A Complementary Method of PCC for the Construction of Scalp Resting-State EEG Connectome: Maximum Information Coefficient

李沛洋, Yin Tian, Huiling Zhang, Peiyang Li, Yang LiIEEE Access 2019 · DOI

Introduction

Understanding the brain's functional architecture requires mapping the "conversations" between its various regions. These maps, known as connectomes, are typically constructed as mathematical graphs where brain regions serve as nodes and the functional relationships between them are represented as edges. Traditionally, the Pearson Correlation Coefficient (PCC) has been the gold standard for defining these edges. However, as research into the brain's complex, non-linear dynamics deepens, the limitations of PCC have become increasingly apparent.

The Limitations of Linear Correlation

While PCC is computationally efficient and effective for identifying linear associations, it suffers from two significant drawbacks in the context of neuroimaging:

  1. Linear Bias: PCC only measures how well two variables fit a straight line. In a system as complex as the human brain, interactions are often non-linear (e.g., sinusoidal or quadratic). In such cases, PCC may fail to detect a relationship entirely, reporting a correlation of zero.
  2. Outlier Sensitivity: EEG data is inherently noisy. PCC is highly sensitive to outliers, which can skew results and lead to inaccurate network representations.

Because the brain is a quintessentially non-linear system, relying solely on PCC likely results in "missing links" in our functional maps of brain activity.

Methodology: Introducing the Maximum Information Coefficient (MIC)

To address these gaps, researchers have proposed the use of the Maximum Information Coefficient (MIC). Rooted in mutual information theory, MIC works by gridding the scatterplot of two variables in various ways and calculating the maximum possible mutual information. This approach allows it to detect any functional relationship—linear or non-linear—with similar strength, a property known as "equitability."

The workflow for constructing and analyzing brain networks using EEG data, comparing MIC and PCC methods.

The study validated this approach through a two-pronged strategy:

  • Simulation: Three-node mathematical models (Linear, Mixed, and Nonlinear) were used to test if MIC could identify connections that PCC missed.
  • Real-world EEG: Resting-state EEG was recorded from 18 healthy subjects, focusing on the alpha rhythm (8–13 Hz), which is dominant during quiet wakefulness. Connectivity matrices were calculated using 21 electrode locations as nodes.

Results and Insights

The study's findings highlight the strengths of MIC as a "wide-angle lens" for brain connectivity:

  • Capturing Non-linear Dynamics: In the non-linear simulation model, PCC failed to find significant correlations (p > 0.05), leaving nodes disconnected. MIC successfully captured these relationships with high significance (p < 0.05).
  • Network Robustness: Brain networks are typically "thresholded" to filter out weak or noisy connections. MIC-based networks proved more robust at lower thresholds, maintaining higher global efficiency and node density compared to PCC-based networks.
  • Small-World Topology: Both methods confirmed that the human brain operates as a "small-world" network, characterized by high local clustering and short paths between nodes, facilitating efficient information transfer.

Conclusion and Future Directions

The authors emphasize that MIC is not intended to replace PCC but rather to complement it. While PCC is effective for linear associations, MIC acts as a vital safeguard, capturing the complex, "hidden" dynamics that linear measures ignore.

Future work will focus on overcoming the computational intensity of MIC to make it viable for high-density EEG arrays and exploring its utility across other frequency bands like theta, beta, and gamma. Most importantly, applying MIC to clinical populations, such as those with Alzheimer’s or Mild Cognitive Impairment (MCI), may reveal non-linear connectivity markers that serve as earlier diagnostic tools than traditional methods.

Summary updated

BibTeX
@article{2019complementary,
  author = {李沛洋 and Yin Tian and Huiling Zhang and Peiyang Li and Yang Li},
  title = {{A Complementary Method of PCC for the Construction of Scalp Resting-State EEG Connectome: Maximum Information Coefficient}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2897908},
}
Figure