Predicting behavior through dynamic modes in resting-state fMRI data

Shigeyuki Ikeda, Koki Kawano, Soichi Watanabe, Okito Yamashita, Yoshinobu Kawahara

TL;DR

This paper introduces a novel method to predict individual behavioral differences using dynamic modes (DMs) extracted from resting-state fMRI data. They demonstrate that these spatiotemporal patterns, derived via Dynamic Mode Decomposition (DMD), are more effective than traditional methods like ICA, particularly those in lower frequency bands (<0.2 Hz), and show spatial similarities to known brain networks.

What Problem Does This Paper Address?

Resting-state functional connectivity (FC) in the brain is known to be related to individual differences in behavior, intelligence, and personality. However, traditional FC measures are static, capturing only average correlations. While dynamic FC offers a richer picture, extracting meaningful, predictive features from its complex temporal fluctuations has been challenging. Existing methods like sliding window approaches have issues with window length selection, and Hidden Markov Models (HMMs) can be difficult to interpret. This paper aims to bridge this gap by establishing a robust method to predict individual behaviors from the spatiotemporal characteristics of dynamic FC, specifically using Dynamic Mode Decomposition (DMD).

Methods

The core of this study lies in applying Dynamic Mode Decomposition (DMD) to resting-state fMRI (rs-fMRI) data. DMD is a technique that decomposes complex dynamic systems into a set of "dynamic modes" (DMs), which are coherent spatial patterns with specific temporal frequencies and decay rates.

Here's a breakdown of their approach:

  • DMD Application: For each subject and each of their four rs-fMRI runs, DMD was applied to extract DMs. To handle the high dimensionality and ensure sufficient data for DMD, they constructed an augmented data matrix.
  • Gram Matrix Construction: A key challenge was comparing DMs across subjects, as there's no inherent correspondence between them. To overcome this, they used the Grassmann manifold, a mathematical space where subspaces (represented by sets of DMs) can be treated as points. The similarity between two subjects' DMs was then calculated using a projection kernel on this manifold, resulting in a gram matrix for each subject.
  • Frequency Band Analysis: To understand which temporal frequencies are most informative, they computed separate gram matrices using DMs from seven specific frequency bands (0-0.1 Hz up to 0.6-0.7 Hz).
  • Prediction with Gaussian Process Regression (GPR): For each of the 59 behavioral measures, GPR was used to predict individual scores based on the mean gram matrix (averaged across all runs).
  • Validation and Comparison: They validated their approach using data from the Human Connectome Project (HCP). For comparison, they also performed prediction using temporal Independent Component Analysis (ICA), a common method for extracting spatial patterns.
  • Clustering of DMs: To investigate the spatial characteristics of the DMs, they clustered them and compared the resulting centroids to known resting-state networks (RSNs).

Datasets

The study utilized the Human Connectome Project (HCP) 1200-subjects release. This dataset provided both rs-fMRI data and 59 behavioral measures. The authors selected a subset of 829 subjects who met specific criteria for data quality and completeness. No new benchmarks were introduced; the study focused on applying existing methods to a well-established dataset.

Key Insights and Ideas

  • DMD as a Predictive Tool: The study demonstrates that DMs derived from rs-fMRI can effectively predict individual differences in behavior, suggesting that these spatiotemporal patterns capture meaningful information about brain function related to cognition and other traits.
  • Low-Frequency Dominance: The most significant predictive power comes from DMs in lower frequency bands (<0.2 Hz). This highlights the importance of slower fluctuations in the brain's intrinsic activity for understanding individual differences.
  • DMs Mirror RSNs: DMs in the lower frequency bands (<0.2 Hz) exhibit spatial structures similar to well-known resting-state networks (RSNs) like the Default Mode Network (DMN) and Executive Control (EC) network. This suggests that these established brain networks are not static but possess coherent dynamic properties that are behaviorally relevant.
  • DMD Outperforms ICA: DMD proved to be a more powerful tool than temporal ICA for predicting behavioral measures, underscoring its ability to capture the dynamic, spatiotemporal nature of brain activity more effectively.
  • Grassmann Manifold for Subject Comparison: The use of the Grassmann manifold provided a robust solution to the challenge of comparing subject-specific DMs, enabling a direct assessment of similarity without requiring a predefined correspondence.

