BrainODE: Dynamic Brain Signal Analysis via Graph-Aided Neural Ordinary Differential Equations

TL;DR

BrainODE tackles three core data-quality issues in dynamic fMRI signals by learning a continuous-time model of ROI activity with neural ODEs. It jointly estimates latent initial states using CNN-based activity encoding, self-attention for long-range dependencies, and dual graphs to encode temporal and spatial ROI relations, which inform ODE inference via GCNs. The framework is trained end-to-end as an autoencoder, combining a KL-regularized initial-state prior with an ODE-driven trajectory generator to reconstruct signals at arbitrary times, improving downstream brain-network analyses. Empirical results on ABIDE and ABCD show substantial gains in AUC and accuracy across multiple base models, and ablation studies confirm the importance of temporal graphs and position encoding for robust performance under missing and misaligned data. The work advances practical brain-signal re-processing, enabling more reliable connectome-based predictions and offering a path toward integrating real structural graphs in future work.

Abstract

Brain network analysis is vital for understanding the neural interactions regarding brain structures and functions, and identifying potential biomarkers for clinical phenotypes. However, widely used brain signals such as Blood Oxygen Level Dependent (BOLD) time series generated from functional Magnetic Resonance Imaging (fMRI) often manifest three challenges: (1) missing values, (2) irregular samples, and (3) sampling misalignment, due to instrumental limitations, impacting downstream brain network analysis and clinical outcome predictions. In this work, we propose a novel model called BrainODE to achieve continuous modeling of dynamic brain signals using Ordinary Differential Equations (ODE). By learning latent initial values and neural ODE functions from irregular time series, BrainODE effectively reconstructs brain signals at any time point, mitigating the aforementioned three data challenges of brain signals altogether. Comprehensive experimental results on real-world neuroimaging datasets demonstrate the superior performance of BrainODE and its capability of addressing the three data challenges.

Figures (4)

• Figure 1: Motivation of BrainODE. As elaborated in the introduction, the current pipeline ignores data challenges, leading to serious problems. The current methods are unable to fully address these issues. Therefore, we propose BrainODE, which tackles these data challenges, thus making the comprehension of brain networks and potential applications in clinical settings possible.
• Figure 2: Overview of BrainODE. The left panel demonstrates the raw dynamic brain signal from the fMRI sequence where ROI#1 shows an ideal signal and ROI#2-4 illustrate the "missing value", "irregular sample", and "sampling misalignment" challenges, respectively. The middle panel describes the BrainODE framework which leverages Short-Term and Long-Term Time Encoders to learn a latent embedding for each input ROI signal. The embeddings are then refined through two distinct GCN layers, which learn the spatial and temporal relations among ROI channels, aiding in the derivation of initial states for ODE inference. An ODE solver and decoder, shown on the right panel, are utilized to obtain the brain signals at desired regular time steps, which can lead to enhanced performance in subject classification.
• Figure 3: Effectiveness in addressing irregular samples and sampling misalignment. Performances are grouped by offsets in the left figure and grouped by frequencies in the right figure.
• Figure 4: Hyperparameter analysis.