Machine‐learning derived white matter signatures and the risk of dementia, stroke and mortality: the Rotterdam Study

Abstract

AbstractBackgroundWhite matter hyperintensities (WMH), loss of white matter (WM) volume and WM microstructural integrity are important risk indicators of stroke and dementia, albeit of heterogeneous nature. Use of machine learning techniques may unravel different patterns of white matter injury, with distinct underlying pathologies and disease risk.MethodWe measured WMH volume, WM volume, global fractional anisotropy (FA) and global mean diffusivity (MD) in 5205 participants (mean age 64.9 years, 56.0% women) of the population‐based Rotterdam study with brain MRI between 2005 and 2016. We performed hierarchical clustering on age‐standardized imaging parameters to identify separate clusters of white matter injury, and compare determinants across clusters. We then determined the association between clusters and risk of dementia, stroke and mortality, using Cox proportional hazard models adjusted for sex, education, APOE‐ε4 and cardiovascular risk factors.ResultWe identified four distinct white matter signatures: (i) cluster with above‐average microstructural integrity and little WM atrophy and WMH, (ii) cluster with above‐average microstructural integrity and little WMH, yet substantial WM atrophy, (iii) cluster with poor microstructural integrity and substantial WMH, yet little WM atrophy, and (iv) cluster with poor microstructural integrity, substantial WM atrophy, and average WMH load. There were no clear differences between clusters in APOE genotype and cardiometabolic risk factors, except for higher prevalence of hypertension in cluster (iv). Cluster (iii) contained more women, and participants in clusters (iii) and (iv) had more microbleeds and lacunes. During a median 10.7 years of follow‐up, 272, 210, and 844 cases of dementia, stroke, death occurred, respectively. In fully adjusted models, dementia risk was increased for clusters (ii) (HR 1.57 [95%CI:1.04‐2.37]), (iii) (HR 2.43 [95%CI:1.68‐3.52]) and (iv) (HR 1.78 [95%CI:1.22‐2.62], compared to cluster (i). Moderate associations of clusters (iii) and (iv) with incidence stroke attenuated and were no longer statistically significant after adjustment for cardiovascular risk factors, while only cluster (iii) was associated with a small increase in mortality (HR 1.30 [95%CI:1.07‐1.58]).ConclusionMachine‐learning derived white matter signatures are differentially associated with dementia, stroke and mortality. Further research is needed to pinpoint differences in underlying pathology, and assess potential benefit of clusters for risk stratification.