Predicting disability progression and cognitive worsening in multiple sclerosis using patterns of grey matter volumes

Abstract

Objective: In multiple sclerosis (MS), MRI measures at the whole brain or regional level are only modestly associated with disability, while network-based measures are emerging as promising prognostic markers. We sought to demonstrate whether data-driven patterns of covarying regional grey matter (GM) volumes predict future disability in secondary progressive MS (SPMS).

Methods: We used cross-sectional structural MRI, and baseline and longitudinal data of Expanded Disability Status Scale, Nine-Hole Peg Test (9HPT) and Symbol Digit Modalities Test (SDMT), from a clinical trial in 988 people with SPMS. We processed T1-weighted scans to obtain GM probability maps and applied spatial independent component analysis (ICA). We repeated ICA on 400 healthy controls. We used survival models to determine whether baseline patterns of covarying GM volume measures predict cognitive and motor worsening.

Results: We identified 15 patterns of regionally covarying GM features. Compared with whole brain GM, deep GM and lesion volumes, some ICA components correlated more closely with clinical outcomes. A mainly basal ganglia component had the highest correlations at baseline with the SDMT and was associated with cognitive worsening (HR=1.29, 95% CI 1.09 to 1.52, p<0.005). Two ICA components were associated with 9HPT worsening (HR=1.30, 95% CI 1.06 to 1.60, p<0.01 and HR=1.21, 95% CI 1.01 to 1.45, p<0.05). ICA measures could better predict SDMT and 9HPT worsening (C-index=0.69-0.71) compared with models including only whole and regional MRI measures (C-index=0.65-0.69, p value for all comparison <0.05).

Conclusions: The disability progression was better predicted by some of the covarying GM regions patterns, than by single regional or whole-brain measures. ICA, which may represent structural brain networks, can be applied to clinical trials and may play a role in stratifying participants who have the most potential to show a treatment effect.

Figure 1

Visual representation of our image-analysis pipeline. Aiming to identify data-driven network-based measures of covarying GM volumes, we initially preprocessed our data as in Eshaghi et al (N4 bias field correction, lesion filling, brain segmentation and parcellation). We created a customised template from all the available scans from 39 randomly selected subjects. After having resampled those scans to an isotropic space, we created 39 single subject templates, and from those an average study-specific template. We registered the T1 lesion filled scans to the template and diffeomorphically transformed the GM segmentation maps to the template using the warping matrix generated from the previous step. We modulated the GM segmentation maps by the Jacobian determinants in order to account for possible deformations to the original volumes occurred after the non-linear transformation. We applied an 8 mm smoothing kernel to account for intersubject variability and applied a whole brain mask to constrain the following analysis at the level of the brain. Aiming to prove the stability of our results, we randomly divided our cohort into four folds. For each fold and for the entire cohort, we generated a 4D image by concatenating the available GM maps and ran fast ICA on each of those inputs allowing for 20 components to be identified. For each fold and for the entire cohort, we generated a 4D image by concatenating the 20 generated ICA components and ran cross-sectional correlations between those inputs to identify which components were stable and could be implemented for statistical analysis. 4D, four-dimensional; ANTs, advanced normalisation tools; GM, grey matter; ICA, independent component analysis.

