[Paper Review] NeuroSSM: Multiscale Differential State-Space Modeling for Context-Aware fMRI Analysis

Furkan Genç, Boran İsmet Macun+3 more|arXiv (Cornell University)|Jan 3, 2026

Functional Brain Connectivity Studies0 citations

TL;DR

NeuroSSM introduces a multiscale selective state-space model with a parallel differentiating branch to analyze raw fMRI BOLD time series, achieving efficient, end-to-end context-aware representations that capture fast transients and slow trends.

ABSTRACT

Accurate fMRI analysis requires sensitivity to temporal structure across multiple scales, as BOLD signals encode cognitive processes that emerge from fast transient dynamics to slower, large-scale fluctuations. Existing deep learning (DL) approaches to temporal modeling face challenges in jointly capturing these dynamics over long fMRI time series. Among current DL models, transformers address long-range dependencies by explicitly modeling pairwise interactions through attention, but the associated quadratic computational cost limits effective integration of temporal dependencies across long fMRI sequences. Selective state-space models (SSMs) instead model long-range temporal dependencies implicitly through latent state evolution in a dynamical system, enabling efficient propagation of dependencies over time. However, recent SSM-based approaches for fMRI commonly operate on derived functional connectivity representations and employ single-scale temporal processing. These design choices constrain the ability to jointly represent fast transient dynamics and slower global trends within a single model. We propose NeuroSSM, a selective state-space architecture designed for end-to-end analysis of raw BOLD signals in fMRI time series. NeuroSSM addresses the above limitations through two complementary design components: a multiscale state-space backbone that captures fast and slow dynamics concurrently, and a parallel differencing branch that increases sensitivity to transient state changes. Experiments on clinical and non-clinical datasets demonstrate that NeuroSSM achieves competitive performance and efficiency against state-of-the-art fMRI analysis methods.

Motivation & Objective

• Motivate the need for models that capture multi-scale temporal structure in fMRI signals, from fast transients to slow global trends.
• Propose NeuroSSM, a selective state-space architecture that processes raw BOLD time series without relying on FC features.
• Introduce a multiscale backbone plus a parallel differencing branch to jointly model fast and slow dynamics.
• Demonstrate competitive predictive performance on large public fMRI datasets across clinical and non-clinical tasks.

Proposed method

• Process raw, ROI-parcellated BOLD signals with a stack of NeuroSSM modules.
• Within each module, apply a Multiscale Differential State-Space Block (MSD-SSB) across multiple time scales.
• Use a Dual-Stream State-Space Module (Dual-SSM) to concurrently handle time-rescaled data and its first-order temporal differences.
• Fuse scale outputs via a simple residual sum and apply layer normalization and GELU activation.
• Pool temporally and classify with a linear head to obtain subject- or task-level predictions.
• Maintain linear computational cost with sequence length due to diagonal-state SSMs and parallel scale streams.

Experimental results

Research questions

• RQ1 Can NeuroSSM jointly model fast transient dynamics and slow global trends in raw fMRI time series?
• RQ2 Does adding a parallel differencing stream improve sensitivity to rapid brain-state changes?
• RQ3 How does NeuroSSM perform against transformer, CNN/GNN-based, and traditional SSM baselines on HCP and PPMI datasets?
• RQ4 Is NeuroSSM robust under varying training data sizes and across clinical vs. non-clinical tasks?

Key findings

• NeuroSSM achieves competitive or superior accuracy, F1, and AUC across HCP-Rest, HCP-Task, and PPMI compared to baselines.
• Ablation shows both multiscale processing and the differential input stream improve performance, with the full model performing best.
• NeuroSSM shows stronger learning efficiency, particularly in low-data regimes, relative to transformer and SSM baselines.
• Across datasets, NeuroSSM maintains robust performance and competitive AUC, indicating reliability in heterogeneous data.
• Temporal-resolution ablations indicate a 3-scale setup (τ = 1,2,3) is effective without unnecessary complexity.