Understanding psychophysiological interaction and its relations to beta series correlation

Xin Di, Zhiguo Zhang, Bharat B. BiswalbioRxiv 2018 · DOI

TL;DR

This paper clarifies the relationship between two fMRI analysis techniques, Psychophysiological Interaction (PPI) and Beta Series Correlation (BSC), for studying task-modulated brain connectivity. It demonstrates that while both methods can detect condition-specific changes in connectivity, they capture different aspects, with PPI focusing on changes in effective connectivity (regression slopes) and BSC on functional connectivity (correlations).

What Problem Does This Paper Address?

Understanding how brain connectivity changes during specific tasks is crucial for deciphering brain function. Psychophysiological Interaction (PPI) and Beta Series Correlation (BSC) are two popular methods used to investigate these task-modulated connectivity changes in fMRI data. However, there have been persistent misunderstandings regarding what each method measures and how they relate to each other, particularly in event-related designs. This paper aims to resolve these ambiguities by providing a clear explanation of both methods, their underlying principles, and their empirical comparison.

Methods

The authors delve into the theoretical underpinnings of both PPI and BSC. For PPI, they explain how it models the interaction between a physiological time series (from a seed region) and a psychological variable (representing task conditions) within a General Linear Model (GLM). They detail the deconvolution process used to estimate these interactions at the "neuronal level" rather than the BOLD signal level. For BSC, they describe how it models each trial's activation separately to estimate trial-by-trial variability and then computes correlations between brain regions to infer functional connectivity. The paper also discusses variations of PPI, such as the "generalized" PPI, and compares different correlation measures for BSC.

Datasets

The study utilizes an fMRI dataset from an event-related stop signal task, previously reported in Di & Biswal (2019). This dataset includes data from 114 healthy subjects performing the task and a subset of 103 subjects with resting-state fMRI data.

Key Insights and Ideas

  • PPI measures effective connectivity differences, BSC measures functional connectivity: This is the core distinction. PPI quantifies how the relationship (regression slope) between two brain regions changes due to a task, while BSC quantifies how the correlation between their activity levels changes.
  • Deconvolution is crucial for neuronal-level PPI: The paper highlights the importance of deconvolution to move beyond BOLD signal artifacts and estimate interactions at the neuronal level, especially for event-related designs.
  • "Generalized" PPI offers a robust approach for multiple conditions: The authors explain how this method, by modeling conditions relative to all others, provides a more comprehensive way to assess condition-specific connectivity differences.
  • BSC is highly correlated with resting-state connectivity: This suggests that BSC primarily captures intrinsic functional relationships that are less influenced by task demands, which can be a limitation when studying task modulation.
  • Despite theoretical differences, PPI and BSC can reveal similar condition differences: When comparing the differences between task conditions (e.g., Stop vs. Go), both PPI and BSC can yield similar results, indicating their complementary nature in this context.

Key Results

  • When analyzing single conditions (e.g., Go or Stop), PPI and BSC matrices showed low correlations, reflecting their different measurement principles.
  • However, when comparing the differences between conditions (Stop vs. Go), the PPI and BSC matrices showed high correlations (r=0.73), indicating that both methods can effectively capture task-modulated connectivity changes.
  • The "generalized" PPI and contrast PPI approaches yielded identical results for condition differences, supporting the "generalized" PPI as a reliable method.
  • BSC results were highly similar to resting-state functional connectivity (correlations of 0.91 and 0.92), suggesting that BSC primarily reflects intrinsic connectivity.
  • Among different BSC measures, Pearson's correlation yielded more significant effects than Spearman's correlation or covariance, but still fewer than PPI.

Limitations and Future Work

  • Deconvolution challenges: The paper acknowledges that deconvolution is an ill-posed problem and that its success is critical for both PPI and BSC, especially in event-related designs. Future work could focus on improving deconvolution techniques, perhaps using more sophisticated filters or subject/region-specific HRFs.
  • Event-related design optimization: Optimizing event-related designs could also improve deconvolution and single-trial estimates, leading to better connectivity estimates for both methods.
  • Statistical sensitivity of BSC measures: While Pearson's correlation showed more significant effects than other BSC measures, it still yielded fewer significant results than PPI. Further research is needed to determine the optimal BSC measure and to understand why PPI might offer greater statistical sensitivity for detecting connectivity differences.
  • Statistical sensitivity claims: The paper refutes the claim that BSC is inherently more suitable for event-related data in terms of statistical sensitivity.

Summary updated

Abstract

Abstract Psychophysiological interaction (PPI) was proposed 20 years ago for study of task modulated connectivity on functional MRI (fMRI) data. A few modifications have since been made, but there remain misunderstandings on the method, as well as on its relations to a similar method named beta series correlation (BSC). Here, we explain what PPI measures and its relations to BSC. We first clarify that the interpretation of a regressor in a general linear model depends on not only itself but also on how other effects are modeled. In terms of PPI, it always reflects differences in connectivity between conditions, when the physiological variable is included as a covariate. Secondly, when there are multiple conditions, we explain how PPI models calculated from direct contrast between conditions could generate identical results as contrasting separate PPIs of each condition (a.k.a. “generalized” PPI). Thirdly, we explicit the deconvolution process that is used for PPI calculation, and how is it related to the trial-by-trial modeling for BSC, and illustrate the relations between PPI and those based upon BSC. In particular, when context sensitive changes in effective connectivity are present, they manifest as changes in correlations of observed trial-by-trial activations or functional connectivity. Therefore, BSC and PPI can detect similar connectivity differences. Lastly, we report empirical analyses using PPI and BSC on fMRI data of an event-related stop signal task to illustrate our points.
BibTeX
@misc{di2018understanding,
  author = {Xin Di and Zhiguo Zhang and Bharat B. Biswal},
  title = {{Understanding psychophysiological interaction and its relations to beta series correlation}},
  year = {2018},
  doi = {10.1101/322073},
  howpublished = {bioRxiv},
}

Loading related content...

Figure