A typology of cerebral small vessel disease based on imaging markers

Background Lacunes, microbleeds, enlarged perivascular spaces (EPVS), and white matter hyperintensities (WMH) are brain imaging features of cerebral small vessel disease (SVD). Based on these imaging markers, we aimed to identify subtypes of SVD and to evaluate the validity of these markers as part of clinical ratings and as biomarkers for stroke outcome. Methods In a cross-sectional study, we examined 1207 first-ever anterior circulation ischemic stroke patients (mean age 69.1 ± 15.4 years; mean NIHSS 5.3 ± 6.8). On acute stroke MRI, we assessed the numbers of lacunes and microbleeds and rated EPVS and deep and periventricular WMH. We used unsupervised learning to cluster patients based on these variables. Results We identified five clusters, of which the last three appeared to represent distinct late stages of SVD. The two largest clusters had no to only mild or moderate WMH and EPVS, respectively, and favorable stroke outcome. The third cluster was characterized by the largest number of lacunes and a likewise favorable outcome. The fourth cluster had the highest age, most pronounced WMH, and poor outcome. Showing the worst outcome, the fifth cluster presented pronounced microbleeds and the most severe SVD burden. Conclusion The study confirmed the existence of different SVD types with different relationships to stroke outcome. EPVS and WMH were identified as imaging features of presumably early progression. The number of microbleeds and WMH severity appear to be promising biomarkers for distinguishing clinical subgroups. Further understanding of SVD progression might require consideration of refined SVD features, e.g., for EPVS and type of lacunes. Supplementary Information The online version contains supplementary material available at 10.1007/s00415-023-11831-x.


Supplemental Online Materials Supplementary Table 1: Detailed cluster centroids
Detailed data on the cluster centroids, i.e. average scores across the included variables for each cluster.Data were normalized and partially de-skewed by log transformation before being subjected to the k-means algorithm.Here, data are shown re-transformed into original data space to allow meaningful interpretation.Values are mean (standard deviation).EPVSenlarged perivascular spaces; WMH PVperiventricular white matter hyperintensities; WMH Ddeep white matter hyperintensities.This finding raised the question of whether the underestimation of certain SVD imaging features by 1.5T MRI could have biased our findings.Hence, we re-ran the k-means cluster analysis with the subsample of 441 patients with 3T MRI.We again searched for 5 clusters and exactly replicated all other factors.The results are shown in Supplementary Figure 1 below.Conceptually, the cluster solution closely resembled the original one in the total sample.Each cluster in the replication could be assigned to one of the original clusters.Differences were mostly minimal and can be explained by the varying sensitivity of 1.5T vs. 3T MRI as well as slightly different cluster borders.In conclusion, any potential underestimation of SVD features should not have affected the main results of our study.Still, magnetic field strength appears to be a factor to consider in studies on imaging markers of SVD, as a small but significant difference in sensitivity between field strengths exists.

Supplemental analysisestimation of disease progression from cross-sectional data
In an additional analysis, we aimed to infer the typical disease progression of SVD markers with Bayesian tree models that are capable to represent the progression of binary disease events from cross-sectional data (Beerenwinkel et al., 2005).Studies on the progression of genetic aberrations in tumours (Ketter et al., 2007) and HIV (Beerenwinkel et al., 2005) utilized mutagenetic trees to model disease progression from cross-sectional data.A patient's disease progression within such a model, represented by a so-called genetic progression score, was identified as a biomarker for cancer survival (Rahnenführer et al., 2005).Likewise, models that convert the simple binary assessment of SVD into progression scores might bear the potential to improve biomarkers over simple scorings.
Based on the conditional probabilities between events (i.e. the presence of a type of SVD feature), this technique generates interpretable, directed graph models that visualize the typical order in which SVD features appear.We estimated models with the binary SVD variables for lacunes, microbleeds, EPVS, deep WMH, and periventricular WMH with the Rtreemix package (Bogojeska et al., 2008) in Rsoftware.To account for potentially divergent typical progressions, we allowed for a mixture model of multiple trees (Beerenwinkel et al., 2005) and selected the number of parallel models by evaluation of maximum likelihood.Additionally, we computed for each patient the so-called genetic progression score, a measure that accounts for a patient's disease progression within the estimated tree.The validity of this measure compared to the SVD MRI burden score was assessed by their correlations with stroke outcome, as measured by 3 months post-stroke modified Rankin Scale.
Correlations were statistically compared with the R cocor package (Diedenhofen et al., 2015).All statistics were performed at a two-tailed alpha level of p = 0.05.

Results
Maximum likelihood evaluation suggested that three tree graph models are sufficient to describe the disease progression underlying the cross-sectional binarized SVD data.
As such mixture model includes a star graph to represent a noise component this leaves two interpretable trees (supplementary Figure below).The first tree suggested EPVS as the first disease stage followed by WMH.

Supplementary Table 2: Detailed statistics for the comparison between clusters Detailed
statistics for the omnibus tests for the comparison of demographic and clinical data between clusters, as shown in Figure3.mRSmodifiedRankinScale.See Figure3for the results of post hoc tests.

the impact of MRI field strength on SVD typology
We assessed features of SVD by MRI scanners with a field strength of either 1.5T