Beyond Focal Lesions: Dynamical Network Effects of White Matter Hyperintensities

Abstract

White matter (WM) tracts shape the brain's dynamical activity and their damage (e.g., white matter hyperintensities, WMH) yields relevant functional alterations, ultimately leading to cognitive symptoms. The mechanisms linking the structural damage caused by WMH to the arising alterations of brain dynamics is currently unknown. To estimate the impact of WMH on brain dynamics, we combine neural-mass whole-brain modeling with a virtual-lesioning (disconnectome) approach informed by empirical data. We account for the heterogeneous effects of WMH either on inter-regional communication (i.e., edges) or on dynamics (i.e., nodes) and create models of their local versus global, and edge versus nodal effects using a large fMRI dataset comprising 188 non-demented individuals (120 cognitively normal, 68 with mild cognitive impairment) with varying degrees of WMH. We show that, although WMH mainly determine local damage to specific WM tracts, these lesions yield relevant global dynamical effects by reducing the overall synchronization of the brain through a reduction of global coupling. Alterations of local nodal dynamics through disconnections are less relevant and present only at later stages of WMH damage. Exploratory analyses suggest that education might play a beneficial role in counteracting the reduction in global coupling associated with WMH. This study provides generative models linking the structural damage caused by WMH to alterations in brain dynamics. These models might be used to evaluate the detrimental effects of WMH on brain dynamics in a subject-specific manner. Furthermore, it validates the use of whole-brain modeling for hypothesis-testing of structure-function relationships in diseased states characterized by empirical disconnections.

Keywords: connectivity; dementia; disconnectome; fMRI; neural‐mass modeling; white matter hyperintensities; whole‐brain modeling.

FIGURE 1

Preprocessing workflow of the study. (A) fMRI data preprocessing was performed using the reproducible containerized versions of fMRIPrep (Esteban et al. 2019) and XCP (Adebimpe et al. 2023). (B) WMH segmentation was performed in subject space via automated deep learning software and visually checked to avoid inconsistencies. Infratentorial WMH were removed. (C) MNI‐registered WMH masks were used as input for the Lesion Quantification Toolkit to calculate subject‐specific damage matrices, representing the percentage of damage to each regional connection (represented by an “X” of different dimensions on the damaged tract) and a node disconnection vector, summarizing, for each brain region, the extent of regional disconnection from the whole‐brain network (represented by different shades of red; gray defines nodes that were not disconnected by WMH).

FIGURE 2

Schematics of the modeling pipeline. (A) In a whole‐brain Hopf model, each regional dynamics over time is driven by its intrinsic frequency (not shown, as this was estimated from the empirical data and not modified in relation to WMH) and by a bifurcation parameter, describing the transition from asynchronous noisy behavior (< 0) to full oscillations (> 0), with zero referred to as the critical bifurcation point. The overall activity of the network is derived from the sum of the local activity plus the weighted (by the strength of inter‐regional connection, e.g., number of tracts) sum of the activities of all regions connected to it, scaled by a global coupling parameter (G). Using this framework, a baseline model was constructed based on a normative SC and tuned to the empirical data of subjects without cognitive impairment and without WMH (Baseline model). (B) This panel displays the four WMH‐weighted whole‐brain models, categorized into homogeneous and heterogeneous types according to the hypotheses outlined in the main text. These models investigate whether WMH effects are localized to specific brain regions (heterogeneous) or distributed broadly across the brain (homogeneous). For the homogeneous models, the log‐transformed volume of WMH is applied to uniformly reduce either connectivity or bifurcation parameters uniformly across the brain. In the homogeneous Node Disconnectivity (NDC) model, the bifurcation parameters in all regions are uniformly decreased, represented by red nodes with the same shade. In the homogeneous structural disconnectivity (SDC) model, connectivity across all tracts is uniformly reduced, shown as uniformly thinner red lines. In the heterogeneous models, the NDC model links WMH to changes in bifurcation parameters, using the node disconnectivity vector to inform how these parameters are altered in specific regions, depicted as red nodes of varying shades. Meanwhile, the SDC version simulates how WMH reduce inter‐regional communication along specific tracts. This is achieved by using the damage matrix from the Lesion Quantification Toolkit to decrease connectivity only in those specific tracts, illustrated with red lines of varying thickness. The brain network illustration also includes gray nodes, which represent regions where WMH do not impact bifurcation parameters (matching the baseline model), and blue tracts, indicating connections not affected by WMH. These models were employed to simulate whole‐brain BOLD activity and were subsequently compared against empirical data and a baseline model without WMH information. (C) The dynamics of phase coherence matrices (phFCD) was chosen as the fitting measurement between simulated and empirical data and compared with the Kolmogorov–Smirnov distance (KSD, see Methods for a full description). Lower values of KSD represent a better fit.

FIGURE 3

(A) Maximum intensity projections along the sagittal (left), axial (middle) and coronal (right) planes of WMH probability maps. (B) Histograms of the distribution of the group‐averaged phase functional connectivity dynamics (phFCD) in the groups without (blue) and with (orange) relevant WMH. The distribution of the phFCD is significantly shifted towards lower phFCD values in the group with relevant WMH, suggesting lower synchrony (p < 0.001). (C) Boxplots summarizing model comparisons between the baseline (white) and the homogeneous (light blue) and heterogeneous models (dark blue). The same baseline model (in white) is shown twice for better visualization of the comparison. Structural disconnectivity models (SDC), assessing the impact of WMH on structural connections are grouped on the left. Node disconnectivity (NDC) models, showing the effects of WMH on bifurcation parameters are shown as grouped boxplots on the right. The boxplots represent the Kolmogorov–Smirnov distance (KSD) between the empirical and simulated phase functional connectivity dynamics for the whole group with WMH. Note that after Benjamini–Hochberg correction only the homogeneous SDC remained statistically significant. (D) Scatterplots depicting the correlation between WMH volume (log‐transformed, along the x axes) and the percentage of improvement in the model fit of the considered model (homogenous/heterogeneous SDC/NDC) compared to the baseline model (r refers to post hoc Spearman's rank‐order partial correlation corrected for age, p = p‐value). (E) The same model comparison as in (C) was evaluated in the group with the highest Fazekas score of 3. All comparisons remained statistically significant also after Benjamini–Hochberg correction. *0.01 < p < 0.05; **0.001 < p < = 0.01.

FIGURE 4

Scatterplots depicting the correlation between demographics factors (age, years of patient education and mini‐mental status examination (MMSE), and the percentage improvement in model performance of the considered model compared to the baseline model. The first row shows the results for the homogeneous structural disconnectivity model (SDC), while the second row shows the results for the heterogeneous SDC. For age, r refers to Spearman's rank correlation, while for patient education and MMSE, r refers to Spearman's partial correlation accounting for age. p = p‐value, not corrected for multiple comparisons.