Tracking causal pathways in TMS-evoked brain responses

Jinming Xiao, Qing Yin, Lei Li, Yao Meng, Xiaobo Liu, Wanrou Hu, Xinyue Huang, Yu Feng, Xiaolong Shan, Weixing Zhao, Peng Wang, Xiaotian Wang, Youyi Li, Huafu Chen, Xujun DuanPLOS Computational Biology 2025 · DOI

TL;DR

This paper introduces a novel framework combining Transcranial Magnetic Stimulation (TMS) with Electroencephalography (EEG) and advanced computational methods to map the dynamic, causal flow of brain activity. By analyzing how a targeted brain perturbation propagates through neural networks in real-time, the study reveals a distinct "diffusion" and "re-entrant" pattern of information flow, offering crucial insights into brain network mechanisms and potential therapeutic applications of TMS.

What Problem Does This Paper Address?

Understanding how information flows and causal relationships are established within the complex network of the human brain is fundamental to neuroscience. Traditional methods often rely on observational data (like resting-state fMRI), which can infer statistical dependencies but struggle to establish true causality. While techniques like Granger Causality and Dynamic Causal Modeling exist, they often analyze data over extended time windows, potentially missing the rapid, dynamic, and stage-dependent nature of brain activity. Specifically, researchers want to understand how a localized brain perturbation, like that induced by TMS, influences activity across the entire brain network and how this influence unfolds over time. This is particularly relevant for understanding the mechanisms behind TMS-based therapies.

Methods

The researchers developed a novel framework that integrates several key techniques:

  • Concurrent TMS-EEG: This is the core experimental paradigm. TMS is used to precisely and temporarily "perturb" a specific brain region (in this case, the left primary motor cortex, M1), while EEG simultaneously records the brain's electrical response with millisecond-level precision. This allows for the direct observation of causal effects.
  • Source Reconstruction: Raw EEG signals are transformed into source-level activity, estimating where in the brain the electrical signals originate. This provides a more anatomically precise understanding of brain activity.
  • Sparse Non-negative Matrix Factorization (sNMF): This dimensionality reduction technique is applied to the source-reconstructed TMS-evoked activity. sNMF decomposes the complex, high-dimensional brain activity into a smaller set of "co-activation modules" (representing spatially distinct networks that tend to activate together) and their corresponding "time-varying weights" (representing how active these modules are over time). This allows for a more interpretable, low-dimensional representation of brain dynamics.
  • Hierarchical Temporal Framework: To capture the dynamic nature of brain responses, the researchers divided the TMS-evoked activity into "states" (short, temporally stable configurations of activity) using the Greedy State Boundary Search (GSBS) algorithm. These states were then grouped into coarser "propagation stages" (e.g., "Activity Diffusion Stage" and "Activity Re-entrant Stage") based on a first-order Markov assumption.
  • Stage-Dependent Phase Slope Index (PSI): This effective connectivity measure was applied within each identified propagation stage. PSI estimates the direction of information flow between brain regions by analyzing phase differences in their signals, particularly how these differences change with frequency. By calculating PSI for each stage, the researchers could track how causal influences evolved over time.

Datasets

The study utilized a publicly available concurrent TMS-EEG dataset collected by the Rogasch group. This dataset consists of data from 20 healthy individuals who received single-pulse TMS over the left primary motor cortex (M1) while their brain activity was recorded. The data was used for both the development and validation of the proposed framework.

Key Insights and Ideas

  • Dynamic, Stage-Dependent Propagation: The most significant insight is that TMS-evoked activity doesn't spread uniformly. Instead, it unfolds in distinct temporal stages, characterized by different patterns of causal influence. This highlights the inadequacy of time-cumulative analyses for fully understanding brain dynamics.
  • The "Activity Diffusion Stage": In the initial phase, the stimulated region (left M1) acts as a primary source, with activity diffusing to both ipsilateral and contralateral brain regions. This suggests a rapid spread of the initial perturbation.
  • The "Activity Re-entrant Stage": A crucial finding is the subsequent "re-entrant" pathway. Here, contralateral brain regions (specifically, right M1) become sources, sending signals back to the originally stimulated region. This suggests a feedback loop and functional integration between hemispheres.
  • Hemispheric Lateralization: The study found pronounced hemispheric lateralization in the spatial distribution of co-activation modules, with distinct modules associated with the stimulated left hemisphere and its contralateral right hemisphere counterpart.
  • Robust Neurobiological Basis: The identified co-activation modules showed dense structural and functional connections within their constituent brain regions, supporting their neurobiological validity.

Key Results

  • The framework successfully decomposed TMS-evoked activity into 10 co-activation modules, revealing distinct spatial patterns.
  • Time-cumulative PSI analysis showed initial propagation from left-hemisphere dominant modules to right-hemisphere dominant modules.
  • Stage-dependent PSI analysis revealed two primary propagation stages:
  • Activity Diffusion Stage: Modules 1 and 2 (ipsilateral to stimulation) acted as sources, propagating activity to both hemispheres.
  • Activity Re-entrant Stage: Modules 5 and 6 (contralateral to stimulation) acted as sources, propagating activity back to the stimulated modules (1 and 2).
  • The study provides preliminary evidence for a loop-like propagation pattern: "stimulation brain regions - contralateral brain regions - stimulation brain regions."

Limitations and Future Work

  • Sensory-Evoked Responses: The authors acknowledge that raw scalp responses might be contaminated by sensory-evoked potentials (e.g., muscle artifacts), though they argue their preprocessing and analysis mitigate this.
  • Volume Conduction: Like other EEG/MEG studies, volume conduction remains a potential limitation, potentially introducing spurious correlations and reducing spatial specificity. The use of dSPM normalization and PSI is noted as a way to address this, but it cannot be entirely eliminated.
  • Single Stimulation Target: The current findings are based on stimulation of the left M1. Future work should validate these findings with different stimulation targets.
  • Data-Driven Stage Identification: The current stage segmentation relies on prior knowledge and a Markov assumption. Future research could explore more data-driven methods for identifying temporal stages and calculating effective connectivity.
  • Clinical Applications: The authors propose that this framework could be used to evaluate the network-level influence of TMS in clinical settings, providing insights into symptom improvement and refining therapeutic strategies.

Summary updated

BibTeX
@article{xiao2025tracking,
  author = {Jinming Xiao and Qing Yin and Lei Li and Yao Meng and Xiaobo Liu and Wanrou Hu and Xinyue Huang and Yu Feng and Xiaolong Shan and Weixing Zhao and Peng Wang and Xiaotian Wang and Youyi Li and Huafu Chen and Xujun Duan},
  title = {{Tracking causal pathways in TMS-evoked brain responses}},
  journal = {PLOS Computational Biology},
  year = {2025},
  doi = {10.1371/journal.pcbi.1013316},
}
Figure