Abstract
Psychophysiological interaction (PPI) is a regression-based method to study task-modulated brain connectivity. Despite its popularity in functional MRI (fMRI) studies, its reliability and reproducibility have not been extensively evaluated. We investigated these factors during a simple visual task and examined the effect of deconvolution on PPI results using a large open-access dataset (n = 138). Our analysis revealed that while group-level PPI effects are reproducible, individual-level reliability is surprisingly low. Furthermore, we found that the deconvolution step may be unnecessary for block-designed data, as both methods yielded similar results, with the non-deconvolution approach often showing larger effect sizes.
Introduction
In functional MRI (fMRI), researchers aim to understand not just which brain regions are active, but how they communicate. Psychophysiological Interaction (PPI) is a widely used method to model how task demands (the "psychological" factor) change the connectivity between brain regions (the "physiological" factor). However, a technical hurdle exists: the BOLD signal measured by fMRI is a slow, indirect measure of neural activity. To account for this, many researchers use "deconvolution" to estimate the underlying neural signal before calculating the interaction. This study evaluates whether this complex step is truly necessary and how reliable PPI is across different scanning sessions.
The Deconvolution Dilemma
The standard PPI approach attempts to recover the neuronal signal through deconvolution before computing the interaction term. However, this process is mathematically complex and sensitive to noise. The authors compared this to a simpler "non-deconvolution" method where the task design is convolved with the hemodynamic response function (HRF) and multiplied directly by the BOLD signal.
For block designs, the convolved task signal is so slow that it remains highly correlated with the raw signal. As the design cycle increases, the correlation between the PPI terms generated by both methods approaches 1.0, suggesting that deconvolution provides little additional information for these types of experiments.
Methods and Datasets
The study utilized the Enhanced Nathan Kline Institute—Rockland Sample (NKI-RS), analyzing data from 138 subjects performing a simple visual checkerboard task (20s "on", 20s "off"). Each subject was scanned twice with different repetition times (TR): a "fast" scan at 645 ms and a "standard" scan at 1,400 ms.
To assess the results, the researchers used:
- Dice Coefficients: To measure the overlap of significant brain maps between different scanning runs (reproducibility).
- Intraclass Correlation (ICC): To measure the consistency of effects for individual subjects across runs (reliability).
Key Results
Reproducibility vs. Reliability
At the group level, the results were highly reproducible; the same brain regions showed up in both runs. However, at the individual level, the reliability was strikingly low. While standard brain activation showed ICCs between 0.2 and 0.7, PPI effects hovered near zero (~0.07). This suggests that while group averages are stable, we should be cautious when using PPI to compare individual differences.
Methodological Comparison
Both the deconvolution and non-deconvolution methods produced very similar group-level maps. In fact, the non-deconvolution method often showed larger effect sizes and better overlap between runs. The study also identified the posterior fusiform gyrus (FG1) as a major "hub" that showed reduced connectivity with many other visual areas during the task, suggesting a high degree of functional segregation during simple visual stimulation.
Conclusion for the Practitioner
If you are conducting a standard block-design fMRI study, the evidence suggests you can likely skip the deconvolution step in your PPI analysis without losing significant information. However, researchers must remain wary of using PPI for individual-subject correlations or clinical predictions unless they have very long scan times and high-quality data to overcome the low individual-level reliability of the signal.
Summary updated