Clifton Frilot II1, David E. McCarty2, Paul Y. Kim3, Andrew A. Marino4
1School of Allied Health Professions, LSU Health Sciences Center, Shreveport, LA
2Division of Sleep Medicine, Department of Neurology, LSU Health, Shreveport, LA
3Centenary College, Shreveport, LA
4ABR Analytics, Belcher, LA
Presented at SLEEP 2015, the 29th Annual Meeting of the Associated Professional Sleep Societies, Seattle, WA, June, 2015
Recurrence analysis (RA) is a new method for quantifying non-randomness in the EEG. Applied to the sleep-staged EEG (SS-EEG), RA yields 16 markers for stage-dependent sleep depth and fragmentation. We postulated that RA could accurately identify the presence of temporal information (“signature”) in the SS-EEG specifically associated with OSA, as assessed by RA’s capability to reliably identify the subjects in a cohort consisting of those who did and did not have OSA.
Five independent subcohorts were formed from a cohort of 266 subjects consisting of 115 without OSA (AHI < 5), 115 with mild OSA (AHI 5–15), 50 with moderate OSA (AHI 16–30), and 23 with severe OSA (AHI > 30). Staged PSGs were obtained from the National Sleep Research Resource. RA markers computed from the C3 EEG were analyzed by discriminant analysis to identify marker combinations that reliably classified individual subjects into No-OSA and OSA (AHI ≥ 5) groups (each N = 25). The OSA group composition was 40% mild, 40% moderate, 20% severe. Classification accuracy (CA) was assessed by comparing the classification result with ground truth.
In five successive independent analyses, an ABR-produced EEG biomarker function accurately identified the subjects with OSA; the respective accuracies were 82%, 82%, 86%, 88%, and 90%. Classification accuracy for 28 subjects who were not selected in any of the 5 subcohorts was 93%, as assessed using a combination of the five previously determined biomarker functions, indicating that new subjects could be correctly diagnosed based on a comparison of their sleep EEG with those in the parent cohort.
OSA subjects (AHI ≥ 5) were identified by RA, using only a single EEG lead. The SS-EEG contained a signature detectable by RA that allowed a reliable binary classification of the individual subjects in a cohort that included a group with no OSA and an equally-sized group of subjects who exhibited a wide range of disorder severity.