Evaluating cognitive phenotypes and associations with neuroimaging biomarkers of dementia in a sample of aging individuals using unsupervised machine learning

Abstract Background Otherwise cognitively normal individuals can experience age‐related cognitive dysfunction and may present with heterogenous patterns of cognitive impairment. There is limited knowledge regarding the presence and stability of cognitive subtypes observed in individuals with mild cognitive impairment and dementia (e.g., dysexecutive, non‐amnestic). Further, the extent to which cognitive subtypes related to age‐ and dementia‐related neuroimaging biomarkers is uncertain. Method This study sought to evaluate the presence and composition of cognitive subtypes in a large diverse community sample of aging individuals without dementia (NC; Baseline N = 337; Table 1). Participants enrolled in the Wake Forest ADRC and completed an annual battery of cognitive assessments (Uniform Data Set; NACC), the Mini‐Mental State Exam (MMSE), and Clinical Diagnostic Rating Scale (CDR). Unsupervised machine learning (e.g., clustering: hierarchical, k‐means, and k‐medoids algorithms; 1000 bootstrap iterations; majority voting rule) was applied to the battery of cognitive assessments. Demographic characteristics and neuroimaging measures of amyloid deposition (global PiB PET SUVr; Aβ‐PET), total brain volume (BVOL), diffusion‐weighted fractional anisotropy (FA), and NODDI freewater (FW) were compared across prospective clusters using general linear models. Highly correlated cognitive metrics (r>.85) were removed prior to analysis. Result Cluster analyses using baseline cognitive scores suggested NC participants could be partitioned into two to eight clusters; two primary clusters representing one large homogenous group of relatively unimpaired individuals and another heterogenous cluster composed of 4‐7 possible cognitive subgroups (Figure 1). The two‐cluster solution partitioned NC participants into higher and lower cognitive performers. Across cluster solutions, lower performing individuals were older, had higher Aβ‐PET, and lower total brain volume on average (p<.05; Figure 2). Conclusion We did not observe well‐defined clusters representing distinct cognitive subtypes reported in previous studies (e.g., dysexecutive, amnestic, visuospatial). However, a heterogenous subset of NC participants exhibited substantial variability across assessments. Three‐ and four‐cluster solutions identified individuals with greater impairment in executive functioning, and verbal learning and memory respectively. Findings highlight the need to model complex patterns of cognitive abilities in otherwise healthy older adults. Future longitudinal work will assess how health information and neuroimaging biomarkers of dementia contribute to the composition and stability of cognitively defined subgroups.