This work presents a novel approach for detection of (non-apnea based) sleep arousals in a large sized polysomnographic dataset. Existing literature (and our preliminary analysis) indicates that sleep arousals can be classified, with a reasonable degree of accuracy, by training a classifier on features extracted from EEG channels. However, it is difficult to differentiate between apnea and non-apnea arousals using EEG information alone.
We propose a two stage classification approach to discriminate between apnea and non-apnea based arousals. During the first stage, we derive spatio-temporal features from the EEG channels using the dynamic mode decomposition. Dynamic mode decomposition (DMD) is a data-driven dimensionality reduction technique, originally used in fluid mechanics and features derived from DMD have been successfully used to detect sleep spindles and sleep stage classification from electrocorticographic data. The proposed first stage classifier employs spatio-temporal mode powers in EEG channels to detect arousals.
After arousal detection, a dedicated classifier is employed to localize and discard arousals resulting from apnea events. This classifier employs features extracted from the envelope of the airflow signal to identify apneas events; which correspond to regions where the airflow envelope exhibits values substantially lower than it would under normal breathing. Since, the airflow signal tends to suffer from breathing artifacts, therefore, a modified version of the (classical) Hilbert transform based envelope detector is employed for envelope extraction. This, modified, detector replaces the low-pass-filter at the end of a Hilbert detector with a cascade of a standard median filter and a recursive median filter. This suppresses large amplitude jumps while preserving edge information thus enhancing immunity to breathing artifacts and irregularities. Using this approach, we were able to achieve AUROC=0.731 on our validation dataset.