No document available.
Abstract :
[en] Arousals during sleep are transient accelerations of the EEG signal typically detected by visual inspection of the sleep recording. Such visual identification is a time-consuming, subjective process that prevents comparability across scorers, studies and research centres. We developed an algorithm, which automatically detects arousals in whole-night EEG recordings, based on time and frequency analysis with adapted thresholds derived from individual data.
We performed automatic arousals detection over 35 sleep recordings of young (µ=24.07±3, N=18) and older (µ=61.38±6, N=17) healthy individuals, and compared it against human raters (HR) detection from two research centres. We assessed performance of the automatic algorithm using generalized linear mixed models with Cohen’s kappa as dependent variable. Performance of automatic detection was compared to a gold standard, composed of either all arousals found by any of the HR (inclusive detection – ID) or only those common to both HR (conservative detection – CD).
Comparison between human scorers revealed a high variability in the number of arousals detected (µ=71±32 vs 111±50). Although many more arousals were automatically detected (µ=200 ± 43), agreement of automatic detection against human detection was high, as reflected by very large Cohen’s kappa values (κ=.93 for ID, .94 for CD). Importantly, automatic detection was correlated to human detection (r=.38, p=.025 for CD). Algorithm performance was not significantly influenced by sleep stage (p=.74 for ID; p=.97 for CD), age (p=.12 for ID; p=.91 for CD) or sex (p=.10 for ID; p=.21 for CD). We further found that relative power in the theta and alpha bands were, respectively, higher and lower (p<.0001) for arousals that were only detected by the algorithm, arguably making them less obvious for the human eye.
Our results show that the automated algorithm is performing at least equally as well as HR. While the automatic method detects most of HR events, it finds many more events that bear the characteristics of AASM arousals, but are missed by visual inspection of the EEG. This is seen for other micro-events detectors such as spindle detectors. In conclusion, our algorithm a reliable tool for automatic detection of arousals.
