[Paper Review] BrainCast: A Spatio-Temporal Forecasting Model for Whole-Brain fMRI Time Series Prediction

Yunlong Gao, Jinbo Yang+6 more|arXiv (Cornell University)|Mar 9, 2026

Functional Brain Connectivity Studies0 citations

TL;DR

BrainCast jointly models inter-ROI spatial interactions and intra-ROI temporal dynamics to forecast whole-brain fMRI time series, improving predictive accuracy over baselines and boosting downstream cognitive ability prediction.

ABSTRACT

Functional magnetic resonance imaging (fMRI) enables noninvasive investigation of brain function, while short clinical scan durations, arising from human and non-human factors, usually lead to reduced data quality and limited statistical power for neuroimaging research. In this paper, we propose BrainCast, a novel spatio-temporal forecasting framework specifically tailored for whole-brain fMRI time series forecasting, to extend informative fMRI time series without additional data acquisition. It formulates fMRI time series forecasting as a multivariate time series prediction task and jointly models temporal dynamics within regions of interest (ROIs) and spatial interactions across ROIs. Specifically, BrainCast integrates a Spatial Interaction Awareness module to characterize inter-ROI dependencies via embedding every ROI time series as a token, a Temporal Feature Refinement module to capture intrinsic neural dynamics within each ROI by enhancing both low- and high-energy temporal components of fMRI time series at the ROI level, and a Spatio-temporal Pattern Alignment module to combine spatial and temporal representations for producing informative whole-brain features. Experimental results on resting-state and task fMRI datasets from the Human Connectome Project demonstrate the superiority of BrainCast over state-of-the-art time series forecasting baselines. Moreover, fMRI time series extended by BrainCast improve downstream cognitive ability prediction, highlighting the clinical and neuroscientific impact brought by whole-brain fMRI time series forecasting in scenarios with restricted scan durations.

Motivation & Objective

• Motivate extending and forecasting whole-brain fMRI time series when data acquisition is limited.
• Formulate fMRI forecasting as a multivariate time series prediction problem across ROIs.
• Jointly model inter-ROI spatial dependencies and intra-ROI temporal dynamics.
• Develop a unified architecture (SIA, TFR, SPA) to produce informative whole-brain features for forecasting.

Proposed method

• Embed each ROI time series as a token via a shared MLP to form X_E in R^{N×D}.
• Use SIAformer layers with Fourier Analysis Network, self-attention, and FFN to learn inter-ROI dependencies and produce H^{spat}.
• Apply Amplifier as the Temporal Feature Refinement module to equalize and decompose energy in the temporal spectrum, capturing low- and high-energy components to form H^{temp}.
• Align spatial and temporal representations through Spatio-temporal Pattern Alignment to obtain H^{global}.
• Forecast future fMRI time series with a linear prediction head from H^{global} to X̂_F, optimized by MSE loss.

Experimental results

Research questions

• RQ1 Can whole-brain fMRI time series forecasting improve data quality under short scan durations?
• RQ2 How can inter-ROI spatial interactions and intra-ROI temporal dynamics be jointly modeled for accurate forecasting?
• RQ3 Do spatio-temporal representations improve downstream cognitive ability prediction when fMRI data are limited?
• RQ4 What is the contribution of each module (SIA, TFR, SPA) to forecasting performance?

Key findings

• BrainCast outperforms state-of-the-art baselines on HCP resting-state and task fMRI datasets across multiple metrics.
• Extending fMRI time series with BrainCast leads to better cognitive ability prediction.
• Ablation shows all three modules (SIA, TFR, SPA) contribute to performance, with TFR having the largest individual impact.
• Visualized cortical attention maps reveal meaningful, task-dependent spatial patterns and hemispheric differences between rest and task states.
• Transformer-based baselines generally outperform RNN/LSTM and some GNN baselines in this task.