Estimating the Number of Sources in Magnetoencephalography Using Spiked Population Eigenvalues

Zhigang Yao, Ye Zhang, Zhidong Bai, William F. EddyJournal of the American Statistical Association 2017 · DOI

TL;DR

This paper introduces a novel method for accurately estimating the number of underlying neural sources in Magnetoencephalography (MEG) data, even in the presence of significant noise. By focusing on "spiked population eigenvalues" of a transformed data matrix, their approach offers a more robust solution than traditional methods, particularly in low signal-to-noise ratio scenarios.

What Problem Does This Paper Address?

Magnetoencephalography (MEG) is a powerful tool for understanding brain activity by measuring magnetic fields produced by neural currents. However, a key challenge in analyzing MEG data is determining the number of underlying neural sources (e.g., groups of active neurons). Traditional methods like Principal Component Analysis (PCA) and information criteria rely on the distribution of eigenvalues from the sample covariance matrix. The problem is that these sample eigenvalues are highly sensitive to noise, especially when the signal-to-noise ratio (SNR) is low, leading to inaccurate estimates of the number of sources. This paper aims to overcome this limitation by developing a method that estimates the number of sources based on the population eigenvalues, which are more robust to noise.

Methods

The core of the paper's contribution lies in a novel framework for estimating the "intrinsic dimensionality" (ID) of the MEG data. This ID is defined as the number of "spiked population eigenvalues" of a transformed data matrix. The key steps involve:

  • Optimal SNR Transformation: They introduce a transformation that maximizes a signal-to-noise ratio (SNR) rescaling functional. This transformation aims to enhance the signal while suppressing noise.
  • Spiked Covariance Model: The problem is reframed using a "spiked covariance model," where the population covariance matrix is assumed to have a few large eigenvalues (spikes) and the rest are bulk eigenvalues.
  • Estimating Population Eigenvalues: Crucially, instead of relying on noisy sample eigenvalues, they develop methods to estimate the true population eigenvalues. This involves algorithms (Algorithm 1, 2, and 3) that iteratively refine thresholds to distinguish between spiked and bulk eigenvalues.
  • Intrinsic Dimensionality (ID): The number of estimated spiked population eigenvalues directly corresponds to the intrinsic dimensionality, which is then interpreted as the number of neural sources.

Datasets

  • Simulated Data: The authors generated synthetic MEG data with varying SNRs to test their method's performance against established techniques. This simulated data allowed them to control the ground truth (number of dipoles) and SNR.
  • Empty Room MEG Data: A dataset from an empty MEG room was used to test the method's ability to detect known and potentially unknown sources in a real-world scenario. This data was collected from a 306-channel system.
  • Brain-Controlled Interfaces (BCI) Data: A real MEG dataset from a BCI experiment was analyzed to investigate the dynamics of source activity over time. This data involved 102 magnetometers.

Key Insights and Ideas

  • Population vs. Sample Eigenvalues: The most significant insight is that relying on sample eigenvalues for source estimation is inherently flawed in noisy MEG data. Estimating population eigenvalues provides a much more stable and accurate basis for determining the number of sources.
  • Optimal SNR Transformation is Key: The proposed transformation is not just a preprocessing step; it's fundamental to revealing the underlying structure of the data by effectively separating signal from noise.
  • Spiked Eigenvalues as Source Indicators: The number of "spiked" population eigenvalues directly correlates with the number of active neural sources, offering a principled way to determine dimensionality.
  • Robustness to Low SNR: The method demonstrates remarkable resilience in identifying the correct number of sources even when the SNR is extremely low, a critical advantage for real-world MEG data.

Key Results

  • Simulation Studies: In simulations with varying SNRs, their Spiked Population Eigenvalue (SPE) method consistently recovered the correct number of dipoles (four in their primary simulation), outperforming PCA, AIC, MDL, and EIF, especially at low SNRs. For instance, at an SNR of 0.001, SPE correctly identified four dipoles, while other methods struggled or overestimated.
  • Empty Room Data: Analysis of the empty room data revealed the presence of a known 60 Hz source and suggested the existence of one or two additional sources below 50 Hz, demonstrating the method's ability to detect unexpected signals.
  • BCI Data: The BCI data analysis showed dynamic changes in the estimated number of sources during different task periods (imagined vs. overt movement), highlighting the method's utility in tracking neural activity over time.

Limitations and Future Work

  • Noise Estimation Accuracy: The optimal SNR transformation's effectiveness relies on accurate noise estimation. The paper acknowledges that noise structure in neurological experiments, particularly MEG, can be very complex, and further investigation into noise estimation methods is needed.
  • Dynamic Estimation: While the paper shows stable results by estimating sources sequentially, the authors suggest that a truly dynamic implementation of the estimation is preferable, especially for data where the number of sources changes over time (like in BCI data).
  • Criterion Selection: The paper notes that while the number of spiked population eigenvalues is a useful guide, the issue of assessing which criterion is most informative for deciding on the intrinsic dimensionality deserves further scrutiny.
  • Further Investigation of Noise Influence: The authors point to future work in investigating the impact of noise estimation on the estimated spiked eigenvalues.

Summary updated

Abstract

Magnetoencephalography (MEG) is an advanced imaging technique used to measure the magnetic fields outside the human head produced by the electrical activity inside the brain. Various source localization methods in MEG require the knowledge of the underlying active sources, which are identified by a priori. Common methods used to estimate the number of sources include principal component analysis or information criterion methods, both of which make use of the eigenvalue distribution of the data, thus avoiding solving the time-consuming inverse problem. Unfortunately, all these methods are very sensitive to the signal-to-noise ratio (SNR), as examining the sample extreme eigenvalues does not necessarily reflect the perturbation of the population ones. To uncover the unknown sources from the very noisy MEG data, we introduce a framework, referred to as the intrinsic dimensionality (ID) of the optimal transformation for the SNR rescaling functional. It is defined as the number of the spiked population eigenvalues of the associated transformed data matrix. It is shown that the ID yields a more reasonable estimate for the number of sources than its sample counterparts, especially when the SNR is small. By means of examples, we illustrate that the new method is able to capture the number of signal sources in MEG that can escape PCA or other information criterion-based methods. Supplementary materials for this article are available online.
BibTeX
@article{yao2017estimating,
  author = {Zhigang Yao and Ye Zhang and Zhidong Bai and William F. Eddy},
  title = {{Estimating the Number of Sources in Magnetoencephalography Using Spiked Population Eigenvalues}},
  journal = {Journal of the American Statistical Association},
  year = {2017},
  doi = {10.1080/01621459.2017.1341411},
}
Figure