Key Results

  • Successful Prediction: The study successfully predicted individual differences in seven out of 59 behavioral measures using DMD-derived DMs.
  • DMD vs. ICA: DMD showed significantly better prediction accuracy (Pearson correlation coefficient) compared to temporal ICA across the behavioral measures.
  • Frequency Band Contribution: DMs from frequency bands below 0.2 Hz were found to be the primary contributors to successful predictions. Specifically, the 0-0.1 Hz and 0.1-0.2 Hz bands showed significant results.
  • Cognitive Measures Dominate: Most of the significantly predicted behavioral measures belonged to the cognition category.
  • RSN Similarity: Magnitude maps of DMs from lower frequency bands (<0.2 Hz) showed strong cosine similarities with established RSNs, indicating that these dynamic modes reflect the underlying organization of known brain networks.

Limitations and Future Work

  • Spatial-Only Prediction: The current study primarily used the spatial characteristics of DMs for prediction. The authors acknowledge that temporal characteristics of DMD also contain valuable information and suggest future work should explore kernels that incorporate both spatial and temporal information.
  • Parcellation Dependency: While DMD can be applied without parcellation, the study used a predefined functional atlas. Future work could explore the impact of different parcellation schemes or a parcellation-free approach.
  • Single-Site Data: The study relied solely on HCP data. Investigating the generalizability of DMD on multi-site rs-fMRI data is recommended due to potential site-specific variations.
  • Phase Map Similarity: The authors note a potential issue with using cosine similarity to compare phase maps with RSNs, as phase is cyclic. They suggest that further research is needed to find more appropriate similarity measures for phase information.
  • Biomarker Potential: The authors propose that future studies should investigate whether DMs can serve as biomarkers for psychiatric diseases, which could lead to a deeper understanding of their clinical relevance.

Summary updated

Abstract

ABSTRACT Dynamic properties of resting-state functional connectivity (FC) provide rich information on brainbehavior relationships. Dynamic mode decomposition (DMD) has been used as a method to characterize FC dynamics. However, it remains unclear whether dynamic modes (DMs), spatial-temporal coherent patterns computed by DMD, provide information about individual behavioral differences. This study established a methodological approach to predict individual differences in behavior using DMs. Furthermore, we investigated the contribution of DMs within each of seven specific frequency bands (0-0.1,…,0.6–0.7 Hz) for prediction. To validate our approach, we confirmed whether each of 59 behavioral measures could be predicted by performing multivariate pattern analysis on a gram matrix, which was created using subject-specific DMs computed from resting-state functional magnetic resonance imaging (rs-fMRI) data of individuals. The prediction was successful, and DMD outperformed temporal independent component analysis, a conventional data decomposition method for extracting spatial activity patterns. Most of the behavioral measures that showed significant prediction accuracies in a permutation test were cognitive-behavioral measures. Our results suggested that DMs within frequency bands &lt;0.2 Hz primarily contributed to prediction. In addition, we found that DMs &lt;0.2 Hz had spatial structures similar to several common resting-state networks. We demonstrated the effectiveness of DMs, indicating that DMD is a key approach for extracting spatiotemporal features from rs-fMRI data.
BibTeX
@article{ikeda2022predicting,
  author = {Shigeyuki Ikeda and Koki Kawano and Soichi Watanabe and Okito Yamashita and Yoshinobu Kawahara},
  title = {{Predicting behavior through dynamic modes in resting-state fMRI data}},
  journal = {NeuroImage},
  year = {2022},
  doi = {10.1016/j.neuroimage.2021.118801},
}
Figure