The Impact of Denoising Approaches on the Relationship Between Alzheimer's Disease Diagnostic Status and Network Topology Measures

Abstract

Graph theory provides a promising technique to investigate Alzheimer's disease (AD)-related alterations in brain network properties. However, there are discrepancies in the reported disruptions that occur to network topology across the AD continuum. In this study, we examined whether diagnostic group differences in graph metrics are attributed to differences in denoising approach used in fMRI processing. Resting state data from 60 cognitively normal (CN), 55 Mild Cognitive Impairment (MCI), and 38 AD participants from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database were denoised using 11 pipelines including combinations of confound regression (head motion parameters, white matter [WM], cerebrospinal fluid [CSF], global signal), volume censoring (scrubbing, spike regression), and component-based noise removal (Independent Component Analysis-based Automatic Removal of Motion Artifacts [ICA-AROMA], anatomical and temporal component correction). Graph metrics representing network segregation (clustering coefficient, modularity, local efficiency), network integration (largest connected component, path length, global efficiency), and small-worldness were calculated. The results revealed that diagnostic group differences in modularity and local efficiency were dependent on denoising approach, especially in high-parameter regression models in combination with censoring methods (36 parameters and spike regressor or volume censoring). Independent of denoising approach, CN exhibited more segregated (clustering coefficient) but less integrated (largest component, path length, global efficiency) networks than MCI and AD. Independent of diagnosis, denoising strategy significantly affected the magnitude of all metrics, particularly models including global signal regression. Collectively, these results suggest that the directionality of the diagnostic differences in network topology, particularly in global metrics of network segregation, can vary based upon the denoising approach employed, although the effect size is small. Transparent reporting of preprocessing decisions is critical for the accurate interpretation of graph theoretical findings in the context of AD and a better understanding of the mechanisms underlying pathological aging.