not. The placement of the manually scored arousal events on the signal is not always accurate. These two facts make the automatic detection of arousals from EEG signals particularly challenging.