Data-driven topic analysis of high density EEG reveals concomitant superficial sleep during deep sleep in insomnia

Introduction: An intriguing problem in the evaluation of insomnia is that the subjective feeling of insomnia in many cases does not correlate with what is documented by polysomnography (PSG) as evaluated with standard sleep scoring. Although many resolutions to the misperception of sleep in insomnia have been proposed, only few questioned whether the ‘gold standard’ of sleep staging could be the main obstacle in the search for a robust marker of the disorder. Essentially, the assumptions that sleep staging is discrete and constant for predefined periods do not accommodate pathological brain signals. In current study, we used our previously validated data-driven sleep model (Koch et al. J. Neurosci. Methods 2014:235:130–7) that allows latent brain states (referred to as topics) to occur simultaneously within each epoch, in order to obtain a more realistic description of whether their separation and co-occurrence differ between people with insomnia and controls.

Materials and methods: We applied a validated data-driven topic model on sleep EEG of 55 people with Insomnia Disorder (ID, 14 males/41 females, 47.8±12.9 years) and 64 control subjects (CON, 22 males/42 females, 45.4±14.7 years). In the topic modeling approach, spectral EEG and EOG correlation measures in 1-second windows were calculated and used to define 3-second sleep structures. By counting the presence of the sleep structures, each sleep epoch was expressed as a mixture of probabilities of latent sleep states using a Bayesian topic model (Koch et al. J. Neurosci. Methods 2014:235:130–7). Subsequently, each sleep epoch was described independently as a probability distribution of six topics (topic T1-T6), where the dominant topic in an epoch was related but not equal to a classical sleep stage. Using the topic time series of the entire night, we computed the overall proportion of each topic. We calculated a measure of their temporal separation versus co-occurrence within 30-second epochs as follows: for each set of epochs where the dominant topic was the same for at least three epochs, we calculated the normalized proportion of presence among the remaining topics.

Results: Between-group comparisons revealed no significant differences for the overall proportions of topics, much like the often-reported lack of differences between cases and controls on classical measures like total time of sleep and sleep stages. However, in epochs where T1 (a N3-related topic) was dominant and stable, cases with ID showed a much higher co-occurrence of T4 (a N1-related topic). In these epochs, the mean co-occurrence proportion was 9.16% in ID compared to 4.42% in CON.

Conclusions: As compared to controls, N3 sleep in people with insomnia contains twice as much concurrent N1 sleep-related EEG signatures, suggesting continued hyperarousal even during their deepest sleep. Topic modeling is a powerful approach for analyzing concomitant vigilance states.

Acknowledgements: The ESRS Short-Term Research Fellowship 2017 awarded to J.A.E. Christensen supported this project.