Manifold alteration between major depressive disorder and healthy control subjects using dynamic mode decomposition in resting-state fMRI data

Hidenori Endo, Shigeyuki Ikeda, Kenichiro Harada, Hirotaka Yamagata, Toshio Matsubara, Koji Matsuo, Yoshinobu Kawahara, Okito YamashitaFrontiers in Psychiatry 2024 · DOI

TL;DR

This paper introduces a novel method using Dynamic Mode Decomposition (DMD) and t-SNE to analyze the dynamic brain states captured by resting-state fMRI in individuals with Major Depressive Disorder (MDD) and healthy controls (HCs). They found that MDD patients don't develop entirely new brain dynamics but rather tend to "sink" into specific existing dynamic modes, particularly those associated with the default mode and sensory-motor networks, while showing reduced engagement in visual and saliency networks.

What Problem Does This Paper Address?

Major Depressive Disorder (MDD) is a significant global health issue, and understanding its underlying neural mechanisms is crucial for developing better treatments. While previous studies have used resting-state fMRI (rsfMRI) to examine static functional connectivity and energy landscapes, these methods often overlook the temporal dynamics of brain activity. Specifically, they don't fully capture how brain states evolve over time or how individuals might get "stuck" in certain patterns. This paper aims to address this gap by investigating the temporal dynamics of brain activity in MDD patients compared to healthy controls, focusing on how these dynamic modes differ and what this implies about the disorder's neural underpinnings.

Methods

The core of this study lies in the application of Dynamic Mode Decomposition (DMD) to rsfMRI data. DMD is a data-driven technique that decomposes complex, time-varying systems into a set of coherent spatiotemporal patterns (dynamic modes or DMs) and their associated temporal dynamics (eigenvalues).

Here's a breakdown of their approach:

  1. DMD Application: BOLD signals from each subject were fed into a one-stacked time-delay coordinates DMD (tdcDMD) algorithm. This process extracts DMs, which represent coherent spatial patterns of brain activity, and discrete-time eigenvalues, which capture the temporal evolution (frequency and decay rate) of these patterns.
  2. Dimensionality Reduction with t-SNE: To visualize and compare the complex DMs across subjects, the researchers used t-distributed Stochastic Neighbor Embedding (t-SNE). This algorithm reduces the high-dimensional DM data into a 2D manifold, allowing for easier visualization of how DMs cluster and separate between groups.
  3. Density Estimation and Ratio Estimation: Kernel density estimation was applied to the 2D manifolds to visualize the distribution of DMs for both HCs and MDD groups. Crucially, they then employed Kernel Density Ratio Estimation (specifically, RuLSIF) to quantify the differences in these distributions. This method highlights areas where the density of DMs is significantly higher in one group compared to the other, indicating a tendency to "sink" into specific states.
  4. Statistical Significance: Permutation tests were used to identify areas in the 2D manifold that showed statistically significant differences in density ratios (p < 0.05), ensuring that observed differences were not due to random chance.

What's novel here is the combination of DMD for capturing dynamic brain states with t-SNE and density ratio estimation for a robust, data-driven comparison of these states between clinical groups, without requiring pre-defined network assignments or signal binarization.

Datasets

The study utilized a large, multi-site dataset from the Japanese Strategic Research Program for the Promotion of Brain Science (SRPBS) dataset (27). This dataset included a total of 845 participants: 543 healthy controls (HCs) and 302 individuals with Major Depressive Disorder (MDD). The MDD group had an average Beck Depression Inventory-II (BDI-II) score of 28.1 ± 10.5, while HCs scored 7.5 ± 6.3. The data was collected from multiple sites with both unified and non-unified protocols, which the authors addressed during preprocessing.

Key Insights and Ideas

  • "Sinking" into States, Not Creating New Ones: The most striking finding is that MDD doesn't appear to create entirely new brain dynamics. Instead, individuals with MDD tend to "sink" into specific, pre-existing dynamic modes more frequently than healthy controls. This suggests a potential loss of flexibility in brain state transitions.
  • Network Specificity of Altered Dynamics: The "sinking" behavior is not random. MDD patients showed increased occurrence in dynamic modes associated with the default mode network (DMN) and sensory-motor networks (SMN). Conversely, they showed reduced occurrence in modes linked to the visual network (VN) and saliency network (SN).
  • Temporal Dynamics Matter: This study emphasizes that looking beyond static connectivity is crucial. The temporal evolution and stability of brain states, as captured by DMD, reveal differences that static measures might miss.
  • Potential for Biomarkers: The method's ability to differentiate between MDD and HCs suggests its potential for developing biomarkers for diagnosis and treatment monitoring.

Key Results

  • Differential Occurrence of Dynamic Modes: The study identified specific dynamic modes that were either increased or decreased in occurrence in MDD patients compared to HCs.
  • Increased Occurrence: Modes resembling the DMN and SMN were more prevalent in MDD.
  • Reduced Occurrence: Modes resembling the VN and SN were less prevalent in MDD.
  • Manifold Visualization: t-SNE visualizations clearly showed distinct clustering patterns for HCs and MDD on the 2D manifold of DMs, with MDD points appearing more concentrated in certain regions.
  • Density Ratio Differences: Kernel density ratio estimation revealed significant differences in DM distribution between the groups, supporting the idea of MDD patients "sinking" into specific states.
  • Classification Potential: The authors mention that individual-level classification using this DMD-based approach achieved balanced accuracy slightly better than previous studies using static functional connectivity (sFC).

Limitations and Future Work

  • Computational Cost: The authors acknowledge that t-SNE, while effective for visualization, is computationally intensive, requiring significant time and memory for large datasets. This limits the exploration of all possible parameters.
  • Parameter Selection: Due to computational constraints, the search for optimal parameters for t-SNE was limited. Other parameters, like early exaggeration factor and learning rate, were set using heuristic methods.
  • ROI Selection: The study excluded the cerebellum and white matter from analysis due to potential noise and artifacts. Future work could explore methods to incorporate these regions.
  • Integration with Structural Data: The authors highlight the importance of investigating the relationship between these dynamic modes and structural brain abnormalities (cortical thickness and surface area) to deepen the understanding of MDD mechanisms.
  • Subtyping MDD: The study did not differentiate between subtypes of MDD (e.g., melancholic vs. non-melancholic), which could explain some discrepancies with previous studies using energy landscape methods. Future research could explore these subtypes.
  • Broader Applications: The methodology could be extended to other psychiatric disorders and neurological conditions.

Summary updated

Abstract

To the best of our knowledge, this study showed initial evidence of the alteration of occurrence of the dynamic modes between MDD and HCs. To deepen understanding of how the alteration of the dynamic modes emerges from the structure, it is vital to investigate the relationship between the dynamic modes, cortical thickness, and surface areas.
BibTeX
@inproceedings{endo2024manifold,
  author = {Hidenori Endo and Shigeyuki Ikeda and Kenichiro Harada and Hirotaka Yamagata and Toshio Matsubara and Koji Matsuo and Yoshinobu Kawahara and Okito Yamashita},
  title = {{Manifold alteration between major depressive disorder and healthy control subjects using dynamic mode decomposition in resting-state fMRI data}},
  booktitle = {Frontiers in Psychiatry},
  year = {2024},
  doi = {10.3389/fpsyt.2024.1288808},
}
Figure