Interpretable Hierarchical RNNs for rs-fMRI: Promise and Limits of Individualized Brain Dynamics

Abstract

Modeling individual brain dynamics from resting-state fMRI (rs-fMRI) remains challenging due to substantial inter-subject variability, noise, and limited data length per subject. Here, we systematically evaluate whether hierarchical shallow piecewise-linear recurrent neural networks (shPLRNNs), recently introduced as interpretable dynamical system reconstruction models, can generate individualized rs-fMRI time series while preserving subject-specific functional connectivity structure. We applied the framework to 1,423 rs-fMRI samples from healthy participants of the Marburg-Munster Affective Disorders Cohort Study (MACS). Simulated rs-fMRI data reproduced substantial empirical FC structure, with comparable reconstruction accuracy on the validation and held-out test sets. Generalization to unseen individuals was heterogeneous and strongly depended on how typical a subjects connectivity pattern was relative to the training cohort, with template similarity explaining 37% of variance in reconstruction accuracy. Learned subject-specific parameters exhibited significant test-retest stability and higher within-subject than between-subject similarity on longitudinal data from two different timepoints, supporting their interpretation as individualized dynamical markers. Associations between individual parameters and demographic or cognitive variables were statistically significant but modest in effect size, and predictive performance remained below that obtained using empirical rs-fMRI features directly. Empirical FC was used as a reference for static subject information rather than as a target to be outperformed. Together, these results suggest that hierarchical shPLRNNs can extract meaningful and partially stable individual-specific dynamical structure from rs-fMRI data. The findings delineate key trade-offs between model expressivity, generalization and subject specificity, and point to directions for future methodological refinement in individualized brain modeling. Graphical AbstractA hierarchical dynamical RNN captures substantial individual rs-fMRI functional connectivity structure using compact subject-specific parameters embedded in shared population dynamics. The resulting representations generalize to held-out subjects and show test-retest stability, but only modest associations with phenotypic variables. O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=150 SRC="FIGDIR/small/713153v3_ufig1.gif" ALT="Figure 1"> View larger version (61K): org.highwire.dtl.DTLVardef@1b029edorg.highwire.dtl.DTLVardef@90bdd6org.highwire.dtl.DTLVardef@9f6b77org.highwire.dtl.DTLVardef@486c9e_HPS_FORMAT_FIGEXP M_FIG C_FIG