Clustering Analysis based on Regional WMHs Volumes to Explain Cognitive Complaints in Memory Clinic Patients

Abstract

White matter hyperintensities (WMHs) are a common imaging finding in memory clinic patients. They are known to contribute to cognitive impairment. There are different WMHs patterns, and these have varying effects on cognition. This heterogeneity may limit the assessment of the contribution of WMHs to cognitive complaints at the individual level. This study aimed to cluster memory clinic patients based on regional WMHs volumes to better explain cognitive variance associated with WMHs burden compared to unclustered regional WMHs or total WMHs volume(s). Pooled and harmonized individual patient data from 11 international memory clinic cohorts were used. Cognitive function was measured in four cognitive domains. Regional WMHs volumes in 122 tracts of the Julich Brain Atlas were calculated from imaging data. A Mixture-of-Experts (MoE) framework was used to cluster patients based on regional WMHs volumes and to simultaneously predict cognitive performance using these volumes together with age, sex, education, brain parenchymal fraction (BPF), and total WMHs volume. A total of 3525 patients with a mean age of 71.5 years (SD=8.6) were included. Model optimization identified nine clusters as optimal for predicting cognitive scores across domains. Among 144 pairwise comparisons, 122 clusters had distinct WMH patterns (cosine similarity < 0.7). Overall, three of the 36 clusters exhibited WMHs patterns that have been previously associated with Alzheimer's disease and vascular risk factors. The overall performance of all prediction models was low, ranging from 0.013 to 0.085 R². Models that included clusters did not explain more cognitive variance than models that did not. None of the regions had a consistent effect on or were consistently important for cognition across clusters. In 36 of the 180 combinations of covariates and clusters, covariates had a more consistent effect on cognitive performance. The identified clusters based on regional WMHs volumes did not help explain more of the cognitive variance attributed to WMHs burden in memory clinic patients compared to unclustered regional WMHs volumes or total WMHs volume.