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

Abstract

Brain atrophy, largely driven by tau deposition, is the most proximal correlate of cognitive decline in Alzheimer's disease (AD). Understanding the heterogeneity and longitudinal progression of brain atrophy during the disease course will play a key role in understanding the mechanisms of AD.The aim of this study is to propose a framework for longitudinal clustering that: 1) incorporates simultaneous clustering of longitudinal multivariate neuroimaging measures, 2) leverages information of individuals with irregularly sampled observations (different sampling times), 3) compares clusters with a control group, 4) allows the study and fixation of potential confounding effects, 5) can provide visualization of the resulting clusters for interpretation, 6) measures the uncertainty of clustering. We aimed to include amyloid-β positive AD patients and amyloid-β negative cognitively unimpaired (CU) subjects with longitudinal data, three sMRI scans over two years. Cortical thickness and subcortical volume measures from the longitudinal stream of FreeSurfer 6.0 pipeline were used as input for cluster analysis.Using the proposed methodology, we found 3 distinct longitudinal brain atrophy patterns in AD patients: a typical diffuse AD pattern (n=34, 47.2%), and 2 atypical AD patterns: Minimal atrophy (n=23 31.9%) and Hippocampal sparing (n=9, 12.5%). We also identified outlier observations (n=3, 4.2%) and observations with uncertain classification (n=3, 4.2%). The clusters of AD patients differed not only in regional distributions of atrophy at baseline, but also in atrophy progression over time, age at AD onset, cognitive deficits at baseline and cognitive decline over time.A framework for the longitudinal assessment of variability in cohorts with several neuroimaging measures was successfully developed and the results show that it can be used to understand heterogeneity in the context of AD.

Multi-scale semi-supervised clustering of brain images: deriving disease subtypes

“…Sub4 (164 MCI & 34 AD) has the highest proportion of MCI. The normal anatomy subtype has been confirmed in previous works both from semi-supervised and unsupervised methods (Dong et al, 2016b; Ezzati et al, 2020; Jung et al, 2016; Nettiksimmons et al, 2014; Ota et al, 2016; Poulakis et al, 2020, 2018; Ten Kate et al, 2018; Yang et al, 2020). Sub2 showed typical AD-like neuroanatomical patterns with diffuse atrophy over the whole brain, with the largest effect size in the hippocampus and medial temporal lobe.…”

“…The proposed method seamlessly integrates multi-scale representation learning and semi-supervised clustering in a coherent framework via a double-cyclic optimization procedure to yield scale-agnostic delineation of heterogeneous disease patterns. In contrast to existing unsupervised approaches presented in (Ezzati et al, 2020, 2020; Jeon et al, 2019, 2019; Jung et al, 2016; Lubeiro et al, 2016; Nettiksimmons et al, 2014; Ota et al, 2016; Pan et al, 2020; Park et al, 2017; Planchuelo-Gómez et al, 2020; Poulakis et al, 2020, 2020, 2018; Sugihara et al, 2016; Ten Kate et al, 2018), MAGIC is a semi-supervised approach, leveraging the patient-control dichotomy to drive subtypes that reflect distinct pathological processes. In contrast to the existing state-of-the-art semi-supervised clustering method (i.e., HYDRA), MAGIC can accurately delineate effect patterns that are both global and focal, thanks to its multi-scale optimization routine.…”

“…However, there still exists a gap between these models and human interpretability, especially for clinicians who frequently seek discrete disease subtypes (Miotto et al, 2018). Thus, many recent efforts to discover the heterogeneous nature of brain diseases have investigated different clustering algorithms (Chand et al, 2020; Dong et al, 2016a, 2016b; Dwyer et al, 2018; Ezzati et al, 2020; Filipovych et al, 2012; Honnorat et al, 2019; Jeon et al, 2019; Jung et al, 2016; Lubeiro et al, 2016; Nettiksimmons et al, 2014; Ota et al, 2016; Pan et al, 2020; Park et al, 2017; Planchuelo-Gómez et al, 2020; Poulakis et al, 2020, 2020, 2018; Sugihara et al, 2016; Ten Kate et al, 2018; Varol et al, 2017; Young et al, 2018; Zhang et al, 2016). These methods can be divided into two categories depending on whether the clustering algorithm is unsupervised or semi-supervised b .…”

Towards harmonizing subtyping methods for neuroimaging studies in Alzheimer’s disease

“…Studies from our lab used visual rating scales of brain atrophy in medial temporal, frontal, and posterior cortices (7)(8)(9)(10)(11)(12)(13), and determined abnormality based on clinical cut points (14). We also used an unsupervised clustering method in another cross-sectional study (15), which has recently been extended for subtyping on longitudinal data (16). Other groups used different unsupervised clustering methods (17)(18)(19)(20)(21)(22)(23)(24), highlighting the methodological variability across studies.…”

“…Surprisingly, no head-to-head comparison of subtyping methods has been published so far. Such a comparison arises as an urgent and important step towards facilitating consistent progress in this field, especially with the current surge in subtyping studies using longitudinal sMRI (16,26,27) and tau PET (3,25,28). To illustrate this problem and substantiate our claim for harmonizing subtyping methods, we applied subtyping methods based on five previous studies (1,(5)(6)(7)15,25), on sMRI and tau PET data, in the same cohort.…”