Data-Driven Analysis of EEG Reveals Concomitant Superficial Sleep During Deep Sleep in Insomnia Disorder

Abstract

Study Objectives: The subjective suffering of people with Insomnia Disorder (ID) is insufficiently accounted for by traditional sleep classification, which presumes a strict sequential occurrence of global brain states. Recent studies challenged this presumption by showing concurrent sleep- and wake-type neuronal activity. We hypothesized enhanced co-occurrence of diverging EEG vigilance signatures during sleep in ID. Methods: Electroencephalography (EEG) in 55 cases with ID and 64 controls without sleep complaints was subjected to a Latent Dirichlet Allocation topic model describing each 30 s epoch as a mixture of six vigilance states called Topics (T), ranked from N3-related T1 and T2 to wakefulness-related T6. For each stable epoch we determined topic dominance (the probability of the most likely topic), topic co-occurrence (the probability of the remaining topics), and epoch-to-epoch transition probabilities. Results: In stable epochs where the N1-related T4 was dominant, T4 was more dominant in ID than in controls, and patients showed an almost doubled co-occurrence of T4 during epochs where the N3-related T1 was dominant. Furthermore, patients had a higher probability of switching from T1- to T4-dominated epochs, at the cost of switching to N3-related T2-dominated epochs, and a higher probability of switching from N2-related T3- to wakefulness-related T6-dominated epochs. Conclusion: Even during their deepest sleep, the EEG of people with ID express more N1-related vigilance signatures than good sleepers do. People with ID are moreover more likely to switch from deep to light sleep and from N2 sleep to wakefulness. The findings suggest that hyperarousal never rests in ID.

Keywords: data-driven analysis; indiscrete labeling of sleep; insomnia; latent Dirichlet allocation; polysomnography; topic modeling; vigilance states.

Figure 1

Symbolization of raw EEG for the purpose of Latent Dirichlet Allocation (LDA) topic modeling. The raw data of two EEG signals of which one is depicted in (A), were bandpass filtered into the classical frequency bands (δ, θ, α, and β). (B) One bandpass filtered signal is shown. (C) The power in each 1 s window was calculated and summarized in a distribution across the entire recording. (D) According to quintiles of this distribution, the power in each 1 s window was categorized as either “Very low,” “Low,” “Median,” “High,” or “Extreme,” effectively creating a vector of letters. (E) For each 3 s sliding window, three consecutive letters were concatenated to create words indicating the spectral power-level development. (F) Finally, a word distribution was created for each 30 s epoch by counting how often each of the possible words occurred (5 categories and 3 letters: 5³ = 125 words per frequency-band per EEG channel). In addition to the EEG words, an additional 192 words were created in a similar fashion [4 categories and 3 letters: 4³ = 64 words per EOG signal (left and right) plus 64 words for the cross-correlation between the EOG signals].

Figure 2

Estimation of concurrent vigilance states in each 30 s epoch. The combined word distributions for each 30 s epoch were inserted into a Latent Dirichlet Allocation (LDA) model, which returns a mixture of topic probabilities for each epoch. The LDA topic model was previously trained to learn the particular distribution of words for six topics (Koch et al., 2014). (A) The observed word distributions of one 30 s epoch is inserted into the LDA topic model. (B) The observed word distributions of the 30 s epoch are depicted in transparent gray bars, and the expected word distributions for each topic in colored bars. (C) By comparing the observed and expected word distributions, the LDA topic model returns the probability that the observed word distributions comes from the word distribution of each topic. Each 30 s epoch is thereby represented as a mixture of six topics. (D) A topic diagram obtained by repeating this procedure for each 30 s epoch in the recording. Each vertical bin in the topic diagram is a mixture of six colors, where the hight of each stacked color represents the probability of a topic. Colors codes range from dark blue (T1) to red (T6). (E) For comparison, the manual scored hypnogram is presented below the topic-diagram.

Figure 3

Examples of topic diagrams. Topic diagrams of two normal sleepers (A,B) and two insomniac patients (C,D). Each 30 s sleep epoch is represented as a multi-colored vertical bin where the heigth of each stacked color presents the probability that the topic is present. For comparison, the manual scored hypnograms are presented below each diagram. Color codes range from dark blue (T1, mostly seen in deep N3 sleep) to red (T6, mostly seen in wakefulness). Orange (T5) is mostly seen during REM sleep. Note, however, that there is no one-to-one matching of topics and conventional top-down defined qualitative sleep stages. Subtle differences might be seen comparing the two controls (A,B) and two cases with insomnia (C,D): In the epochs dominated by the light sleep related topic (yellow), the light sleep related topic is generally stronger (taller yellow bars) for the cases with insomnia as compared to controls.

Figure 4

Light sleep-related EEG signatures are more abundant during light and deep sleep in ID than in controls. (A) Violin plots for each group (horizontal axis) of the probability of light sleep-related topic T4 when this is stable and dominant (vertical axis). (B) Normalized co-occurrence of topic T4 when deep sleep-related topic T1 is stable and dominant (vertical axis).

Figure 5

Increased probability for transitions from deep sleep to light sleep in ID as compared to controls. Markovian state diagram for topic transitions. The red arrow indicates a higher probability for participants with ID compared to controls, whereas the blue arrow indicates a lower probability for participants with ID compared to controls. The topics are indicated with colors and arranged in a way so the vertical axis denotes deepness of sleep.