Deep Learning patterns of Aβ accumulation in Alzheimer’s Disease

Abstract

AbstractBackgroundAccumulation of amyloid (Aß) plaques is an early pathologic change of Alzheimer’s Disease (AD). However, the mechanisms and pathways by which amyloid spreads across the cerebrum are not fully understood. Deriving distinct dimensions of amyloid deposition and their associations with biomedical factors could be useful to understand how amyloid propagates. In this analysis, we identified two distinct subtypes of progression based on spatiotemporal variations using a recently developed data‐driven, deep learning clustering method called Surreal‐GAN1.MethodWe used data from 482 A+ and 801 A‐ subjects with 18F‐florbetapir PET from the Alzheimer’s Disease Neuroimaging Initiative (ADNI; N = 832) and 11C‐PiB from the Preclinical AD consortium (PAC; N = 451). Standardized uptake value ratio (SUVR) maps (cerebellar GM reference) were transformed to reference space. We derived amyloid status using Gaussian mixture modelling of mean total cortical SUVR for each of 4 sites. We then applied non‐negative matrix factorization2 on the SUVR maps to identify data‐driven common patterns of deposition across the sample. Next, we applied Surreal‐GAN on a total of 1283 subjects to determine the amyloid spatiotemporal subtypes. Correlations with clinical and cognitive variables were evaluated.ResultsParticipants were 30%/22% female and had mean age 74.9 (±7.4)/ 67.9(±9.1) years for ADNI and PAC, respectively. We identified 24 components from NMF (Fig 1A) after excluding 6 as non‐target features. Surreal‐GAN showed optimal agreement indices for two patterns. While both patterns shared involvement of posterior cingulate/precuneus and inferior frontal cortex, pattern r1 shows relatively more occipital deposition and pattern r2 shows more frontal deposition (Fig 1B), as seen when comparing strongly r1 vs strongly r2 participants using voxel‐based morphometry (Fig 2). Both were associated with worse global cognitive function and worsening neurodegeneration (Table 1). r1 correlates with white matter hyperintensity volume, a measure of cerebrovascular disease, while pattern r2 is associated with the presence of APOE e4 alleles.ConclusionEmploying a data‐driven deep learning‐based approach, we derived more than one dissociable pattern of amyloid deposition, potentially differently influenced by vascular disease (r1/occipital) and genetics (r1/frontal). These findings could inform identification of subgroups of individuals who may or may not be responsive to particular clinical intervention.