Evaluating Dementia Neuroanatomic Risk of a Clinical Cohort Using MRI

Abstract

AbstractBackgroundMRI is a widely used modality to evaluate dementia risk. Here, we extend previous work by estimating the Alzheimer’s disease pattern similarity (AD‐PS) scores for participants of a community‐dwelling older adult cohort to evaluate neuroanatomic risk of AD using MRI.MethodBaseline MRI data were available from 522 participants from the Wake Forest Alzheimer’s Disease Research Center’s (ADRC) Clinical Core with mean age 70.2, 81.2% White and 17.6% African American. Of those, 253 were cognitively impaired (dementia and mild cognitive impairment (MCI)) and 138 underwent [11C]PiB PET (MCI = 48 and dementia = 19). Global PiB SUVR was averaged from a cortical region of interest sensitive to early AD. Participants were classified as amyloid positivity using a previously defined threshold (≥1.21 SUVR). AD‐PS scores were estimated based on GM probability maps according to previously reported machine learning methodology[1]. High‐dimensional classifiers were trained using ADNI MRI data. To generate the AD‐PS scores, MRI data from the ADRC cohort was provided as input to the classifiers.We investigated associations that AD‐PS scores [range 0‐1] had with amyloid positivity and cognitive impairment. All analyses were adjusted for age, sex, race and education. AD‐PS scores discrimination of participants with dementia versus normal cognition (CN) was compared to hippocampal volume and cortical thickness in temporal regions based meta‐ROI which are sensitive to AD.ResultAD‐PS scores were strongly associated with cognitive impairment: OR = 4.1, CI: 95% [2.9‐5.9], p<0.0001. In addition, they were associated with amyloid positivity with OR = 2.8,CI: 95% [1.7‐4.7], p<0.001, N = 138. These amyloid positivity results did not hold in analyses limited to CN participants only(N = 71). Performance when discriminating patients with dementia from CN individuals was highest for AD‐PS score (AUC = 92.2, CI: 95% [86.9‐97.5]), followed by hippocampal volume (AUC = 86.5, CI: 95% [81.1‐91.2]) and cortical thickness in temporal regions based meta‐ROI (AUC = 81.5 CI: 95% [74.4‐88.6], See‐Figure).ConclusionThis work highlights further the generalization performance of our machine learning algorithm. AD‐PS scores, a measure of dementia neuroanatomic risk, strongly discriminated individuals with dementia from CN individuals in the Wake Forest ADRC.