Interpretable hierarchical RNNs for rs-fMRI: promise and limits of individualised brain dynamics

Introduction: Individual differences in intrinsic brain dynamics are highly relevant to translational psychiatry, but modelling them from resting-state functional magnetic resonance imaging (rs-fMRI) is difficult because of short time series, noise and marked inter-individual variability. Interpretable hierarchical recurrent neural networks may address this challenge by jointly learning shared cohort-level dynamics and compact subject-specific representations, thereby enabling partial pooling across participants while preserving individual variation.

Aims: We evaluated whether a hierarchical shallow piecewise-linear recurrent neural network (hier-shPLRNN) [1] could (i) simulate individual resting-state brain dynamics and preserve subject-specific functional connectivity (FC), and (ii) yield stable individual parameters associated with demographic and cognitive variables. By combining shared group-level recurrent dynamics with low-dimensional subject-specific parameters, the approach is designed to be data-efficient and attractive for short resting-state scans.

Methods: Rs-fMRI data from healthy participants in the Marburg-Münster Affective Disorders Cohort Study [2] were analysed, yielding 1,423 samples. Data underwent standard preprocessing and was split into training, test and validation sets. The hier-shPLRNN modelled each regional time series using shared group-level recurrent dynamics and an individual parameter vector mapped through shared weights. Hyperparameters were optimised with Optuna, identifying 20 subject-specific parameters and a hidden size of 1294 as the best configuration. Models were trained using generalised teacher forcing. For unseen subjects, group-level parameters were fixed and only individual vectors were fine-tuned, testing transfer of learned cohort dynamics to new individuals with minimal adaptation. Model fit was assessed using Pearson correlation between empirical and simulated FC matrices and dynamic time warping between empirical and simulated multivariate time series. Subject identification matched simulated to empirical FC across participants. Test-retest stability was evaluated in 308 subjects with two sessions using permutation tests, Mann-Whitney U tests and effect sizes. Associations with sex, age, body mass index, education and intelligence quotient were tested using false-discovery-rate-controlled mass-univariate models and nested cross-validated machine learning.

Results: The model reproduced empirical FC with mean correlations of r=0.63 in training, r=0.51 in the test set and r=0.51 in independent validation; corresponding normalised dynamic time warping values were 0.16, 0.20 and 0.16. Validation reconstruction depended substantially on how typical a participant’s connectivity pattern was, with similarity to the training-set mean FC explaining 37% of variance in reconstruction accuracy. Subject identification exceeded chance but remained limited: Top-1, Top-3 and Top-5 accuracies were 10.16%, 18.84% and 23.92% among 945 subjects. Individual parameters showed significant test-retest stability, with higher within-subject than between-subject similarity. Associations with sex, age and body mass index were statistically significant but modest, and machine-learning prediction from hier-shPLRNN parameters remained below prediction from empirical imaging features.

Conclusions: Hier-shPLRNNs can retrieve meaningful cohort-level dynamics and stable low-dimensional individual factors from short resting-state scans. However, current individualisation remains constrained: reconstruction is partly driven by population-level connectivity structure, subject identification is limited, and phenotypic associations are modest. These findings support hier-shPLRNNs as promising but still developing tools for personalised applied neuroscience.

References

1. Brenner M, Weber E, Koppe G, Durstewitz D. LEARNING INTERPRETABLE HIERARCHICAL DYNAMI- CAL SYSTEMS MODELS FROM TIME SERIES DATA. 2025; Available from: https://openreview.net/forum?id=Vp2OAxMs2s
2. Vogelbacher C, Möbius TWD, Sommer J, Schuster V, Dannlowski U, Kircher T, et al. The Marburg-Münster Affective Disorders Cohort Study (MACS): A quality assurance protocol for MR neuroimaging data. NeuroImage. 2018 May;172:450–60.