Group-level component analyses of EEG: validation and evaluation

René J. Huster, Sergey Plis, Vince D. CalhounFrontiers in Neuroscience 2015 · DOI

Abstract

Multi-subject or group-level component analysis provides a data-driven approach to study properties of brain networks. Algorithms for group-level data decomposition of functional magnetic resonance imaging data have been brought forward more than a decade ago and have significantly matured since. Similar applications for electroencephalographic data are at a comparatively early stage of development though, and their sensitivity to topographic variability of the electroencephalogram or loose time-locking of neuronal responses has not yet been assessed. This study investigates the performance of independent component analysis (ICA) and second order blind source identification (SOBI) for data decomposition, and their combination with either temporal or spatial concatenation of data sets, for multi-subject analyses of electroencephalographic data. Analyses of simulated sources with different spatial, frequency, and time-locking profiles, revealed that temporal concatenation of data sets with either ICA or SOBI served well to reconstruct sources with both strict and loose time-locking, whereas performance decreased in the presence of topographical variability. The opposite pattern was found with a spatial concatenation of subject-specific data sets. This study proofs that procedures for group-level decomposition of electroencephalographic data can be considered valid and promising approaches to infer the latent structure of multi-subject data sets. Yet, specific implementations need further adaptations to optimally address sources of inter-subject and inter-trial variance commonly found in EEG recordings.

Introduction

In electroencephalography (EEG) research, identifying common neural sources across a group of subjects is a fundamental challenge. Unlike fMRI, where group-level analysis is well-established, EEG methods must contend with significant inter-subject variability. Two primary hurdles complicate this process: temporal jitter (neural responses that are not perfectly phase-locked to a stimulus) and topographic variability (differences in how neural sources project to the scalp due to unique brain anatomy and electrode placement).

This study, published in Frontiers in Neuroscience, evaluates how different algorithmic approaches and data concatenation strategies handle these "messy" real-world factors to provide a robust framework for group-level EEG analysis.

Methodology

The researchers evaluated four primary combinations of techniques based on two algorithms and two data-stacking strategies:

  • Algorithms: ICA (Infomax), which maximizes statistical independence, and SOBI (Second Order Blind Identification), which utilizes time-lagged correlations.
  • Concatenation Strategies:
    • Temporal (tc): Data from all subjects are lined up side-by-side. This approach assumes that all subjects share identical scalp maps.
  • Multilevel (ml): A two-step process involving an initial PCA reduction for each subject followed by vertical stacking for group analysis. This allows for subject-specific variation in topographies.

To establish a "ground truth," the study utilized simulations of 15 virtual subjects. Three distinct neural sources were modeled—Alpha (10 Hz), Beta (20 Hz), and Gamma (40 Hz)—and embedded in realistic noise profiles derived from actual human resting-state EEG.

Illustration of simulated neural sources and the distinction between time-locked and non-time-locked (jittered) activity across trials.

Key Results and Discussion

The evaluation revealed a critical trade-off between the two concatenation strategies, dictated by the nature of the neural activity and the anatomical consistency of the group.

The Impact of Temporal Jitter

Temporal concatenation methods (tcICA/tcSOBI) proved significantly more effective at capturing "jittery" or non-phase-locked neural responses. In conditions with high temporal jitter (200ms), these methods significantly outperformed multilevel versions (p < 0.001). The multilevel approach relies on a group-level PCA that acts similarly to an average; if signals are not perfectly aligned in time, this step can inadvertently wash out the induced activity.

Handling Topographic Variability

Conversely, multilevel methods (mlICA/mlSOBI) were far superior when dealing with anatomical differences. By allowing for subject-specific scalp maps in the first stage of decomposition, they preserved individual anatomical "fingerprints." When topographies varied between subjects, multilevel methods explained up to 60% more variance for high-frequency sources compared to temporal concatenation, which assumes a rigid, shared mapping.

Frequency Sensitivity

High-frequency signals, such as Gamma (40 Hz), were found to be much more sensitive to jitter than lower frequencies. A 50ms timing difference is a minor phase shift for a 10Hz wave but represents two full cycles for a 40Hz wave, leading to massive signal cancellation in multilevel models that lack precise alignment.

Conclusion

This study demonstrates that group-level decomposition is a valid and promising approach for multi-subject EEG analysis. However, researchers must choose their strategy based on their specific goals: temporal concatenation is ideal for studying jittery, induced responses, while multilevel approaches are necessary to account for significant anatomical and topographic variability. Future developments may include jitter-correction techniques and improved model-order estimation to further enhance the sensitivity of these tools.

Summary updated

Abstract

Multi-subject or group-level component analysis provides a data-driven approach to study properties of brain networks. Algorithms for group-level data decomposition of functional magnetic resonance imaging data have been brought forward more than a decade ago and have significantly matured since. Similar applications for electroencephalographic data are at a comparatively early stage of development though, and their sensitivity to topographic variability of the electroencephalogram or loose time-locking of neuronal responses has not yet been assessed. This study investigates the performance of independent component analysis (ICA) and second order blind source identification (SOBI) for data decomposition, and their combination with either temporal or spatial concatenation of data sets, for multi-subject analyses of electroencephalographic data. Analyses of simulated sources with different spatial, frequency, and time-locking profiles, revealed that temporal concatenation of data sets with either ICA or SOBI served well to reconstruct sources with both strict and loose time-locking, whereas performance decreased in the presence of topographical variability. The opposite pattern was found with a spatial concatenation of subject-specific data sets. This study proofs that procedures for group-level decomposition of electroencephalographic data can be considered valid and promising approaches to infer the latent structure of multi-subject data sets. Yet, specific implementations need further adaptations to optimally address sources of inter-subject and inter-trial variance commonly found in EEG recordings.
BibTeX
@inproceedings{huster2015group,
  author = {René J. Huster and Sergey Plis and Vince D. Calhoun},
  title = {{Group-level component analyses of EEG: validation and evaluation}},
  booktitle = {Frontiers in Neuroscience},
  year = {2015},
  doi = {10.3389/fnins.2015.00254},
}
Figure