Stimulus-Induced Narrowband Gamma Oscillations are Test-Retest Reliable in Human EEG

Abstract

Visual stimulus-induced gamma oscillations in electroencephalogram (EEG) recordings have been recently shown to be compromised in subjects with preclinical Alzheimer's Disease (AD), suggesting that gamma could be an inexpensive biomarker for AD diagnosis provided its characteristics remain consistent across multiple recordings. Previous magnetoencephalography studies in young subjects have reported consistent gamma power over recordings separated by a few weeks to months. Here, we assessed the consistency of stimulus-induced slow (20-35 Hz) and fast gamma (36-66 Hz) oscillations in subjects (n = 40) (age: 50-88 years) in EEG recordings separated by a year, and tested the consistency in the magnitude of gamma power, its temporal evolution and spectral profile. Gamma had distinct spectral/temporal characteristics across subjects, which remained consistent across recordings (average intraclass correlation of ~0.7). Alpha (8-12 Hz) and steady-state-visually evoked-potentials were also reliable. We further tested how EEG features can be used to identify 2 recordings as belonging to the same versus different subjects and found high classifier performance (AUC of ~0.89), with temporal evolution of slow gamma and spectral profile being most informative. These results suggest that EEG gamma oscillations are reliable across sessions separated over long durations and can also be a potential tool for subject identification.

Keywords: EEG; gamma oscillations; healthy aging; intersubject variability; test–retest reliability.

Figure 1

Gamma oscillations remain consistent across baseline and follow-up sessions. Change in time–frequency power spectra for the 2 recording sessions, baseline (left), and follow-up (right) of 20 female subjects to static gratings (“gamma” experiment). Subjects are ordered vertically based on the decreasing average slow and fast gamma power in visual electrodes, starting from the top left. The color bar on the top left denotes the log power ratio in dB units. The numbers indicate test–retest interval in days.

Figure 2

Same as Figure 1 for 20 male subjects.

Figure 3

Temporal evolution of band powers in slow gamma, fast gamma, and alpha range across baseline and follow-up remain consistent in males and females. Band power as a function of time, computed from averaging baseline corrected time–frequency spectra across relevant frequency bins for slow gamma, fast gamma, and alpha bands. Traces are plotted in black for baseline (Y0) and gray for follow-up (Y1) for 20 female and 20 male subjects. Same subject order as in previous plots.

Figure 4

Power spectral densities (PSDs) are consistent across baseline and follow-up in males and females. The change in PSD during stimulus period relative to baseline PSD for baseline (Y0; black) and follow-up (Y1; gray) for 20 female and 20 male subjects.

Figure 5

Change in slow gamma, fast gamma, alpha power and SSVEP at 32 Hz power are correlated across baseline and follow-up. (A)–(C) Scatter plots of the average change in power within slow gamma, fast gamma, and alpha bands for baseline and follow-up sessions. (D) The average change in power at SSVEP frequency at 32 Hz. Pearson’s correlation coefficient and P-value are mentioned on the top left and bottom right in each panel. Note the difference in axis limits across the plots.

Figure 6

Change in PSD profile and temporal profiles of band powers are more correlated within self-pairs than cross-pairs. (A)–(C) Correlation between the band power time series traces (shown in Fig. 3) computed between 2 sessions of the same subject (self-pair) plotted against the median correlation between traces for all other subjects (cross-pair), for slow gamma (A), fast gamma, (B) and alpha (C) bands. (D) Same plot for the change in PSD traces shown in Figure 4. There are 40 data points in each plot. The median of these data is shown by the black box with error bars indicating median absolute deviation (MAD).

Figure 7

Temporal slow gamma and spectral correlations emerge as dominant features in classification. (A) The 7 features provided as inputs for the subject distinctness analysis are plotted separately for cross-pairs (cross individual comparisons) and self-pairs (same individual comparisons). Change in power (first 3 features) are in units of dB. Temporal and spectral correlations (other 4 features) are unitless. (B) Z-scored features are plotted in the same way as panel (A). The values mentioned in the figure insets denote the classifier weight of the respective feature in separating out self-pairs from the cross-pairs.

Figure 8

Features from stimulus period perform better than resting state features. The AUC values obtained using various classifiers, constructed based on the feature set listed on the x-axis. First, the performance of each of the stimulus-induced features is considered individually and subsequently collectively, and then the baseline (resting state) features are considered, first individually and subsequently collectively, and finally combined with the stimulus-induced features.