Enhancing cognitive performance prediction by white matter hyperintensity connectivity assessment

Abstract

White matter hyperintensities of presumed vascular origin (WMH) are associated with cognitive impairment and are a key imaging marker in evaluating brain health. However, WMH volume alone does not fully account for the extent of cognitive deficits and the mechanisms linking WMH to these deficits remain unclear. Lesion network mapping (LNM) enables us to infer if brain networks are connected to lesions and could be a promising technique for enhancing our understanding of the role of WMH in cognitive disorders. Our study employed LNM to test the following hypotheses: (i) LNM-informed markers surpass WMH volumes in predicting cognitive performance; and (ii) WMH contributing to cognitive impairment map to specific brain networks. We analysed cross-sectional data of 3485 patients from 10 memory clinic cohorts within the Meta VCI Map Consortium, using harmonized test results in four cognitive domains and WMH segmentations. WMH segmentations were registered to a standard space and mapped onto existing normative structural and functional brain connectome data. We employed LNM to quantify WMH connectivity to 480 atlas-based grey and white matter regions of interest (ROI), resulting in ROI-level structural and functional LNM scores. We compared the capacity of total and regional WMH volumes and LNM scores in predicting cognitive function using ridge regression models in a nested cross-validation. LNM scores predicted performance in three cognitive domains (attention/executive function, information processing speed, and verbal memory) significantly better than WMH volumes. LNM scores did not improve prediction for language functions. ROI-level analysis revealed that higher LNM scores, representing greater connectivity to WMH, in grey and white matter regions of the dorsal and ventral attention networks were associated with lower cognitive performance. Measures of WMH-related brain network connectivity significantly improve the prediction of current cognitive performance in memory clinic patients compared to WMH volume as a traditional imaging marker of cerebrovascular disease. This highlights the crucial role of network integrity, particularly in attention-related brain regions, improving our understanding of vascular contributions to cognitive impairment. Moving forward, refining WMH information with connectivity data could contribute to patient-tailored therapeutic interventions and facilitate the identification of subgroups at risk of cognitive disorders.

Keywords: cerebral small vessel disease; dementia; lesion network mapping; magnetic resonance imaging; vascular cognitive impairment; white matter hyperintensities.

Figure 1

Methodology. (A) Data from 10 memory clinic cohorts of the Meta VCI Map Consortium were used including harmonized cognitive scores and WMH segmentations in MNI space. For functional lesion network mapping (fLNM) we employed the GSP1000 normative functional connectome comprising resting state fMRI (rsfMRI) data from 1000 healthy GSP participants. For structural lesion network mapping (sLNM), we used the HCP32 normative structural connectome based on diffusion-weighted imaging data from 32 healthy HCP participants, detailing the fibre bundle architecture. (B) LNM was performed to quantify the functional and structural connectivity of white matter hyperintensities of presumed vascular origin (WMH) to multiple regions of interest (ROIs) (Schaefer400 × 7 cortical, Melbourne Subcortical Atlas subcortical, HCP1065 white matter areas). For this, voxel-level functional and structural connectivity maps were computed for each ROI, reflecting resting state blood oxygen level-dependent (BOLD) correlations or anatomical connection strength via tractography streamlines, respectively. ROIwise LNM scores were derived by averaging voxel-level connectivity indices within the normalized WMH masks, considering only positive correlation coefficients for functional mapping. This resulted in a matrix for both fLNM and sLNM scores per ROI per patient (n_ROIs × n_patients). The matrices shown in the figure are populated with random data only serving as a visual aid. (C) The fLNM and sLNM scores across patients were used in predictive models to estimate cognitive domain scores (predictive modelling analysis) and analysed in permutation-based general linear models to identify regions significantly influencing the cognitive domain-WMH disconnectivity relationship at the ROI level (ROI-level inferential statistics). GSP = Genomic Superstruct Project; HCP = Human Connectome Project.

Figure 2

Predictive modelling analysis. Violin plots illustrate prediction outcomes across cognitive domains. Each violin displays the distribution of Pearson correlations (between actual and predicted cognitive domain performance; 10-fold cross-validation × 10 repeats = 100 folds → 100 Pearson correlations) for a model informed by a different feature set. The higher the Pearson correlation, the higher the prediction performance. Blue = demographics (age, sex and education); orange = total WMH volume + demographics; green = tract-level WMH volumes + demographics; red = sLNM scores + demographics; purple = fLNM scores + demographics; brown = sLNM scores + fLNM scores + demographics. Average Pearson correlations are indicated above each violin, with coloured dots showing training score averages. Geometric symbols denote t-test results comparing LNM-based models against demographics and WMH volume-based models: filled triangle ▴ indicates significantly higher Pearson correlation than demographics, filled square ▪ indicates significantly higher Pearson correlation than WMH volume + demographics; and filled pentagon ⬟ indicates significantly higher Pearson correlation than tract-level WMH volume + demographics. Below the violin plots, performance curves display the average Pearson correlations across folds, for subsets randomly sampled in sizes ranging from 20% to 100% of the entire dataset. Line colours match the corresponding violin plots in A, which display predictive modelling results for the full sample size. Again, higher Pearson correlation indicates higher prediction performance. fLNM = functional lesion network mapping; sLNM = structural lesion network mapping; WMH = white matter hyperintensities of presumed vascular origin.

Figure 3

Inferential statistics results of cortical and subcortical grey matter. Anatomical plots on the left display the regional relationship between lesion network mapping (LNM) scores and cognitive domain scores. Regions of interest (ROIs) in which LNM scores across participants were significantly associated with cognitive domain scores after family-wise error rate correction are highlighted by colours encoding β-coefficients from general linear models: a negative β (red) denotes that a higher regional LNM score, i.e. higher WMH connectivity, is associated to a lower performance in individual cognitive domains; a positive β (blue) indicates that a higher regional LNM score is linked to a higher cognitive domain performance. Bar plots on the right display the corresponding β-coefficients averaged in the canonical (Yeo) resting state functional connectivity networks. The brain in the bottom right indicates the regional distribution of the canonical resting state networks with colours corresponding to the bars. Statistical significance was assessed using spin permutations. Each row corresponds with a different combination of lesion network mapping modality and cognitive domain: (A) fLNM—attention/executive function; (B) fLNM—information processing speed; (C) fLNM—verbal memory; (D) sLNM—attention/executive function; (E) sLNM—information processing speed; and (F) sLNM—verbal memory. fLNM = functional lesion network mapping; P_spin = P-value derived from spin permutations; sLNM = structural lesion network mapping; WMH = white matter hyperintensities of presumed vascular origin.

Figure 4

Inferential statistics results of white matter tracts. Radar plots displaying the top 10 strongest linear associations (standardized β) for the functional (A) and structural (B) lesion network mapping (LNM) scores in each tract in association with cognitive domain scores. Strongest associations are shown at the 3 o’clock position, decreasing in strength counterclockwise. Red dots indicate a negative association (higher LNM score − lower cognitive domain score) and blue dots indicate a positive association (higher LNM score − higher cognitive domain score). Faintly coloured dots indicate non-significant associations. Tracts with a significant association are displayed below the radar plots in alphabetical order. For paired tracts only left side examples are visualized. AF = arcuate fascicle; C = cingulate; CBT = corticobulbar tract; CPT = corticopontine tract; CS = corticostriatal pathway; F = fornix; FAT = frontal aslant tract; MdLF = middle longitudinal fasciculus; SLF = superior longitudinal fasciculus; UF = uncinate fasciculus; fLNM = functional lesion network mapping; IPS = information processing speed; n.s. = non-significant; P = P-value; sLNM = structural lesion network mapping.