DeepGraphDMD: Interpretable Spatio-Temporal Decomposition of Non-linear Functional Brain Network Dynamics

Md Asadullah Turja, Martin Styner, Guorong Wu

Abstract

Functional brain dynamics is supported by parallel and overlapping functional network modes that are associated with specific neural circuits. Decomposing these network modes from fMRI data and finding their temporal characteristics is challenging due to their time-varying nature and the non-linearity of the functional dynamics. Dynamic Mode Decomposition (DMD) algorithms have been quite popular for solving this decomposition problem in recent years. In this work, we apply GraphDMD -- an extension of the DMD for network data -- to extract the dynamic network modes and their temporal characteristics from the fMRI time series in an interpretable manner. GraphDMD, however, regards the underlying system as a linear dynamical system that is sub-optimal for extracting the network modes from non-linear functional data. In this work, we develop a generalized version of the GraphDMD algorithm -- DeepGraphDMD -- applicable to arbitrary non-linear graph dynamical systems. DeepGraphDMD is an autoencoder-based deep learning model that learns Koopman eigenfunctions for graph data and embeds the non-linear graph dynamics into a latent linear space. We show the effectiveness of our method in both simulated data and the HCP resting-state fMRI data. In the HCP data, DeepGraphDMD provides novel insights into cognitive brain functions by discovering two major network modes related to fluid and crystallized intelligence.

Introduction

The human brain is a dynamic system where regions form temporary, oscillating networks even at rest. Understanding these functional dynamics is a cornerstone of modern neuroscience, yet researchers face a fundamental conflict between modeling power and interpretability. Traditional methods like Principal Component Analysis (PCA) or Independent Component Analysis (ICA) assume static networks, failing to capture the flickering, time-varying nature of neural activity. While newer methods like Dynamic Mode Decomposition (DMD) account for temporal movement, they typically assume linear dynamics, which is a poor fit for the complex, non-linear synchronizations observed in fMRI data.

Deep learning offers the power to model non-linearity but often results in "black box" models that are difficult for neuroscientists to interpret. DeepGraphDMD addresses this by bridging the gap: it leverages the expressive power of deep neural networks while maintaining the transparency of linear decomposition methods.

Methodology

DeepGraphDMD is built upon Koopman Operator Theory, which posits that a complex, non-linear system can be "lifted" into a higher-dimensional space where its dynamics behave linearly. The framework uses a deep autoencoder to learn these Koopman eigenfunctions for graph-structured data.

Overview of the DeepGraphDMD framework, showing the transformation of non-linear graph dynamics into a latent linear space for interpretable mode extraction.

Key Technical Innovations

  • Preserving Edge Identity: To ensure the model identifies recognizable brain networks, DeepGraphDMD embeds individual regions of interest (ROIs) first and then computes their relationships (Pearson correlation) in the latent space. This preserves the identity of connections between specific nodes.
  • Sparse Koopman Operator: To manage the computational complexity of $O(N^2)$ interactions between brain regions, the authors introduced a sparse operator. This operator focuses on interactions between edges that share a common endpoint, aligning the model with biological plausibility and reducing noise.
  • Window-based Analysis: Recognizing that brain states are transient, the model is applied to sliding windows of fMRI data, allowing it to capture the evolution of network modes over time.

Results and Discussion

The researchers validated DeepGraphDMD using both synthetic simulations and high-resolution resting-state fMRI data from 840 subjects in the Human Connectome Project (HCP) Dense Connectome dataset.

(a) Comparison of mode recovery in simulated data against PCA and ICA. (b-c) Visualization of discovered brain network modes related to cognitive functions.

Simulation Performance

In controlled simulations with known ground-truth network modes, DeepGraphDMD achieved a Pearson correlation of 0.98, significantly outperforming PCA (0.81) and ICA (0.88). This demonstrates the model's superior ability to "de-mix" overlapping signals that traditional methods conflate.

Insights into Human Intelligence

When applied to the HCP dataset, DeepGraphDMD identified specific network modes that correlate strongly with cognitive performance. Notably, the model discovered a specific mode in the 0.08–0.12 Hz frequency band that was invisible to static methods but highly predictive of intelligence.

Key findings include:

  • Fluid Intelligence: DeepGraphDMD improved the prediction of problem-solving and reasoning scores by approximately 28% compared to traditional static connectivity methods.
  • Crystallized Intelligence: The model showed consistent performance gains over linear DMD and ICA in predicting knowledge-based cognitive scores.
  • Network Identification: The framework successfully isolated two major network modes—primarily involving the Frontoparietal and Default Mode Networks—that serve as the primary drivers of these cognitive correlations.

Conclusion and Future Directions

DeepGraphDMD demonstrates that while linear approximations are useful, accounting for non-linearity via the Koopman operator yields cleaner, more biologically relevant brain networks. While the model requires careful hyperparameter tuning and significant computational resources compared to linear methods, its ability to uncover hidden, oscillating modes provides a powerful new tool for clinical neuroscience.

Future work may extend this framework to study non-linear neural disruptions in clinical populations, such as individuals with Alzheimer’s disease or schizophrenia, where traditional linear models may overlook critical dynamic biomarkers.

Summary updated

Abstract

Functional brain dynamics is supported by parallel and overlapping functional network modes that are associated with specific neural circuits. Decomposing these network modes from fMRI data and finding their temporal characteristics is challenging due to their time-varying nature and the non-linearity of the functional dynamics. Dynamic Mode Decomposition (DMD) algorithms have been quite popular for solving this decomposition problem in recent years. In this work, we apply GraphDMD -- an extension of the DMD for network data -- to extract the dynamic network modes and their temporal characteristics from the fMRI time series in an interpretable manner. GraphDMD, however, regards the underlying system as a linear dynamical system that is sub-optimal for extracting the network modes from non-linear functional data. In this work, we develop a generalized version of the GraphDMD algorithm -- DeepGraphDMD -- applicable to arbitrary non-linear graph dynamical systems. DeepGraphDMD is an autoencoder-based deep learning model that learns Koopman eigenfunctions for graph data and embeds the non-linear graph dynamics into a latent linear space. We show the effectiveness of our method in both simulated data and the HCP resting-state fMRI data. In the HCP data, DeepGraphDMD provides novel insights into cognitive brain functions by discovering two major network modes related to fluid and crystallized intelligence.
BibTeX
@inproceedings{turja2023deepgraphdmd,
  author = {Md Asadullah Turja and Martin Styner and Guorong Wu},
  title = {{DeepGraphDMD: Interpretable Spatio-Temporal Decomposition of Non-linear Functional Brain Network Dynamics}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2023},
  doi = {10.1007/978-3-031-43993-3\_35},
}
Figure