Severe obstructive sleep apnea event detection from EEG recordings

Sleep apnea is a serious disorder where breathing stops frequently during sleep. Changes in brain activities that occur during apnea can be detected with an electroencephalogram (EEG). Although accurate detection of apnea events is very important, there is currently no algorithm that can efficiently measure the onset and end of apnea events based only on electroencephalogram signals. The number and duration of apnea events are used to calculate apnea-hypopnea index (AHI) and mean apnea-hypopnea duration (MAD), that are indicators of obstructive sleep apnea severity. Previous apnea detection algorithms usually focus on the classification of apnea patients and not specific apnea events, or perform a frame-by-frame analysis and classify each frame based on the global characteristics of the frame, instead of locating the onsets and ends of apnea events. Thus, the clinical significance of EEG signals for apnea detection is limited to sleep staging. The purpose of this study is to propose a method for sleep apnea event detection and event duration evaluation using Convolutional Recurrent Neural Networks, based only on EEG signals. Reference and estimated AHI are strongly correlated (r=0.88, p<0.001), whereas the sensitivity and positive predicted value for the individual events detection is 0.73 and 0.78, respectively. Reference and estimated MAD values are very highly correlated (r=0.91, p<0.001), and the absolute error between them is 2.05 ± 1.66 s. The proposed method has high accuracy in detecting individual apnea events from EEG signals, especially in severe apnea cases.