Benchmarking foundation models for resting-state functional magnetic resonance imaging time series in cognitive aging research

Readout choice for foundation-model representations of resting-state fMRI time series is linked to improved selected concurrent cognitive associations, but primary representations show no consistent superiority over functional connectivity.

• Why it matters: Smaller cognitive aging studies need better brain-derived features, yet it is unclear whether foundation-model embeddings improve concurrent cognition, predict follow-up cognition, or characterize cognitive change beyond baseline measures.
• What they did: The study benchmarked nine configurations from five model families on 3,103 scans from 1,653 participants across two observational and five intervention aging cohorts, comparing frozen representations, readout changes, post-training, and embedding alignment.
• The result: Primary frozen representations did not consistently improve concurrent associations or diagnostic classification, longitudinal analyses found no added value beyond baseline clinical measures, and readout changes improved selected concurrent associations without updating model weights.

AI summary of the abstract · may contain errors · not medical advice · read the paper.