Heterogeneous patterns of brain atrophy in Alzheimer's disease

Abstract

There is increasing evidence showing that brain atrophy varies between patients with Alzheimer's disease (AD), suggesting that different anatomical patterns might exist within the same disorder. We investigated AD heterogeneity based on cortical and subcortical atrophy patterns in 299 AD subjects from 2 multicenter cohorts. Clusters of patients and important discriminative features were determined using random forest pairwise similarity, multidimensional scaling, and distance-based hierarchical clustering. We discovered 2 typical (72.2%) and 3 atypical (28.8%) subtypes with significantly different demographic, clinical, and cognitive characteristics, and different rates of cognitive decline. In contrast to previous studies, our unsupervised random forest approach based on cortical and subcortical volume measures and their linear and nonlinear interactions revealed more typical AD subtypes with important anatomically discriminative features, while the prevalence of atypical cases was lower. The hippocampal-sparing and typical AD subtypes exhibited worse clinical progression in visuospatial, memory, and executive cognitive functions. Our findings suggest there is substantial heterogeneity in AD that has an impact on how patients function and progress over time.

Fully Bayesian longitudinal unsupervised learning for the assessment and visualization of AD heterogeneity and progression

“…We included all subjects with longitudinal sMRI data and available CSF data (101 AD and 113CU) from our previously published cross-sectional study on AD subtypes (Poulakis et al, 2018). This was done to be able to compare the cross-sectional and the longitudinal clustering approach in a proper way in the same set of participants.…”

“…A schematic representation of the proposed analytical framework is portrayed in Figure 1. The time from the first visit (baseline) was defined as a random effect for the sake of comparability with the previous literature on AD subtypes were only one observation for each subject is included (Noh et al, 2014;Poulakis et al, 2018). Therefore, the intercept of the model will correspond to the atrophy levels on the first visit and the slope will show how these atrophy levels change over the months The statistical model that we chose to employ has all the features that were described above and its original specification and hyperparameter choices can be found in the supplementary material of that study (Arnošt Komárek & Komárková, 2013).…”

“…When samples include only one diagnosis, a common use of clustering methods is to investigate whether the neuroimaging measures of interest show heterogeneous patterns within that same diagnostic label. Several studies have investigated the heterogeneity in AD with the aim to define disease specific subtypes (Byun et al, 2015(Byun et al, , 2015Corlier et al, 2018;Park et al, 2017;Poulakis et al, 2018;Schwarz et al, 2018;Whitwell et al, 2012;Young et al, 2017). When samples include more than one diagnosis, the main aim of unsupervised clustering methods is to investigate whether neuroimaging markers can be used to distinguish between the diagnostic classes without specifying them with a label.…”

“…Previous clustering studies grouped AD patients based on sMRI features from a single time-point (Dong et al, 2016(Dong et al, , 2017Ferreira et al, 2017;Noh et al, 2014;Park et al, 2017;Poulakis et al, 2018;Varol et al, 2017). Their conclusions were based on a single observation in time and the chance that those clusters reflect different stages of the disease and not particular patterns of atrophy (distinct AD subtypes) cannot be excluded.…”

“…The designed framework is applied to longitudinal sMRI data of mainly amyloid-β (Aβ) positive AD patients and Aβ negative cognitively unimpaired (CU) subjects over a period of two years (three sMRI time points). To validate the results from our new longitudinal clustering framework, we included all data with longitudinal information from our previous cross-sectional clustering study (Poulakis et al, 2018). This allows us to compare the results from cross-sectional and longitudinal clustering in the same dataset.…”