The sleep EEG envelope is a novel, neuronal firing-based human biomarker

Abstract

Sleep EEG reflects voltage differences relative to a reference, while its spectrum reflects its composition of various frequencies. In contrast, the envelope of the sleep EEG reflects the instantaneous amplitude of oscillations, while its spectrum reflects the rhythmicity of the occurrence of these oscillations. The sleep EEG spectrum is known to relate to demographic, psychological and clinical characteristics, but the envelope spectrum has been rarely studied. In study 1, we demonstrate in human invasive data from cortex-penetrating microelectrodes and subdural grids that the sleep EEG envelope spectrum reflects neuronal firing. In study 2, we demonstrate that the scalp EEG envelope spectrum is stable within individuals. A multivariate learning algorithm could predict age (r = 0.6) and sex (r = 0.5) from the EEG envelope spectrum. With age, oscillations shifted from a 4-5 s rhythm to faster rhythms. Our results demonstrate that the sleep envelope spectrum is a promising biomarker of demographic and disease-related phenotypes.

Figure 1

The principle of EEG envelope spectrum analysis. (A) Shows a simulated EEG signal, consisting of the sum of a 2 Hz sinusoid modulated by a 0.2 Hz carrier frequency, a 12 Hz sinusoid modulated by a 1 Hz carrier frequency, and pink noise. Overlain blue and red lines show the instantaneous amplitude or envelope (modulus of the Hilbert transform) of the delta (1–4 Hz) and sigma (10–16 Hz) frequency ranges, respectively. (B) Shows the power spectral density of the original signal (left) and the delta (middle) and sigma (right) envelopes. Note that the carrier frequencies are accurately recovered from spectral analysis of the envelopes (with some impurities due to added noise and the fact that the modulus of the Hilbert transform of a modulated signal is not fully sinusoidal). The spectrum of the envelope reveals periodic fluctuations in the amplitude of higher-frequency activities.

Figure 2

Coupling between EEG envelope in the cortical surface and MUA within the adjacent cortex in NREM sleep. (A) Cross-correlation of the two signals. The horizontal axis indicates time lags, the vertical axis indicates IME channel (N = 23, deeper channels shown at the bottom), while the color axis indicates correlation coefficients. Black outlines show statistically significant results after FDR correction. (B) Magnitude-squared coherence between the two signals. Overlain lines represent individual IME channels. Because of the large number of channels and no substantial between-channel differences, no particular pattern in color coding was used. Dots indicate statistical significance after FDR correction on the corresponding channel. Deeper channels are shown at the top. (C) Mean MUA amplitude (in within-segment z-scores) by ECoG envelope phase bins. Dots indicate statistical significance after FDR correction on the corresponding channel. A sinusoid is overlain in the low delta subplot for illustration. On (B,C), for better visibility only IME channels are shown where at least one data point reached significance.

Figure 3

An illustration of EEG envelopes, the colliding window method and its results. (A) Illustrates the colliding window method. (B) Shows a single epoch of illustrative envelope and MUA data (ECoG low delta envelope and smoothed MUA from the fifth IME channel located in cortical layer III). The Pearson correlation of the two signals is shown for reference. Both the ECoG envelope and the MUA is detrended and demeaned, but not z-transformed. (C) Shows the distribution of available sleep data after artifact rejection using the colliding window method. For each participant, black lines mark the data segments used in analysis. The lower panel shows the total number of participants with available data as a function of time after recording start. Note the lack of systematic undersampling of any part of the night. (D) Illustrates the log-transformed envelope spectra. All data was z-transformed by frequency band to eliminate mean differences. The frequency axis is shown on a log scale to enhance the low frequency ranges which are of particular interest. Note spectral peaks at ~ 0.05–0.06 Hz, ~ 0.25 Hz and ~ 1 Hz, the latter most prominent in the beta range.

Figure 4

The reliability of the sleep EEG envelope spectrum. (A) Shows raincloud plots) by vigilance state and reliability type, showing raw data overlain with box plots on the left side and kernel density curves on the right side. Data from all frequency bands, envelope frequency bins and scalp channels are pooled for estimating the box plot, while individual instances are shown as data points. (B) Illustrates reliability by envelope frequency bin. Data from all frequency bands and scalp channels are pooled, shading indicates 95% confidence intervals of the mean.

Figure 5

Correlations between the NREM envelope spectrum and age, sex and general cognitive ability (IQ). Colored lines represent correlation coefficients by scalp channel. Color codes indicate scalp region, with individual channels from the same region shown with the same color. Black horizontal lines show the threshold of conventional (p = 0.05) significance. Colored dots (with color coding identical to lines) above the lines indicate a statistically significant correlation after FDR correction on the corresponding channel.

Figure 6

Correlations between the REM envelope spectrum and age, sex and general cognitive ability (IQ). Colored lines represent correlation coefficients by scalp channel. Color codes indicate scalp region, with individual channels from the same region shown with the same color. Black horizontal lines show the threshold of conventional (p = 0.05) significance. Colored dots (with color coding identical to lines) above the lines indicate a statistically significant correlation after FDR correction on the corresponding channel.

Figure 7

The performance of elastic net regression models predicting age, sex and IQ from the envelope spectrum. Topographic plots illustrate the correlation between predicted and actual phenotypes in the validation sample. (Elastic net regression models were run separately for each channel). The correlation for channels on which the elastic net model did not converge is set to 0 and not counted towards the average performance described in the text.