Disentangling The Spatio-Temporal Heterogeneity of Alzheimer’s Disease Using A Deep Predictive Stratification Network

Abstract

Alzheimer’s disease (AD) is clinically heterogeneous in presentation and progression, demonstrating variable topographic distributions of clinical phenotypes, progression rate, and underlying neuro-degeneration mechanisms. Although striking efforts have been made to disentangle the massive heterogeneity in AD by identifying latent clusters with similar imaging or phenotype patterns, such unsupervised clustering techniques often yield sub-optimal stratification results that do not agree with clinical manifestations. To address this limitation, we present a novel deep predictive stratification network (DPS-Net) to learn the best feature representations from neuroimages, which allows us to identify latent fine-grained clusters (aka subtypes) with greater neuroscientific insight. The driving force of DPS-Net is a series of clinical outcomes from different cognitive domains (such as language and memory), which we consider as the benchmark to alleviate the heterogeneity issue of neurodegeneration pathways in the AD population. Since subject-specific longitudinal change is more relevant to disease progression, we propose to identify the latent subtypes from longitudinal neuroimaging data. Because AD manifests disconnection syndrome, we have applied our datadriven subtyping approach to longitudinal structural connectivity networks from the ADNI database. Our deep neural network identified more separated and clinically backed subtypes than conventional unsupervised methods used to solve the subtyping task– indicating its great applicability in future neuroimaging studies.