Current gaps in downstream applications of fMRI foundation models: lessons learned from BrainLM & BrainHarmonix on Prevent-AD

Wang, Pichet Binette, Sainath et al. evaluated two recently released fMRI foundation models, BrainLM and BrainHarmonix, using data from PreventAD, a cohort of adults with a family history of Alzheimer's disease. The sample included 343 quality-controlled participants, 218 of whom underwent amyloid PET imaging. The study compared frozen and fine-tuned model features against two baselines—functional connectivity and a "majority class" dummy classifier—with prediction targets of sex, age, progression to mild cognitive impairment (MCI), and β-amyloid burden. The paper reports that both models required substantial modifications to their public code: BrainLM's built-in data preparation pipeline could not be reliably fine-tuned, and fine-tuning was achieved only after switching to the standard Nilearn pipeline; BrainHarmonix's documentation did not specify how to normalize new data. For sex prediction, functional connectivity outperformed all foundation model features; for age-group prediction, BrainHarmonix's structural encoder features were slightly higher than functional connectivity but not significantly so; for MCI and β-amyloid prediction, no feature exceeded the dummy classifier in accuracy, with AUC close to chance…