Abstract
Background Previous studies have leveraged machine learning (ML) analytics to detect brain atrophy patterns that are specific to Alzheimer's disease (AD) and to distinguish them from structural changes related to brain aging. These tools have been primarily developed and validated in White individuals; therefore, their applicability in diverse samples remains uncertain. Herein, we explore the generalizability of ML indices in a racially and ethnically diverse neuroimaging sample. Method MRI data from four studies (FHS, NACC, CHS, NOMAS) including a total of 6,700 scans were processed using robust statistical harmonization pipelines to remove scanner‐related differences and were utilized to derive ML indices that quantify AD‐like (SPARE‐AD) and aging‐related (SPARE‐BA) brain atrophy patterns. A global cognitive score was computed as a weighted sum of each study's cognitive tests, using the first principal component loadings from a PCA containing the normalized scores of the cognitively normal participants as weights. We examined the associations of SPARE‐AD with cognitive impairment (defined as prevalent mild cognitive impairment or dementia) and the associations of SPARE‐BA with global cognition in the total sample and across race and ethnicity using mixed effects models adjusted for age and sex. Result Mean age (SD) was 70.58 (10.97) years; 59% of participants were women (Table 1). A >50% probability for spatial abnormalities consistent with early AD on brain MRI based on the SPARE‐AD index was associated with a tenfold increase in the odds for cognitive impairment (OR[95%CI] = 10.37[8.76,12.27]). Similar associations were observed across race and ethnicity (Table 2). A 1‐year increase in SPARE‐BA‐predicted brain age was associated with an 8.6% of a standard deviation decrease in global cognitive score (β[95%CI] = ‐0.086[‐0.142,‐0.03]). This association was not significant in Hispanic individuals (Table 3). In individuals <75 years of age, associations between SPARE‐BA and global cognition were more consistent across race and ethnicity (Table 3). Conclusion Robust harmonization techniques and ML model training using diverse neuroimaging datasets can facilitate the expansion of ML indices to different racial and ethnic groups. Older Hispanic individuals may have a different relationship between aging‐related brain atrophy patterns and cognition, which might indicate differences in underlying neuropathologies or other unidentified factors.