TL;DR
This paper introduces a novel, data-driven approach to discover recurring patterns (motifs) in brain connectivity networks. By leveraging the gSpan algorithm, they overcome limitations of previous methods that required pre-defined motif sizes, enabling the identification of more complex and adaptive substructures. This method allows for a deeper understanding of how brain networks organize and change during different cognitive states.
What Problem Does This Paper Address?
Analyzing brain connectivity, particularly functional connectivity derived from electroencephalography (EEG), is crucial for understanding brain function. A key challenge is identifying recurring patterns of interconnections, known as "network motifs," which are thought to be fundamental building blocks of complex networks. Traditional methods for motif discovery in brain networks have a significant limitation: they require researchers to pre-specify the size of the motifs to be detected (e.g., motifs of 2, 3, or 4 nodes). This is problematic because the number of possible motifs grows exponentially with size, making it computationally prohibitive to search for larger, potentially more informative, patterns. Furthermore, these methods often rely on a fixed repertoire of motifs, hindering the discovery of novel or adaptive substructures that might emerge in response to different cognitive demands. This paper aims to address this by proposing a method that can discover motifs of any size in a data-driven manner, allowing for a more flexible and comprehensive analysis of brain network organization.
Methods
The core of this paper's methodology lies in adapting the gSpan (graph-based Substructure pattern mining) algorithm for brain connectivity analysis. gSpan is a powerful algorithm for discovering frequent subgraph patterns in a collection of graphs without requiring a pre-defined motif size.
Here's a breakdown of their approach:
- Data-Driven Motif Discovery: Instead of defining motifs beforehand, gSpan explores a hierarchical search space (represented as a DFS Code Tree) to identify patterns that appear frequently across a dataset of brain networks. This "bottom-up" approach allows for the discovery of motifs of varying complexity.
- Functional Connectivity Graphs (FCGs): Brain activity data (EEG) is transformed into FCGs. For static analysis, these are weighted adjacency matrices representing connectivity strength between brain regions. For time-varying analysis, they capture how connectivity evolves over time.
- Sparsification and Binarization: To reduce computational burden and focus on essential connections, the weighted FCGs are thresholded (e.g., at 70% of edges) and binarized (edges are either present or absent). This step is crucial for applying gSpan, which typically works on unweighted graphs.
- Contrastive Learning Scheme: A key innovation is their "contrastive learning" approach. This involves:
- Single Condition Analysis: Identifying motifs that are common within a specific brain state (e.g., resting state).
- Contrasting Conditions: Identifying motifs that are significantly more prevalent in one condition (e.g., mental arithmetic) compared to another (e.g., resting state). This is achieved by scoring motifs from the first condition based on their frequency in the second condition, allowing them to rank motifs by their discriminative power.
- Significance Testing: To ensure the discovered motifs are not random occurrences, they employ a surrogate-assisted analysis. This involves comparing the observed motif occurrences against those found in randomly shuffled networks to calculate P-values.
Datasets
The study utilizes functional connectivity graphs derived from electroencephalography (EEG) recordings of 18 healthy adult subjects. These recordings were collected under two conditions:
- Resting State: Subjects were at rest.
- Mental Calculations: Subjects were performing multiplications.
The data was processed for five standard frequency bands: $\theta$ (4-8 Hz), $\alpha_1$ (8-10 Hz), $\alpha_2$ (10-13 Hz), $\beta$ (13-30 Hz), and $\gamma$ (30-45 Hz). Both static (representing longer periods of activity) and time-varying (representing short millisecond-long snapshots) connectivity graphs were analyzed.
Key Insights and Ideas
- Adaptive Motif Discovery: The ability to discover motifs of arbitrary size is a significant advancement, moving beyond the limitations of pre-defined motif sizes and allowing for the identification of more complex and potentially more functionally relevant substructures.
- Data-Driven Repertoire: Instead of relying on a fixed set of known motifs, this method generates a "motif repertoire" directly from the data, making it more adaptable to different types of brain networks and cognitive states.
- Contrastive Learning for Task Specificity: The contrastive learning approach is powerful for identifying motifs that are specifically associated with a particular cognitive task (like mental arithmetic), effectively highlighting brain network changes that occur during specific mental operations.
- Temporal Dynamics of Motifs: Analyzing time-varying graphs reveals that some motifs are transient, appearing and disappearing over time. This "come and go" behavior is subject-dependent and can provide insights into the dynamic recruitment of brain areas during cognitive tasks.
- Topographic Relevance: The discovered motifs are not just abstract graph structures; they are mapped back to the physical locations of brain sensors, allowing for interpretation of their functional significance based on known brain anatomy and function.
Key Results
- Significant Resting State Motifs: The study identified numerous significant motifs (P < 0.001) present in at least 90% of subjects during resting state across all frequency bands. Certain bands (e.g., $\theta$ and $\gamma$) showed more spatially extended motifs, potentially related to internal processing and integrated thoughts.
- Task-Specific Motifs Identified: Using the contrastive learning approach, they successfully identified motifs specific to mental arithmetic tasks in $\theta$, $\beta$, and $\gamma$ bands. These motifs were localized to specific brain regions (e.g., fronto-central, fronto-central-temporal-parietal) and often showed hemispheric lateralization.
- Transient Nature of Task Motifs: Analysis of time-varying FCGs revealed that task-specific motifs can exhibit dynamic temporal patterns, appearing and disappearing. This dynamic behavior was subject-dependent and suggested the transient engagement of brain networks, such as language areas during multiplication.
- Validation of gSpan Efficiency: The paper demonstrates that gSpan can handle brain network datasets of considerable size, making it a computationally feasible approach for analyzing complex brain connectivity.
Limitations and Future Work
- Binarization Arbitrariness: The authors acknowledge that the binarization step (thresholding to create binary graphs) introduces some arbitrariness. Future work will explore the full exploitation of connectivity weights using methods for weighted motif detection to reduce this arbitrariness and potentially enhance the quality of results.
- Computational Load: While gSpan is efficient, analyzing very large networks (e.g., from MEG data with hundreds of nodes) can still increase computational load.
- Generalizability: The study primarily uses EEG data. Future work could extend this approach to other neuroimaging modalities like fMRI or DTI, which have different network sizes and connectivity characteristics.
- Pairwise vs. N-way Connectivity: The current approach primarily focuses on pairwise connectivity. Future research could investigate n-way connectivity estimators to reveal more complex interaction patterns.
Summary updated