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

Foundation models trained on large neuroimaging datasets could help smaller studies investigate cognition from brain recordings. However, whether the foundation model-derived representations improve assessment of current cognition, prediction of follow-up cognition, or characterization of cognitive change remains unclear. We evaluated nine configurations from five model families that process regional resting-state functional magnetic resonance imaging time series, using 3,103 scans from 1,653 participants across two observational and five intervention aging cohorts. Primary frozen representations showed no consistent superiority over conventional functional connectivity for concurrent cognitive associations or diagnostic classification. Changing the readout improved selected concurrent cognitive associations without updating model weights, indicating that performance depended partly on how representations were extracted. Additional post-training and embedding alignment yielded gains in selected settings but did not consistently improve performance across outcomes. In the primary longitudinal analyses, imaging representations did not establish added value beyond baseline clinical measures. These findings show that readout choice influences concurrent cognitive associations, while the added value of these representations for longitudinal outcomes remains unestablished.