Maximally reliable spatial filtering of steady state visual evoked potentials

Jacek P. Dmochowski, Alex S. Greaves, Anthony M. NorciaNeuroImage 2015 · DOI

TL;DR

This paper introduces "Reliable Components Analysis" (RCA), a novel spatial filtering technique for steady-state visual evoked potentials (SSVEPs). RCA maximizes trial-to-trial reliability, leading to more physiologically plausible components and higher signal-to-noise ratios (SNRs) compared to existing methods like PCA and CSP, especially with limited data.

What Problem Does This Paper Address?

Analyzing brain activity using electroencephalography (EEG) often involves dealing with noisy signals and the inherent spatial smearing caused by volume conduction. Steady-state visual evoked potentials (SSVEPs), generated by presenting a flickering visual stimulus, are a popular tool for studying visual processing. Traditional SSVEP analysis often focuses on individual electrodes or simple linear combinations, which can limit spatial resolution and SNR. While spatial filtering techniques exist to combine electrode signals, they often prioritize maximizing SNR, sometimes at the expense of physiological interpretability. This can lead to components whose spatial patterns don't accurately reflect underlying neural sources, making them difficult to interpret. This paper aims to address this by developing a spatial filtering method that explicitly optimizes for the reliability of the SSVEP signal across trials, leading to components that are both physiologically plausible and have high SNRs.

Methods

The core of the paper is the introduction of "Reliable Components Analysis" (RCA). RCA operates on the fundamental assumption that evoked responses are reproducible across trials. It achieves this by:

  • Trial Aggregation: Instead of stacking trials into a single large matrix, RCA explicitly considers the trial dimension. It forms trial-aggregated data matrices by pairing Fourier coefficients from different trials.
  • Maximizing Trial-to-Trial Covariance: RCA projects the data into a low-dimensional space where the spectral covariance between trials is maximized. This is achieved by solving a generalized eigenvalue problem that optimizes the ratio of across-trial covariance to within-trial covariance.
  • Spatial Filtering: The output of RCA is a set of spatial filters (components) that capture the most reliable neural activity evoked by the stimulus.

The authors compare RCA to two established methods:

  • Principal Component Analysis (PCA): Aims to maximize the proportion of variance explained in the data.
  • Common Spatial Patterns (CSP): Aims to maximize the SNR by maximizing the difference between signal and noise covariance matrices.

Datasets

The paper utilizes two types of datasets for evaluation:

  • Synthetic Data Set: A simulated dataset created using a boundary element model (BEM) of the human head. This simulation allowed for ground-truth measurements of SNR and direct comparison of recovered components with known lead fields of simulated neural sources. The simulation was used to assess the performance of RCA, PCA, and CSP under varying numbers of trials.
  • Real Data Set: EEG data collected from 22 human subjects performing a visual stimulation task. This dataset was used to evaluate the physiological plausibility and SNR of the components recovered by the different methods, particularly by examining contralateralization of scalp topographies with stimulated hemifields and across different stimulus contrasts.

Key Insights and Ideas

  • Reliability as a Primary Metric: The paper strongly advocates for using trial-to-trial reliability as a key metric for extracting meaningful neural signals, especially in cognitive neuroscience where physiological interpretability is crucial.
  • RCA's Physiological Plausibility: RCA consistently recovers components whose scalp topographies closely resemble the expected lead fields of underlying neural sources, unlike some other methods that can produce less interpretable patterns.
  • Robustness to Input SNR: The real-data results suggest RCA is more robust to variations in stimulus contrast (and thus input SNR) compared to PCA, maintaining physiologically plausible components even at low contrasts.
  • Dimensionality Reduction: RCA achieves significant dimensionality reduction, capturing the bulk of trial-to-trial reliability in a small number of components (e.g., 4 components capturing >95% reliability), far fewer than the original number of channels.
  • Trade-off Between SNR and Plausibility: The paper highlights the trade-off between maximizing SNR (where CSP excels) and achieving physiological interpretability (where RCA shows promise).

Key Results

  • Synthetic Data:
    • RCA recovered both the "reliable" and "variable" simulated sources cleanly at 50 trials, while CSP required more trials and PCA's components were often mixtures.
  • RCA showed comparable or lower angular error (indicating better topographic match to ground truth) than PCA and CSP for fewer trials.
  • At low trial counts (≤ 30), RCA yielded the highest SNR. At higher trial counts (≥ 50), CSP achieved the highest SNR, but RCA still outperformed PCA.
  • Real Data:
    • RCA components exhibited clear contralateralization of scalp topographies with stimulated hemifields even at low stimulus contrasts, suggesting physiological relevance.
  • PCA components showed increasing physiological plausibility with contrast but had lower SNRs than the best individual channel.
  • CSP yielded high SNRs but often sacrificed physiological plausibility, especially at low contrasts.
  • RCA provided a median SNR improvement over the best individual channel, particularly at high contrasts, and significantly outperformed PCA in SNR.

Limitations and Future Work

  • Simulation Simplification: The authors acknowledge that the simulation is a simplification of the actual neural environment, using dipolar sources, average conductivity, and uncorrelated noise. However, they argue that the relative performance of methods in simulation is likely to hold in real settings.
  • Learning on Aggregated vs. Individual Data: The paper discusses the choice between learning spatial filters on subject-aggregated data (lower estimation noise) versus individual data (capturing individual differences). For comparing SNR, individual learning was used, which means projected spectra are not directly comparable across subjects.
  • Future Work:
    • The paper suggests that RCA's physiologically plausible scalp projections could be directly used as inputs for source localization algorithms, potentially eliminating the need to choose specific time instants.
  • RCA could be applied to time-domain data for transient evoked potentials, though this would likely require more trials due to the higher dimensionality of time series.
  • The authors propose RCA as a valuable front-end feature selection step for Brain-Computer Interfaces (BCIs), reducing dimensionality while retaining reproducible neural features.

Summary updated

BibTeX
@article{dmochowski2015maximally,
  author = {Jacek P. Dmochowski and Alex S. Greaves and Anthony M. Norcia},
  title = {{Maximally reliable spatial filtering of steady state visual evoked potentials}},
  journal = {NeuroImage},
  year = {2015},
  doi = {10.1016/j.neuroimage.2014.12.078},
}
Figure