Comprehensive evaluation of AT(N) imaging biomarkers for predicting cognition
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background and Objectives
Imaging biomarkers enable in vivo quantification of amyloid, tau, and neurogenerative pathologies that develop in Alzheimer’s Disease (AD). Interest in imaging biomarkers has led to a wide variety of biomarker definitions, some of which potentially offer less predictive value than others. We aimed to assess how different operationalizations of AD imaging biomarkers affect prediction of cognition.
Methods
We included individuals from ADNI who underwent amyloid-PET ([ 18 F]-Florbetapir), tau-PET ([ 18 F]-Flortaucipir), and volumetric MRI imaging. We compiled a large collection of imaging biomarker definitions (42 in total) spanning different pathologies (amyloid, tau, neurodegeneration) and variable types (continuous, binary, non-binary categorical). Using cross-validation, we trained regression models to predict neuropsychological performance, both globally and across different subdomains (Phenotype Harmonization Consortium composites), using different combinations of biomarkers. We also compared these biomarker models to support vector machines (SVMs) trained to predict cognition directly from imaging regions of interest. In a subsample of individuals with CSF biomarker readouts, we repeated experiments comparing the accuracy of models using imaging and fluid biomarkers. Additional analyses tested the predictive strength of imaging biomarkers when limited to specific clinical stages of disease (cognitive unimpaired vs. impaired) and when modeling longitudinal cognitive change.
Results
Our sample included 490 people (247 female) with a mix of no impairment (n=288), mild impairment (n=163), and dementia (n=39). While almost all biomarkers tested were predictive of cognitive performance, we observed substantial variability in accuracy, even for measures of the same pathology. Tau biomarkers were the single most accurate single predictors, though combination of biomarkers spanning multiple pathologies were more accurate overall. SVM models were generally more accurate than models using traditional biomarkers. Incorporating continuous or non-binary categorical biomarkers was beneficial only for tau and neurodegeneration, but not amyloid. Patterns of results were largely consistent when considering different clinical stages of disease, neuropsychological domains, and longitudinal cognition. In the CSF subsample (n=246), imaging biomarkers strongly outperformed CSF versions for cognitive prediction.
Discussion
We demonstrated that different imaging biomarker definitions can lead to variability in downstream predictive tasks. Researchers should consider how their biomarker operationalizations may help or hinder the assessment of disease severity.
