Reproducibility of EEG functional connectivity in Alzheimer's disease

Abstract

Background: Although numerous electroencephalogram (EEG) studies have described differences in functional connectivity in Alzheimer's disease (AD) compared to healthy subjects, there is no general consensus on the methodology of estimating functional connectivity in AD. Inconsistent results are reported due to multiple methodological factors such as diagnostic criteria, small sample sizes and the use of functional connectivity measures sensitive to volume conduction. We aimed to investigate the reproducibility of the disease-associated effects described by commonly used functional connectivity measures with respect to the amyloid, tau and neurodegeneration (A/T/N) criteria.

Methods: Eyes-closed task-free 21-channel EEG was used from patients with probable AD and subjective cognitive decline (SCD), to form two cohorts. Artefact-free epochs were visually selected and several functional connectivity measures (AEC(-c), coherence, imaginary coherence, PLV, PLI, wPLI) were estimated in five frequency bands. Functional connectivity was compared between diagnoses using AN(C)OVA models correcting for sex, age and, additionally, relative power of the frequency band. Another model predicted the Mini-Mental State Exam (MMSE) score of AD patients by functional connectivity estimates. The analysis was repeated in a subpopulation fulfilling the A/T/N criteria, after correction for influencing factors. The analyses were repeated in the second cohort.

Results: Two large cohorts were formed (SCD/AD; n = 197/214 and n = 202/196). Reproducible effects were found for the AEC-c in the alpha and beta frequency bands (p = 6.20 × 10^-7, Cohen's d = - 0.53; p = 5.78 × 10^-4, d = - 0.37) and PLI and wPLI in the theta band (p = 3.81 × 10^-8, d = 0.59; p = 1.62 × 10^-8, d = 0.60, respectively). Only effects of the AEC-c remained significant after statistical correction for the relative power of the selected bandwidth. In addition, alpha band AEC-c correlated with disease severity represented by MMSE score.

Conclusion: The choice of functional connectivity measure and frequency band can have a large impact on the outcome of EEG studies in AD. Our results indicate that in the alpha and beta frequency bands, the effects measured by the AEC-c are reproducible and the most valid in terms of influencing factors, correlation with disease severity and preferable properties such as correction for volume conduction. Phase-based measures with correction for volume conduction, such as the PLI, showed reproducible effects in the theta frequency band.

Keywords: Alzheimer’s disease; EEG; Functional connectivity; Reproducibility.

Fig. 1

Summary of observed differences in ANOVA model 1, shown as effect size (Cohen’s d), between AD and SCD subjects for each of the functional connectivity measures and bandwidths. The significant effect sizes of the comparisons made in the entire cohort 1 (n = 411) are shown. Effects that could not be reproduced in the subset populations or cohort 2 were left out. Red blocks represent a higher and blue blocks a lower level of functional connectivity in AD subjects compared to SCD subjects. The size of the blocks and the number shown in the blocks represent the size of the effect. Results of functional measures susceptible to signal leakage are shaded in grey

Fig. 2

The difference in functional connectivity between SCD and AD subjects per bandwidth is shown for each cohort and subpopulation. Significant effect sizes are shown in Cohen’s d estimated by GLM model 1. Cohort 1: all SCD and AD subjects in cohort 1. Cohort 2: all SCD and AD subjects in cohort 2. Cohort 1 A/T: amyloid-negative/tau-negative SCD versus amyloid-positive/tau-positive AD subjects from cohort 1 (subpopulation 1). Cohort 1 A/T/N: amyloid-negative/tau-negative/MTA < 1 SCD versus amyloid-positive/tau-positive/MTA ≥ 1 AD subjects from cohort 1, excluding any patients with Fazekas > 1 and any potential interfering medication (subpopulation 2)

Fig. 3

Correlation of different functional connectivity measures with the relative power in the corresponding frequency band. The functional connectivity measures are shown on the x-axis and the band power on the y-axis. Only significant correlations are shown, indicated by the correlation coefficient (r) which is also indicated by the colour gradient from r = − 1 (dark red) to r = + 1 (dark blue)

Fig. 4

Summary of observed differences in ANOVA model 2, shown as effect size (Cohen’s d), between AD and SCD subjects for each of the functional connectivity measures and bandwidth. The significant effect sizes of the comparisons made in the entire cohort 1 (n = 411) are shown. Effects that could not be reproduced in the subset populations or cohort 2 were left out. Red blocks represent a higher and blue blocks a lower level of functional connectivity in AD subjects compared to SCD subjects. The size of the blocks and the number shown in the blocks represent the size of the effect. Results of functional measures susceptible to signal leakage are shaded in grey