MAPPING THE HIDDEN GEOMETRY OF RESTING‑STATE BRAIN DYNAMICS WITH RESERVOIR COMPUTING

Brain activity at rest in fMRI is typically characterized using correlational measures between BOLD signals averaged across predefined parcels, a framework known as functional connectivity (FC). This approach implicitly assumes that the underlying dynamics have a dimensionality no greater than the number of parcels. However, emerging evidence indicates that resting-state BOLD activity exhibits a higher intrinsic dimensionality, reflecting a higher-order organization that FC alone cannot capture. In Reservoir Computing (RC), high‑dimensional representations generated by recurrent neural networks (RNNs) are used to reconstruct input signals through a linear decoder. While classical RC focuses on reconstructing signals independently, we investigate whether the latent dynamics of a single reservoir can be decoded through a shared low‑dimensional basis when driven by rs-fMRI time-series from different subjects. A geometric procedure ensures that each decoding mode captures a distinct component of the underlying latent dynamics. This framework provides a unique coordinate system for characterizing each subject via the strengths associated with these modes. We show that gradual changes in the input produce measurable drifts in this coordinate space, revealing latent dimensions of neural‑like dynamics that remain hidden under FC‑based representations. We validated this framework by constructing digital twins of benchmark time-series, and then applied it to resting‑state fMRI data from the Human Connectome Project (HCP). In this shared coordinate space, subjects’ macroscopic characteristics—such as age—could be preliminarily decoded from their latent positions, demonstrating that the inferred latent modes capture meaningful individual‑level structure beyond traditional FC measures.