Atrophy patterns in subjects at increased risk of Alzheimer’s Disease

Abstract

AbstractBackgroundCognitively impaired individuals have multiple contributing conditions leading to heterogeneous atrophy patterns. Unfortunately, neuroimaging data is usually contaminated by noise, which can lead to spurious non‐clinically relevant clusters. Identification of true disease specific clusters requires techniques robust against noise.MethodWe developed an unsupervised approach named COllaborative Clustering based on Adaptive Laplace modeling (CoCAL) to simultaneously cluster individuals and features while removing noise. We derived the distributions of the coefficients produced by non‐negative matrix tri‐factorization (Fig. 1). We applied CoCAL to the PREVENT‐AD dataset, a cohort of cognitively unimpaired participants with a family history of Alzheimer’s Disease (AD). The groups clustered by CoCAL were compared by student’s t‐test with FDR correction, and each cluster is evaluated by Alzheimer’s Progression Score (APS).ResultExperimental results on a synthetic dataset demonstrated superior clustering performance of CoCAL over several benchmark clustering methods including k‐means and conventional matrix factorization based clustering methods. On the PREVENT‐AD dataset, two clusters (G1 and G2) were identified by CoCAL (Table 1). The corresponding t‐test statistic values of regions with p‐values < 0.05 are plotted in Fig. 2. Cluster G1 showed higher APS compared to G2 (APS scores: 0.197 ± 0.0562 [G1] vs ‐0.080 ± 0.0539 [G2], p < 0.001). There were no significant differences in sex, age and APOE ε4 status between the groups.ConclusionWhen applying CoCAL to the PREVENT‐AD cohort, two distinctive atrophy patterns emerge: a parietotemporal cluster (G1) and a fronto‐occipital cluster (G2). G1 corresponds well with classic patterns of cortical atrophy in AD, and participants in this group demonstrate higher APS than G2 participants. In contrast, the greater atrophy in the frontal and occipital lobes in the G2 cluster may reflect either gray matter atrophy from vascular contributions, or the frontal variant of AD. Further investigations into the progression of these different clusters and sub‐networks within the cluster could yield insight for more targeted management of AD.