Spatiotemporal characterisation of information coding in the multiple demand network

Hamid Karimi-Rouzbahani, N. Rich Anina, Alexandra WoolgarImaging 2026 · DOI

TL;DR

This study tackles the challenge of understanding how the brain's "multiple-demand network" (MDN) processes information over time and space. By combining fMRI (high spatial resolution) and MEG (high temporal resolution) data, researchers found that different parts of the MDN specialize in processing different types of information (stimulus details, rules, responses) at distinct times during a task, revealing a dynamic and segregated information processing system.

What Problem Does This Paper Address?

The "multiple-demand network" (MDN), also known as the frontoparietal control network or cognitive control network, is a set of brain regions thought to be crucial for flexible, goal-directed behavior. It's believed to integrate information from various cognitive systems to handle complex tasks. While fMRI studies show that MDN regions are active across many demanding tasks, their precise role in how and when different types of information are encoded remains unclear due to fMRI's limited temporal resolution. Previous research using time-resolved methods like MEG has shown rapid dynamic changes in information coding, but linking these temporal dynamics to the specific spatial locations within the MDN has been difficult. This paper aims to bridge this gap by using a novel fusion technique to investigate the spatiotemporal characteristics of information coding within the MDN and visual areas during a complex task.

Methods

The core of this study lies in its innovative approach to combining neuroimaging data. The researchers employed:

  • Multivariate Pattern Analysis (MVPA): This standard technique was used on both fMRI and MEG data to decode specific types of information (stimulus location, rules, responses) from brain activity.
  • MEG-fMRI Fusion (Commonality Index): This is the novel contribution. They developed a "Commonality Index" that quantifies the shared variance between fMRI data (from specific regions of interest - ROIs), MEG data (at specific time points), and theoretical "model" RDMs representing different task features (stimulus across hemifields, stimulus within hemifields, rules, responses). This allows them to identify when and where specific information is represented in both modalities simultaneously, overcoming the limitations of each individual technique.
  • Regions of Interest (ROIs): They defined specific brain regions, including early visual cortex (EVC), lateral occipital complex (LOC), and key MDN sub-networks (frontoparietal: IPS, IFS; cingulo-opercular: AI/FO, ACC/pre-SMA).

Datasets

The study utilized two datasets:

  • fMRI Data: Collected from 30 participants at Macquarie University performing a visual stimulus-response mapping task.
  • MEG Data: Re-analyzed data from a previous study (Robinson et al., 2022) using the same task paradigm with a different group of participants.

No new benchmarks were introduced; rather, the study leverages existing data and a novel fusion method to analyze them.

Key Insights and Ideas

  • Dynamic Specialization within the MDN: The MDN isn't a monolithic processing unit. Different sub-networks within it (frontoparietal vs. cingulo-opercular) show distinct temporal patterns of information coding, suggesting specialized roles.
  • Prioritization of "Difficult" Information: While early visual areas process basic stimulus features (across hemifields), the MDN, particularly later in the trial, prioritizes more detailed, "difficult" stimulus information (within hemifields) and task rules. This highlights the MDN's role in higher-level discrimination.
  • Temporal Flow of Information: The fusion analysis revealed a potential temporal flow of information, with visual areas processing basic stimulus information earlier, followed by the MDN encoding more complex aspects of the task.
  • Sub-network Dissociation: The frontoparietal sub-network appears to be more dynamic and responsive to transient task demands, while the cingulo-opercular sub-network shows more sustained, "tonic" coding, potentially maintaining task sets.

Key Results

  • Stimulus Information: Early visual cortex (EVC) and LOC showed strong decoding of across-hemifield stimulus information. The MDN, however, showed stronger coding of within-hemifield stimulus information, especially later in the trial.
  • Rule Information: Rule information was primarily encoded in MDN regions, appearing later in the trial and more prominently around the time of response. EVC showed some rule coding, likely reflecting later time points.
  • Response Information: Only the IPS region within the MDN showed reliable response information coding in the fMRI data.
  • MEG-fMRI Fusion: The commonality analysis revealed significant temporal differences in information coding across MDN sub-networks. For instance, the frontoparietal sub-network showed a transition from across- to within-hemifield stimulus dominance, while the cingulo-opercular sub-network exhibited flatter, more sustained patterns. Rule information peaked before the response in most MDN areas.

Limitations and Future Work

  • Limitations:
    • The fusion analysis is only effective when effects are present in both modalities; subtle effects present in only one modality might be missed.
  • The study used separate groups for fMRI and MEG, which might limit the detection of very subtle or less generalizable effects.
  • The ROI-based approach, while providing interpretability, doesn't allow for conclusions about brain areas outside the defined ROIs.
  • Future Work:
    • Exploratory whole-brain analyses could reveal the involvement of other brain regions.
  • Investigating how information coding dynamics change under different circumstances (e.g., memory recall, selective attention) would be valuable.
  • Causal manipulation studies (e.g., TMS) could further establish the necessity of these temporal patterns for task execution.
  • Assessing the generalizability of the MDN's preference for difficult stimulus aspects across other task dimensions (rule, response) is needed.

Summary updated

Abstract

Abstract The multiple-demand network (MDN), a set of highly interconnected, domain-general regions, which is active across a wide variety of cognitively demanding tasks, is thought to support cognitive functions by integrating distinct types of information depending on the task. However, the spatiotemporal characteristics with which each node in the MDN encodes information remains unclear. We collected fMRI and MEG data from separate participants performing a complex visual stimulus-response mapping task. We used multivariate pattern analysis (MVPA) to decode various task-related types of information—stimulus details, motor responses, and mapping rules—in both the MDN and visual areas. We used model-based MEG-fMRI fusion to compare the high temporal resolution data from MEG with high spatial resolution data from fMRI, extracting commonalities that reflect both the timecourse and location with which these different task features were represented. Early on, visual regions encoded information about the visual hemifield of the stimulus, while later, the MDN encoded the fine-grained details of the stimuli within the same hemifield and the task rules. We observed distinct temporal profiles of information coding for the cingulo-opercular vs. frontoparietal sub-networks of the MDN. This study offers insights into the dynamic information processing of the MDN and provides information-coding-based support for at least two sub-networks within the multiple-demand network.
BibTeX
@inproceedings{karimirouzbahani2026spatiotemporal,
  author = {Hamid Karimi-Rouzbahani and N. Rich Anina and Alexandra Woolgar},
  title = {{Spatiotemporal characterisation of information coding in the multiple demand network}},
  booktitle = {Imaging},
  year = {2026},
  doi = {10.1162/imag.a.1171},
}
Figure