Biomarkers.

Area of Science:

• Neuroimaging
• Machine Learning
• Cognitive Neuroscience

Background:

• Machine learning (ML) models identify Alzheimer's disease (AD) and aging brain atrophy patterns.
• Existing ML tools are primarily validated in White populations, limiting their use in diverse groups.
• This study assesses the generalizability of ML indices in a racially and ethnically diverse neuroimaging sample.

Purpose of the Study:

• To evaluate the generalizability of ML-derived brain atrophy indices (SPARE-AD and SPARE-BA) across diverse racial and ethnic groups.
• To examine the association of AD-like and aging-related brain atrophy patterns with cognitive function in a multi-ethnic cohort.

Main Methods:

• Utilized MRI data from 6,700 scans across four studies (FHS, NACC, CHS, NOMAS).
• Applied robust statistical harmonization to remove scanner effects and derive SPARE-AD and SPARE-BA indices.
• Examined associations of SPARE-AD with cognitive impairment and SPARE-BA with global cognition using mixed-effects models across racial/ethnic groups.

Main Results:

• SPARE-AD index indicating >50% probability of AD-like atrophy significantly increased odds of cognitive impairment (OR=10.37), consistent across races.
• A 1-year increase in SPARE-BA-predicted brain age was associated with decreased global cognition (β=-0.086).
• The SPARE-BA association with cognition was not significant in Hispanic individuals and was more consistent across groups in those <75 years old.

Conclusions:

• Harmonization techniques and diverse training data enable ML index expansion to various racial and ethnic groups.
• Older Hispanic individuals may show distinct relationships between aging brain patterns and cognition.
• Potential differences in neuropathology or other factors may underlie observed variations in cognitive associations.