Figure 2

Stable independent component analysis (ICA) components. To determine the stability of the ICA components, we randomly split the sample into four folds and ran the ICA on each of them, as well as on the entire sample. While allowing for 20 components to be identified, cross-sectional correlations proved that only 15 out of the 20 ICA components were stable (emerged in all of the four folds and from the entire sample). The colour bar represents the loading of each component. Most of the identified networks resampled well-known functional systems. Component 3 represents an auditory-like network, spanning mainly the superior temporal gyrus, posterior insular and Heschl’s gyrus (cognition-language-speech network). Component 5 is a sensorimotor-like network, encompassing the precentral gyrus, postcentral gyrus and supramarginal gyrus (action-execution network). Component 6 resamples a cerebellum-like network, involving mainly the cerebellum and fusiform gyrus, temporal and parietal lobe. Component 8 is a cortico-basal ganglia-like network, spanning the brain stem, pons, thalamus, nucleus accumbens, insula, putamen, caudate, pallidum, frontal and temporal lobe. Component 9 represents an executive control-like network, involving mainly medial frontal areas (action planning and inhibition). Component 11 is a visuo-like network, encompassing mainly several regions of the occipital pole and supramarginal, temporal and parietal areas. Component 15 resamples a salience-like network, involving the insula, thalamus and striatus (autonomic reaction to salient stimuli; goal-directed behaviour). Component 17 represents an affective and reward network, encompassing mainly the anterior cingulate, medial orbitofrontal cortex and prefrontal cortex. Component 20 resamples a default mode-like network (DMN-like), spanning mainly the precuneus, posterior cingulate and middle frontal gyrus. The remaining identified networks did not correspond to any major brain functional network, but can be labelled by their predominantly involved brain areas. Component 1 is a superior frontal network, encompassing mainly superior and medial frontal brain areas. Component 2 is a temporal-like network, involving mainly temporal brain regions. Component 7 is a precuneus-like network. Component 12 is an occipito-temporal-like network, spanning mainly the temporal and occipital pole. Component 13 represented a prefrontal cortex-like network, involving mainly frontal and orbitofrontal brain areas. Component 18 is a parieto-temporal-like network, involving mainly temporal and parietal brain areas.

Figure 3

Correlations between baseline ICA components and baseline EDSS, 9HPT and SDM. Among the 15 stable ICA component, baseline SDMT score was more strongly associated with a mainly basal ganglia component (component 8). Among the three clinical tests, (A) SDMT had the highest correlations with ICA networks (mainly with component 8). (B) 9HPT was associated with the factor loading of component 8. 9HPT and SDMT correlated better with some ICA networks rather than with any other regional or whole brain MRI measure. (C) Among all the 15 networks, component 6 (ie, cerebellum, brain stem, pons) had the highest correlation with EDSS. We used the Bonferroni correction to correct for multiple comparisons. CI band is added to the figure. EDSS, Expanded Disability Status Scale; ICA. independent component analysis; SDMT, Symbol Digit Modalities Test; 9HPT, Nine-Hole Peg Test.

Figure 4

Cox regression models predictive of 9HPT worsening. HR of the statistically significant predictors of 9HPT worsening. The figure shows that two GM networks and the volume of the DGM can predict the 9HPT progression. HR >1 indicates that for each SD increase in the corresponding variable there is a higher risk of developing the event. HR <1 indicates that for each SD decrease in the corresponding variable, there is a higher risk of progressing on 9HPT. Error bars represent the CI. P values <0.05 represent a statistically significant relative risk of developing a 9HPT progression comparing subjects for each independent variable shown on the vertical axis. Component 2 encompasses the temporal lobe, middle cingulate gyrus, precentral gyrus medial segment, posterior cingulate gyrus, parietal lobule, inferior and middle temporal gyrus, parahippocampal gyrus, fusiform gyrus and entorhinal area. Component 20 consisted of precuneus, posterior cingulate gyrus, middle and superior frontal gyrus, angular gyrus, superior occipital and superior parietal lobule. 9HPT, Nine-Hole Peg Test; GM, grey matter.

Figure 5

Cox regression models predictive of SDMT worsening. HR of the statistically significant predictors of SDMT worsening in separate Cox regression models. The figure shows that six ICA components, lesion load and the volumes of the thalamus could predict the SDMT progression. HR >1 indicates that for each SD increase in the corresponding variable, there is a higher risk of developing the event. HR >1 indicates that for each SD decrease in the corresponding variable, there is a higher risk of progressing on SDMT. For each SD increase in component 8 (encompassing mainly basal ganglia regions), which is inversely related to GM volumes, there was a 29% higher risk of developing SDMT progression. For each SD decrease in the volume of the thalamus, there is a 18% increased risk of worsening in SDMT. Error bars represent the CI of HR. P values <0.05 represent a statistically significant relative risk of developing an SDMT progression for each independent variable shown on the vertical axis. GM, grey matter; ICA. independent component analysis; SDMT, Symbol Digit Modalities Test.