Exploring the neuromagnetic signatures of cognitive decline from mild cognitive impairment to Alzheimer's disease dementia

Abstract

Background: Developing non-invasive and affordable biomarkers to detect Alzheimer's disease (AD) at a prodromal stage is essential, particularly in the context of new disease-modifying therapies. Mild cognitive impairment (MCI) is a critical stage preceding dementia, but not all patients with MCI will progress to AD. This study explores the potential of magnetoencephalography (MEG) to predict cognitive decline from MCI to AD dementia.

Methods: We analysed resting-state MEG data from the BioFIND dataset including 117 patients with MCI among whom 64 developed AD dementia (AD progression), while 53 remained cognitively stable (stable MCI), using spectral analysis. Logistic regression models estimated the additive explanation of selected clinical, MEG, and MRI variables for AD progression risk. We then built a high-dimensional classification model to combine all modalities and variables of interest.

Findings: MEG 16-38Hz spectral power, particularly over parieto-occipital magnetometers, was significantly reduced in the AD progression group. In logistic regression models, decreased MEG 16-38Hz spectral power and reduced hippocampal volume/total grey matter ratio on MRI were independently linked to higher AD progression risk. The data-driven classification model confirmed, among other factors, the complementary information of MEG covariance (AUC = 0.74, SD = 0.13) and MRI cortical volumes (AUC = 0.77, SD = 0.14) to predict AD progression. Combining all inputs led to markedly improved classification scores (AUC = 0.81, SD = 0.12).

Interpretation: These findings highlight the potential of spectral power and covariance as robust non-invasive electrophysiological biomarkers to predict AD progression, complementing other diagnostic measures, including cognitive scores and MRI.

Funding: This work was supported by: Fondation pour la Recherche Médicale (grant FDM202106013579).

Keywords: Alzheimer's disease (AD); Brain rhythms; Disease progression; Magnetoencephalography (MEG); Mild cognitive impairment (MCI); Spectral power.

Fig. 1

AD progression was characterised by altered spectral power and covariance profiles. Analysis was based on all n = 117 cases. (a) Power spectra averaged over all sensors: upon visual inspection, AD progression was associated with reduced spectral power at baseline in frequencies ranging from 10Hz to 40Hz. (b) Mean spectral power difference between groups (blue line) and 95% confidence interval computed by bootstrap (blue shaded area). (c) TFCE permutation F-test (two-tailed) on mean spectral power, showing significant power difference between AD progression and stable MCI at frequencies ranging from 16Hz to 39.4Hz [T₂: p < 0.05]. The black line represents F values [T₂], the orange line shows the TFCE p-value [T₂]. The shaded green region indicates frequencies with T₂: p < 0.05. (d) Comparison between groups based on covariance using Riemannian-distance MANOVA with dimensionality reduced to 65 components. The black line represents pseudo-F scores [T₄], the orange line the p-value [T₄] corrected using the False Discovery Rate (FDR). The shaded green region indicates frequencies with T₄: p < 0.05. Specifically, significant differences were identified between 1Hz to 4Hz and 10Hz–64Hz (all p-values <0.05). After applying the FDR correction, significant differences remained for frequencies between 1Hz and 2.8Hz, 10.2Hz–10.9Hz, 12.6Hz–64Hz, as shown by the orange indicators. (e) Topographical maps of spectral power difference between groups, showing reduced spectral power over posterior sensors in AD progression group in frequencies from 9.5Hz to 61.8Hz. The white dots indicate significant differences [T₃: p < 0.05]. For simplicity, here we only depict topographies at least including five significant sensors. This result has been obtained by TFCE spatio-temporal cluster permutation test (two-tailed).

Fig. 2

MEG spectral power, MRI, and MMSE offered additive explanations of AD progression. Analysis was based on the subset of n = 104 containing both MEG and MRI. (a) ROC curves of four logistic regression models to predict progression to AD dementia. (b) Marginal effects display of logistic regression model of risk of progressing to AD dementia using the following covariates: MEG 16–38Hz spectral power in parieto-occipital sensors, Hippocampus/Total grey matter ratio, MMSE, education and age. Higher values of MEG 16–38Hz spectral power in left parieto-occipital sensors, higher Hippocampus/Total grey matter ratio and higher MMSE were significantly associated with a reduced risk of progression to AD dementia conditional on all other variables. A higher level of education showed weaker effects in increasing the probability of progression to AD dementia.

