Global signal regression strengthens association between resting-state functional connectivity and behavior

Jingwei Li, Ru Kong, Raphaël Liégeois, Csaba Orban, Yanrui Tan, Nanbo Sun, Avram J. Holmes, Mert R. Sabuncu, Tian Ge, B. T. Thomas Yeo

TL;DR

This study investigates the impact of Global Signal Regression (GSR), a controversial preprocessing step in fMRI, on the relationship between brain connectivity and behavior. The researchers found that GSR generally strengthens these associations, leading to better prediction of behavioral traits from brain activity patterns, even after other advanced denoising techniques.

What Problem Does This Paper Address?

Resting-state functional MRI (rs-fMRI) allows us to study brain connectivity by looking at how different brain regions' activity co-fluctuates when a person is at rest. This connectivity, known as Resting-State Functional Connectivity (RSFC), has been linked to various behavioral traits like cognition, personality, and emotion. However, rs-fMRI data is noisy, and a common preprocessing step, Global Signal Regression (GSR), aims to remove global noise sources like motion and respiration. The controversy arises because GSR also removes potentially valuable neural information and can introduce artificial negative correlations. While much research has focused on GSR's artifact removal capabilities, its impact on the RSFC-behavior link has been less clear, with mixed results in the literature. This paper aims to definitively answer whether GSR helps or hinders our ability to understand how brain connectivity relates to behavior.

Methods

The study employed two primary analytical approaches:

  1. Variance Component Model: This statistical model was used to quantify how much of the variation in behavioral traits could be explained by whole-brain RSFC. This approach helps understand the strength of the association between connectivity and behavior.
  2. Kernel Ridge Regression: This machine learning technique was used to predict individual behavioral scores based on RSFC patterns. This assesses how well RSFC can be used to predict behavior.

To ensure robustness, both methods were applied to the same datasets, and the results were compared. The researchers also performed jackknife resampling and cross-validation to estimate the uncertainty and reliability of their findings.

Datasets

The study utilized two large-scale, publicly available datasets of young, healthy adults:

  • Brain Genomics Superstruct Project (GSP): This dataset provided structural MRI, rs-fMRI, and 23 behavioral measures.
  • Human Connectome Project (HCP): This dataset included structural MRI, rs-fMRI, and 58 behavioral measures. Notably, the HCP data had already undergone denoising with ICA-FIX, allowing the researchers to test GSR's benefit on top of existing advanced denoising.

Key Insights and Ideas

  • GSR is a double-edged sword, but the "sharp edge" is beneficial for behavior: While GSR is known to remove noise, it also removes neural signals. The key question was whether the noise reduction outweighed the signal loss for behavioral prediction. The results strongly suggest it does.
  • Consistency across methods is reassuring: The fact that both a statistical model (variance component) and a predictive model (kernel ridge regression) showed similar improvements with GSR lends significant weight to the findings.
  • Task performance benefits more than self-report: This suggests that GSR might be particularly effective at enhancing the neural signals related to objective cognitive tasks, which are often more sensitive to subtle brain activity differences.
  • GSR's benefits persist even after advanced denoising: The fact that GSR improved results even on HCP data pre-processed with ICA-FIX is a crucial finding, suggesting GSR offers unique advantages beyond standard denoising.

Key Results

  • Increased Variance Explained: GSR increased the proportion of behavioral variance explained by RSFC by an average of 47% in the GSP dataset and 40% in the HCP dataset.
  • Improved Prediction Accuracy: GSR improved the accuracy of predicting behavioral traits from RSFC by an average of 64% in the GSP dataset and 12% in the HCP dataset.
  • Robustness: For a majority of behavioral measures (over 75%), GSR led to significant improvements in both explained variance and prediction accuracy.
  • Task Performance vs. Self-Report: Task-performance measures showed greater GSR-related improvements (average 7.9% increase in explained variance) compared to self-reported measures (4.4%).

Limitations and Future Work

  • Focus on Young Healthy Adults: The study's findings are limited to this demographic. Future work should investigate if GSR's benefits extend to other age groups, clinical populations (e.g., those with autism spectrum disorder or schizophrenia), or lifespan cohorts.
  • Whole-Brain RSFC: The analyses focused on whole-brain connectivity. It remains unclear whether GSR improves associations for specific brain networks or connections. Future research could explore GSR's impact at the network or voxel level.
  • Preprocessing Hyperparameters: The exact choice of preprocessing steps and their parameters can vary. Further exploration into how these choices affect RSFC-behavior associations is warranted.
  • Temporal ICA: The authors suggest investigating the use of temporal Independent Component Analysis (ICA) as an alternative to GSR for removing global artifacts while preserving neural information.

Summary updated

BibTeX
@article{li2019global,
  author = {Jingwei Li and Ru Kong and Raphaël Liégeois and Csaba Orban and Yanrui Tan and Nanbo Sun and Avram J. Holmes and Mert R. Sabuncu and Tian Ge and B. T. Thomas Yeo},
  title = {{Global signal regression strengthens association between resting-state functional connectivity and behavior}},
  journal = {NeuroImage},
  year = {2019},
  doi = {10.1016/j.neuroimage.2019.04.016},
}

Loading related content...

Figure