The Fingerprint of Rapid Eye Movement: Its Algorithmic Detection in the Sleep Electroencephalogram Using a Single...

The strong associations of rapid eye movement (REM) sleep with dreaming and memory consolidation imply the existence of REM-specific brain electrical activity, notwithstanding the visual similarity of the electroencephalograms (EEGs) in REM and wake states. Our goal was to detect REM sleep by means of algorithmic analysis of the EEG. We postulated that novel depth and fragmentation variables, defined in relation to temporal changes in the signal (recurrences), could be statistically combined to allow disambiguation of REM epochs. The cohorts studied were consecutive patients with obstructive sleep apnea (OSA) recruited from a sleep medicine clinic, and clinically normal participants selected randomly from a national database (N = 20 in each cohort). Individual discriminant analyses were performed, for each subject based on 4 recurrence biomarkers, and used to classify every 30-second epoch in the subject's overnight polysomnogram as REM or NotREM (wake or any non-REM sleep stage), using standard clinical staging as ground truth. The primary outcome variable was the accuracy of algorithmic REM classification. Average accuracies of 90% and 87% (initial and cross-validation analyses) were achieved in the OSA cohort; corresponding results in the normal cohort were 87% and 85%. Analysis of brain recurrence allowed identification of REM sleep, disambiguated from wake and all other stages, using only a single EEG lead, in subjects with or without OSA.