Fig. 3

AD-progression showed altered frequency-architecture in non-instantaneous power correlations. Analysis was based on all n = 117 cases. (a) Average power-envelope correlation over all sensors. AD progression was visually associated with increased correlation below 4Hz and decreased correlation around 8Hz. (b) Power envelope correlation difference between groups (blue line) and 95% confidence interval computed by bootstrap (blue shaded area). Testing for differences in average power-envelope correlations revealed significant differences [T₁ < 0.05 uncorrected] at the following frequencies: 1.2Hz, 1.4Hz–1.7Hz, 3.7Hz–4Hz, 8Hz–9.5Hz, 52Hz–64Hz. (c) TFCE permutation F-test (two-tailed) on average power-envelope correlations pointed at a non-significant cluster between 8.3Hz and 8.9Hz [T₂: p < 0.25] for AD progression versus stable MCI. (d) Multivariate detection of group differences in power-envelope correlation matrices between using Riemannian-distance MANOVA. The black line represents pseudo-F scores [T₄], the orange line the p-value [T₄] corrected using the False Discovery Rate (FDR). The shaded green region indicates frequencies with T₄: p < 0.03. The analysis revealed significant differences in power envelope metrics across frequencies ranging from 2.3Hz to 64Hz, suggesting wide-ranging differences in the power-envelope correlation architecture across frequencies (e) Topographical maps of power envelope correlation differences between groups, shown at the five frequencies listed in panel (b). AD progression showed a pattern of decreased power envelope in alpha frequency-range in posterior regions, which did not reach statistical significance.

Fig. 4

AD-progression showed narrow-band alteration of phase interactions. Analysis was based on all n = 117 cases. (a) Average dwPLI over all sensors: AD progression was visually associated with reduced dwPLI around 8–10Hz. (b) dwPLI difference between groups (blue line) and 95% confidence interval computed by bootstrap (blue shaded area). Testing for differences in average power-envelope correlations revealed significant differences [T₁ < 0.05 uncorrected] at the following frequencies: 2Hz, 2.7Hz, 3.1–3.5Hz, 3.9–4.1Hz, 6.3–7.2Hz, 8.6–9.2Hz, 19–19.7Hz, 27.9Hz. (c) TFCE permutation F-test (two-tailed) on average power-envelope correlations pointed at a non-significant cluster between 6.7Hz and 7Hz [T₂: p < 0.25]. (d) Multivariate detection of group differences in dwPLI matrices between using Riemannian-distance MANOVA. The black line represents pseudo-F scores [T₄], the orange line the p-value [T₄] corrected using the False Discovery Rate (FDR). The shaded green regions indicate frequencies with T₄: p < 0.05. Significant differences were identified at the following frequencies: 2.1Hz, 6.5–7Hz, 8.9–10.2Hz, 19–20.4Hz, and 43.7–48.5Hz. After FDR correction, significant differences remained for frequencies between 9.2–9.8Hz and 45.3–46.9Hz (orange indicators). (e) Topographical maps of dwPLI differences between groups, shown at the seven frequencies listed in panel (b). In posterior brain regions, AD progression showed a pattern of decreased dwPLI in alpha band and increased dwPLI in theta band (not statistically significant).

Fig. 5

Data-driven classification model of AD progression. Analysis was based on all n = 117 cases. Panel (a) presents average cross-validated AUC scores, ordered by median scores. Error bars representing the 95% confidence intervals of the median over 100 cross-validation (CV) iterations. For marginal models using single demographic variables or the high-dimensional brain-data as inputs (MEG: Covariance, power envelope and dwPLI matrices along the frequency spectrum, MRI: Freesurfer cortical volumes alongside the full model combining all inputs using stacking (cf. Table 5 for details). One can see that brain-based models were ranked higher than single demographic predictors. The best result was obtained for the full model. Panel (b) presents average conditional permutation importance (CPI) scores capturing the change in the loss function upon removing the unique information given by a variable not shared with the other variables, ordered by median score. Error bars represent the 95% confidence intervals of the median over 100 cross-validation (CV) iterations. P-values referring to testing with pseudo t-statistic [T₅] with Nadeau's and Bengio's corrected t-test. The results suggest that the performance of the full model was based on statistically complementary information from the MEG covariance, MMSE and MRI but also highlighted potential contributions from site effects.