Modeling individual-level long-range brain dynamics from short fMRI scans

Functional magnetic resonance imaging (fMRI), which enables noninvasive measurement of Blood Oxygen Level Dependent (BOLD) brain dynamics, is a cornerstone for studying human brain function and connectivity. In real-world settings, however, fMRI acquisitions are often constrained by limited scanner time and subject compliance. These short scans frequently yield unstable functional connectivity estimates, limiting the reliability of brain-based markers (e.g., connectome) and reducing confidence in downstream tasks. Existing computational approaches only partially address this challenge, as they often struggle to achieve both reliable individual-level reconstruction and strong cross-cohort generalization from short observational windows. Here, we present BOLD-Cast, a computational framework for reconstructing long-range BOLD dynamics from short fMRI scans. BOLD-Cast jointly learns cohort-invariant and subject-specific functional topology to enable individualized reconstruction with explicit temporal anchoring. Trained on UK Biobank data (n > 30000), the framework preserves highest functional connectivity topology with a similarity score of 0.996, while supporting stable reconstruction over sequences up to 2 times the input length. Moreover, our approach demonstrates strong zero-shot generalization on 4 unseen independent datasets (HCP-YA, HCP-A, HCP-D, ABIDE), spanning adolescence to older adulthood and cohorts with neuropsychiatric conditions. We further show that reconstructed fMRI signals improve autism spectrum disorder classification accuracy by 7.2% compared with using short scans alone, and enhance cognition-related prediction performance. Together, these results establish a generalizable framework for improving brain-based inference from short-duration fMRI across heterogeneous cohorts.