Decoding intracranial EEG data with multiple kernel learning method

Jessica Schrouff, Janaı́na Mourão-Miranda, Christophe Phillips, Josef ParviziJournal of Neuroscience Methods 2015 · DOI

TL;DR

This paper introduces a novel Multiple Kernel Learning (MKL) approach to analyze intracranial EEG (iEEG) data, offering a more sensitive and data-driven way to understand brain activity compared to traditional univariate methods. By simultaneously considering multiple dimensions of the iEEG signal (frequency bands, recording sites, and time), MKL can automatically identify and highlight the most informative features for decoding cognitive states, paving the way for new hypotheses about brain function.

What Problem Does This Paper Address?

Analyzing electrophysiological data like intracranial EEG (iEEG) is crucial for understanding brain function, especially during cognitive tasks. Traditionally, researchers rely on univariate methods to examine specific aspects of the signal, such as the power in a particular frequency band at a single recording site. While effective for testing pre-defined hypotheses, these methods can miss subtle, distributed patterns of brain activity that span multiple frequencies, locations, and time points. This limitation hinders a comprehensive understanding of complex cognitive processes. The paper addresses this gap by proposing a multivariate approach that can explore these multiple dimensions simultaneously, offering a more holistic and exploratory way to decode brain states.

Methods

The core of this paper is the application of Multiple Kernel Learning (MKL) to iEEG data. MKL is a machine learning technique that combines multiple "kernels," where each kernel represents a different model or feature set. In this study, each kernel was constructed based on the averaged power of the iEEG signal within a specific frequency band (δ, θ, α, β, low-γ, high-γ) at a particular recording site (electrode).

The MKL algorithm used here is a sparse MKL variant (specifically, the "simpleMKL" implementation). This sparsity constraint is key: it automatically selects only the most relevant kernels (and thus, the most informative electrode-frequency band combinations) that contribute to the final decision boundary, effectively performing automatic feature selection. The algorithm uses a Support Vector Machine (SVM) framework to learn a decision function that discriminates between cognitive states (in this case, "math" vs. "non-math" tasks). The MKL approach then weights these individual kernel decision functions to create a global decision function. This weighting process reveals which kernels (electrode-frequency band combinations) are most important for the discrimination.

The authors also employed a 10-fold cross-validation scheme to assess model performance and a permutation test for statistical significance, ensuring robust evaluation of their findings.

Datasets

The study utilized an iEEG dataset previously collected from three epilepsy patients undergoing presurgical monitoring. This dataset was chosen because it had already been analyzed using univariate methods, allowing for a direct comparison of the proposed MKL approach. The data was recorded during simple true/false judgments of memory sentences and mathematical equations. No new benchmarks were introduced; rather, the existing dataset was re-analyzed with a novel methodology.

Key Insights and Ideas

  • MKL as a powerful exploratory tool: The study demonstrates that MKL can effectively explore the rich, multi-dimensional nature of iEEG signals, going beyond the limitations of univariate analyses.
  • Automatic feature selection: The sparse MKL approach automatically identifies the most discriminative electrode-frequency band combinations, reducing the need for manual feature engineering and guiding hypothesis generation.
  • Distributed information: The findings suggest that information relevant to cognitive tasks is often distributed across multiple frequency bands and recording sites, rather than being confined to a single channel or frequency.
  • Complementary to univariate methods: MKL can highlight potential areas of interest that can then be further investigated with traditional univariate methods, creating a powerful synergy between exploratory and confirmatory analyses.
  • Beyond high-gamma: While high-gamma (HFB) is often the focus in iEEG research, this study shows that other frequency bands (like theta and alpha) can also carry significant discriminative information, especially when considered in combination.

Key Results

  • The MKL model achieved significant classification accuracy for distinguishing between "math" and "non-math" conditions across all subjects and frequency bands.
  • Balanced accuracy ranged from 77% to 89%, with the highest accuracy generally observed in the high-gamma band for two patients.
  • The MKL model's performance was comparable to or slightly better than the best single-frequency band performance, highlighting the benefit of combining information.
  • The sparsity constraint in MKL successfully identified a subset of kernels (electrode-frequency band combinations) that were crucial for discrimination. For instance, in Patient P1, only 5% of kernels were consistently selected across all folds.
  • The anatomical localization of these selected kernels often corresponded to brain regions previously implicated in numerical and memory processing by univariate studies (e.g., IPS and RSC), validating the MKL approach.

Limitations and Future Work

  • Sparsity control: The authors acknowledge that the specific MKL algorithm used enforces sparsity, which might lead to overlooking correlated signals from neighboring electrodes. They suggest that exploring MKL with different regularization techniques (like elastic-net) could address this.
  • Feature types: Currently, kernels were based on signal power. Future work could explore combining different feature types (e.g., raw signal, phase, coherence) within the MKL framework to gain even richer insights.
  • Temporal and frequency resolution: The paper mentions the possibility of investigating finer temporal resolutions or subject-specific frequency bands, which could offer more detailed information about the dynamics of brain activity.
  • Generalizability: While applied to iEEG, the authors suggest the MKL approach is generalizable to other electrophysiological data like scalp EEG or MEG.

Summary updated

Abstract

With a single computation, the contribution of each frequency band in each recording site in the estimated multivariate model can be highlighted, which then allows formulation of hypotheses that can be tested a posteriori with univariate methods if needed.
BibTeX
@article{schrouff2015decoding,
  author = {Jessica Schrouff and Janaı́na Mourão-Miranda and Christophe Phillips and Josef Parvizi},
  title = {{Decoding intracranial EEG data with multiple kernel learning method}},
  journal = {Journal of Neuroscience Methods},
  year = {2015},
  doi = {10.1016/j.jneumeth.2015.11.028},
}
Figure