A new transfer entropy method for measuring directed connectivity from complex-valued fMRI data

Wei-Xing Li, Qiu-Hua Lin, Chaoying Zhang, Yue Han, Vince D. CalhounFrontiers in Neuroscience 2024 · DOI

TL;DR

Researchers have developed a novel Complex-valued Transfer Entropy (CTE) method that utilizes both magnitude and phase data from fMRI scans to map directed connectivity in the brain. By accounting for previously ignored phase information, the method achieves a staggering 95.5% accuracy in classifying schizophrenia patients, significantly outperforming traditional magnitude-only approaches.

What Problem Does This Paper Address?

Functional Magnetic Resonance Imaging (fMRI) is a cornerstone of modern neuroscience, but most researchers only use half of the data it produces. When an fMRI scanner captures an image, it generates "complex-valued" data consisting of both magnitude (the strength of the signal) and phase (the timing or angle of the signal).

Traditionally, the phase data is discarded as "noise," and analysis focuses solely on magnitude. However, this ignores potentially vital information about how brain regions communicate. Furthermore, most existing tools for measuring "directed connectivity"—the causal "who is talking to whom" in the brain—rely on linear models (like Granger Causality) that struggle with the brain's inherently non-linear nature. This paper seeks to bridge this gap by creating a tool that is both complex-valued and non-linear.

Methods

The authors propose Complex-valued Transfer Entropy (CTE). Transfer entropy is a "model-free" approach, meaning it doesn't assume a linear relationship between brain regions; instead, it measures how much information the history of one region provides about the future of another.

The core innovation of CTE lies in how it handles the four-way interaction between two brain regions ($z_1$ and $z_2$):

  1. Magnitude $\to$ Magnitude
  2. Phase $\to$ Phase
  3. Magnitude $\to$ Phase (using Partial Transfer Entropy)
  4. Phase $\to$ Magnitude (using Partial Transfer Entropy)

By using Partial Transfer Entropy (PTE), the algorithm can measure the "cross-talk" between magnitude and phase while controlling for the correlations between them. To make the system robust against the high noise levels typical of fMRI, the authors implemented symbolic processing, which converts continuous signals into discrete symbols, and a shuffling strategy to perform statistical significance tests, ensuring that detected connections aren't just random coincidences.

Datasets

The study validated the CTE method using two distinct approaches:

  • Simulated Signals: The authors generated six types of synthetic signals (Linear L1-L3 and Non-linear N1-N3) with known causal directions and varying levels of Gaussian noise to test the algorithm's accuracy and robustness.
  • Experimental fMRI Data: A resting-state dataset from 80 subjects (40 Healthy Controls and 40 patients with Schizophrenia) collected at the University of New Mexico. The brain was divided into 116 regions using the Anatomical Automatic Labeling (AAL) atlas.

Key Insights and Ideas

  • Phase is Informative, Not Just Noise: The study proves that phase data contains unique causal information. When phase was included, the ability to distinguish between healthy brains and those with schizophrenia improved dramatically.
  • The Power of "Partial" Analysis: Simply adding magnitude results to phase results leads to errors because the two are correlated. Using Partial Transfer Entropy allows the model to isolate the complementary effects of phase on magnitude (and vice versa), which the authors call the "secret sauce" for their high accuracy.
  • Non-linearity is Non-negotiable: The results showed that traditional Granger Causality failed almost entirely (dropping to ~9% accuracy) when faced with non-linear simulated signals, whereas CTE remained highly accurate (>95%).

Key Results

  • Superior Accuracy in Simulation: For non-linear signals with complete complex causality, CTE achieved 18.7% to 85.9% higher accuracy than traditional real-valued algorithms.
  • Schizophrenia Classification: Using directed connectivity as features for a Support Vector Machine (SVM), CTE achieved a classification accuracy of 95.5%. This is significantly higher than:
  • sCTE (Simplified CTE, no cross-causality): 85.3%
    • STE (Symbolic Transfer Entropy): 81.9%
    • Granger Causality: 74.6%
  • Biomarker Identification: The method identified critical "hubs" of abnormal connectivity in schizophrenia, specifically within the Cerebellum (CER), the Right Frontal Parietal (RFP) regions, and the Temporal Lobe (TEM). These findings align with clinical symptoms like hallucinations and cognitive impairment.

Limitations and Future Work

While the results are promising, the authors note a few areas for growth:

  • Computational Complexity: Transfer entropy calculations, especially with complex-valued data and shuffling tests, are more computationally expensive than linear methods.
  • Dynamic Connectivity: This study looked at "static" connectivity over the whole scan. The authors suggest that combining CTE with a sliding window approach could reveal how these causal links change second-by-second.
  • Broader Application: The current study focused on schizophrenia. Future work could apply CTE to other conditions, such as depression or Alzheimer’s, to see if unique complex-valued "causal fingerprints" exist for those disorders as well.

Summary updated

Abstract

The proposed CTE provides a new general method for fully detecting highly predictive directed connectivity from complex-valued fMRI data, with magnitude-only fMRI data as a specific case.
BibTeX
@inproceedings{li2024new,
  author = {Wei-Xing Li and Qiu-Hua Lin and Chaoying Zhang and Yue Han and Vince D. Calhoun},
  title = {{A new transfer entropy method for measuring directed connectivity from complex-valued fMRI data}},
  booktitle = {Frontiers in Neuroscience},
  year = {2024},
  doi = {10.3389/fnins.2024.1423014},
}
Figure