Alzheimer’s disease heterogeneity assessment with MRI biomarkers and unsupervised statistical learning

Abstract

Alzheimer’s disease (AD) is the most common cause of dementia. It is characterized by loss of memory and other cognitive functions. Although it is a heterogeneous condition, it has been studied as one disease for many decades. Neuropathological data and a large body of in vivo neuroimaging literature challenge the hypothesis that AD is a single entity, supporting the hypothesis of AD as a heterogeneous disease. In this thesis, we set out to understand some aspects of the heterogeneity in AD and aging with the help of atrophy and WM integrity markers from magnetic resonance imaging (MRI). The main aim of the thesis was to investigate the potential use of statistical and machine learning models for the assessment of heterogeneous conditions.

In Study I, we utilized whole brain atrophy markers and cross-sectional cluster analysis to characterize the neurodegeneration variability in a large AD dementia cohort (299 amnestic AD patients). The clusters of patients that we discovered presented with distinct atrophy patterns. Some of them exist due to disease severity, but we identified topologically variable atrophy patterns too. Patients of the different clusters had distinct cognitive symptoms and clinical progression. Then, we designed a pipeline that will help us to assess heterogeneous populations when longitudinal neuroimaging and clinical data are available (Study II).We tested this pipeline in atrophy data from a small dataset of AD patients to assess its usefulness in MRI data and heterogeneous conditions. The model fitted the data well and we concluded that it can be used in larger scale analyses. Moreover, larger numbers of participants with long follow-up period should increase its freedom in searching for heterogeneity in longitudinal neuroimaging trajectories. After this methodological study, we used a very large dataset that consisted of neuroimaging, cerebrospinal fluid (CSF), and clinical data. We split our data in discovery and prediction datasets. The discovery dataset included 𝐴𝛽 positive clinically diagnosed AD dementia patients and 𝐴𝛽 negative cognitively unimpaired individuals (CU). Based on this dataset (Study III), we aimed to understand whether the observed heterogeneity in AD is caused by sampling patient’s data at different disease stages, or if it resembles distinct neurodegeneration subtypes. We modelled longitudinal brain atrophy data anchored to the clinical dementia onset.

Our findings show that all the previously reported atrophy subtypes do exist but some of them reflect disease stages rather than subtypes. Most importantly, our modeling managed to summarize the observed heterogeneity in neurodegeneration with two unique pathways (mediotemporal and cortical). These two pathways have distinct cognitive signatures and were evaluated in a large independent AD dataset. Heterogeneity within the pathways exist and is likely caused by a complex interaction between protective/risk factors and concomitant non-AD pathologies. Some findings indicate that WM changes may precede grey matter atrophy in AD. In Study IV we investigated whether more than one WM profile exists in the aging population. We wanted to understand their association with AD pathophysiological changes and relate them to the risk of developing dementia. We discovered four distinct WM integrity patterns with different spatial WM integrity distribution in aging. Those patterns were related to different longitudinal cognitive profiles and specific white matter tracts informed about cluster assignments.

In conclusion, heterogeneity can be observed not only in AD, but also in the population including healthy individuals. In this thesis, we identified distinct pathways of brain atrophy and WM integrity. Understanding the heterogeneous patterns of the different pathophysiological markers during ageing and the course of AD, will ultimately lead to the development of disease modifying (personalized) treatments.