Evidence of Neurovascular Un-Coupling in Mild Alzheimer’s Disease through Multimodal EEG-fNIRS and Multivariate Analysis of Resting-State Data

Figures

• 

  Figure 1. (a) Electroencephalography and functional near-infrared spectroscopy (EEG-fNIRS) probes located on a participant’s head. The fNIRS fiber heads reached the scalp using the spaces between the electrodes of the EEG cap. (b) EEG-fNIRS probes overlaid on a template head. The high-density (128-channel) EEG layout was arranged in agreement with the 10/20 system and the fNIRS probes were positioned with reference to the EEG sensors. (c) Channels’ average sensitivity map of the fNIRS measurements. The sensitivity map was computed using the finite element method to solve the diffusion equation [67] and it is displayed up to an attenuation of 60 dB from its maximum value.

• 

  Figure 2. (a) Pre-processing of the EEG involving band-pass filtering, independent component analysis (ICA) motion correction and power envelope computation. The end results of the EEG pre-processing were artifact-corrected EEG power envelopes for the 128 channels in the three frequency bands of interest (theta, alpha and beta). (b) Pre-processing of the fNIRS involving Optical Densities (OD) computation, wavelet-based motion correction and oxy-hemoglobin (HbO) and deoxy-hemoglobin (HbR) computation. The end results of the fNIRS pre-processing were artifact-corrected HbO and HbR changes in both the long and the short separation channels.

• 

  Figure 3. Algorithm for neurovascular coupling (NC) estimation from EEG-fNIRS recordings using a general linear model (GLM) framework. The independent variables of the GLM were the expected hemodynamic responses (z-scored, 1 Hz resampled) obtained from the convolution of the global EEG power envelopes with the canonical hemodynamic response function (HRF). The global EEG power envelopes were obtained by selecting the 1st principal components (PCs) of bandwise principal component analyses (PCAs). Control for scalp contamination was obtained by considering hemoglobin oscillations in the short separation fNIRS channels as additional independent variables (z-scored, 1 Hz resampled). The dependent variable was either HbO or HbR signal in each long separation fNIRS channel (z-scored, 1 Hz resampled). The analysis provided standardized β-weights depicting the strength of association (NC) between bandwise global EEG power envelopes and HbO and HbR in each long separation channel considered.

• 

  Figure 4. Distributions of the GLM standardized β-weights for the fNIRS long separation channels depicting the linkage of the 1st PCs of the three EEG power envelopes in theta (θ), alpha (α) and beta (β) bands with HbO and HbR changes in AD and HC. The different colors represent different subjects.

• 

  Figure 5. Boxplots showing the average standardized β-weights among the long separation fNIRS channels (global NCs) for AD and HC. (a) Global NC of theta, alpha and beta bands with HbO. (b) Global NC of theta, alpha and beta bands with HbR (* p < 0.05).

• 

  Figure 6. Boxplots showing the EEG and fNIRS average power for AD and HC. (a) Global EEG power in theta (θ), alpha (α) and beta (β) bands. (b) Global fNIRS power in HbO and HbR timecourses.

• 

  Figure 7. Cross-validation (CV) classification outcome of the multivariate data-driven analysis predicting AD and HC from the 6 NC metrics. (a) Receiver operating characteristic (ROC) analysis. (b) Upper panel: model loadings; lower panel: z-scores obtained as loadings normalized by a bootstrap estimate of their standard deviation.

• 

  Figure 8. CV classification outcome of the multivariate data-driven analysis predicting AD and HC from the standalone EEG and fNIRS metrics. (a) ROC analysis employing the 3 global EEG powers. (b) ROC analysis employing the 2 global fNIRS hemoglobin powers.

• 

  Figure 9. Cross-validated (CV) loadings and classification performance of the multivariate data-driven analysis predicting AD and HC from the 6 NC metrics as a function of either recording time or number of EEG channels employed. (a) Left panel: loadings as a function of recording time; right panel: z-scores, obtained as loadings normalized by their standard deviation (estimated using bootstrapping), as a function of recording time. (b) Area under the curve (AUC) of the ROC analysis as a function of recording time. (c) Left panel: loadings as a function of the number of randomly selected EEG channels; right panel: z-scores, obtained as loadings normalized by their standard deviation, as a function of the number of randomly selected EEG channels. (d) AUC of the ROC analysis as a function of the number of randomly selected EEG channels.