TL;DR
This paper introduces a novel framework for estimating whole-brain electrophysiological networks by integrating scalp EEG and intracranial EEG (iEEG) data. Their method uses a state-space model and an Expectation-Maximization algorithm to simultaneously infer brain activity and connectivity, outperforming traditional methods and providing a more comprehensive view of brain dynamics.
What Problem Does This Paper Address?
Understanding the brain's dynamic network activity is crucial for neuroscience and clinical applications. Traditional methods using only scalp EEG for brain network estimation suffer from low accuracy due to the ill-posed nature of source localization. Conversely, invasive iEEG can provide high-resolution data but only covers a partial view of the brain. This leaves a gap in comprehensively mapping the entire brain's dynamic connectivity. The authors aim to bridge this gap by combining the strengths of both non-invasive scalp EEG and invasive iEEG to achieve a more accurate and complete picture of whole-brain electrophysiological networks.
Methods
The core of the proposed method is a principled framework based on state-space models. This framework models the brain's dynamic activity as a latent state that evolves over time, with observed EEG and iEEG signals as noisy measurements.
Key methodological components include:
- Integration of EEG and iEEG: The paper proposes a way to fuse data from both modalities, treating them as complementary sources of information.
- Expectation-Maximization (EM) Algorithm: An EM algorithm is designed to iteratively estimate the model parameters and the latent brain states. This is crucial for inferring both the underlying brain activity (state variables) and the connections between brain regions (connectivity) simultaneously.
- State-Space Model: This mathematical framework is well-suited for modeling dynamic systems with unobserved states, which is ideal for capturing the temporal evolution of brain activity.
- Regularization: To address the ill-posed nature of the inverse problem (estimating source activity from scalp recordings), the method incorporates an L1 regularization term. This encourages sparsity in the connectivity matrix, reflecting the assumption that not all brain regions are directly connected.
- Fixed Interval Smoothing (FIS): For estimating the latent states, the paper utilizes FIS, a technique that retrospectively estimates time series values by considering both past and future observations. This is applied to a combined observation model derived from EEG and iEEG.
- Vector Auto Regressor (VAR) Initialization: The process is initiated using a VAR model to generate initial estimates for the parameters.
Datasets
- Synthetic Data: The framework was initially validated on synthetic data generated to mimic real EEG and iEEG signals. This allowed for controlled evaluation of performance under various conditions.
- Real Data: The method was then applied to a dataset collected from 15 epilepsy patients undergoing intracranial monitoring. This dataset included simultaneous scalp EEG and iEEG recordings during a modified Sternberg working memory task.
Key Insights and Ideas
- Synergy of EEG and iEEG: The most significant insight is that combining scalp EEG and iEEG provides a substantial improvement in brain network estimation accuracy and completeness compared to using either modality alone.
- Dynamic Network Analysis: The framework moves beyond static connectivity measures to capture the dynamic nature of brain networks, revealing how connections change over time during cognitive tasks.
- Whole-Brain Perspective: By integrating data from both superficial and deep brain regions, the method offers a more holistic view of brain connectivity, which is often missed by single-modality approaches.
- Importance of iEEG in Low SNR: The study highlights that iEEG is particularly valuable when scalp EEG signals are noisy (low SNR), helping to maintain accurate connectivity estimation where traditional methods would fail.
Key Results
- Synthetic Data: The proposed method demonstrated improved performance compared to traditional two-step methods using scalp EEG only, validating the importance of iEEG inclusion.
- Real Data (Working Memory Task):
- The framework successfully delineated information flows between cortical and subcortical regions during the encoding and maintenance phases of a working memory task.
- A significant difference was observed in network activity between the encoding and maintenance phases.
- During the encoding phase, information flow was predominantly from cortical to subcortical regions.
- During the maintenance phase, there was a notable shift in information flow, with a reverse direction from subcortical to cortical regions.
- The results showed stronger overall connectivity during the encoding phase compared to the maintenance phase.
Limitations and Future Work
- Atlas Aggregation: The authors acknowledge that potential position deviations during atlas region aggregation in the preprocessing stage might have affected the precise localization of some connections (e.g., temporal lobe to hippocampus).
- No Empty Room Recordings: The study assumed noise characteristics for EEG and iEEG due to the absence of empty room recordings for calibration.
- Future Work: The paper suggests that this framework provides a foundation for future research in brain network analysis and clinical lesion localization, emphasizing the value of simultaneous EEG and iEEG acquisition. The authors also imply that further refinement of preprocessing steps and model parameters could lead to even more precise estimations.
Summary updated