The Multilevel Event-related Deconvolved Signal Analysis (MEDuSA) of trial-to-trial variation in neural task-based activity
Thursday, May 6, 2021 – 4:00 – 5:00 pm
Dr. Michael Hallquist, Associate Professor of Clinical Psychology and Quantitative Psychology, the Developmental Personality Neuroscience (DEPENd) Lab; Department of Psychology and Neuroscience; University of North Carolina at Chapel Hill
The integration of task-based fMRI data with signals from computational models of behavior often depends on a variant of voxelwise general linear model (GLM) analysis. A major challenge limitation of this model-based fMRI approach is that trial-to-trial variation in neural activity is typically ignored, limiting our understanding of brain-behavior relationships that unfold dynamically with learning or repetition. This workshop will present a novel multilevel analysis approach that preserves between-trial neural variability and within-trial dynamics to link task-based fMRI data and trial-varying signals from computational models. I will provide a more general overview of how multilevel models can be applied to brain and behavioral data using signals that vary both within and between individuals.
The Multilevel Event-related Deconvolved Signal Analysis (MEDuSA) framework provides a useful tool for interrogating where (spatially) and when (temporally, and with learning) cognitive processes instantiated by computational models are represented by brain regions measured by fMRI. Conventional model-based fMRI studies of psychopathology often distill individual differences to a single index per person (e.g., sensitivity to reward prediction errors in the striatum). By comparison, the approach presented here allows for more detailed examination of how dynamic cognitive processes such as learning relate to dynamic changes in neural activity as measured by fMRI.
In standard model-based GLMs, signals from computational models are treated as 'parametric modulators' -- dimensionally varying signals that are convolved with a hemodynamic response function (HRF) and correlated with brain activity. This approach assumes that 1) the HRF reasonably represents model-related neural activity; 2) that the value of the parametric modulator and the duration of the event with which it aligns are intertwined; and 3) that a single estimate of signal-related activity (i.e., the estimated regression coefficient or 'beta') summarizes the average within- and between-trial evoked response for each person. The MEDuSA framework presented here provides an analytic approach that overcomes these assumptions and gives researchers access to new information about neural variability.
This approach is still being validated for broader use but was positively received in our initial application to describing learning signals along the long axis of the hippocampus. More generally, given the explosion of interest in computational psychiatry approaches that include fMRI, colleagues and funders are increasingly interested in analyses that go beyond associations between computational model signals and task-related brain activity.
The methods presented here are largely implemented in R, a free open-source program for data analysis (https://cran.r-project.org). Preprocessing of fMRI data is not a focus of this workshop, but we rely on open-source software including FSL and AFNI for this purpose.
Prerequisites: Proficiency with R, particularly for regression models and visualization.
Background knowledge: Voxelwise general linear model analyses (e.g. http://www.fmri4newbies.com/tutorial-2), fundamentals of multilevel modeling (random intercept and slope regression, cross-level interactions; e.g. http://www.bristol.ac.uk/cmm/learning/multilevel-models/), fundamentals of computational reinforcement learning models (e.g. https://cbmm.mit.edu/learning-hub/tutorials/computational-tutorial/reinforcement-learning).
New skills and knowledge required: Basics of deconvolution approaches to fMRI data; Multivariate extensions of multilevel models; Corrections for multiple comparisons under arbitrary dependence among test statistics.
Tutorials online: This is an in-progress series, but we are currently hosting a series of meetings examining the application of multilevel models to experimental data. These provide a useful foundation for MLMs in the experimental context and the materials here will be largely complete by the time of this workshop: https://uncdependlab.github.io/MLM_Tutorial/
Source code: All code related to this approach is freely available here: https://github.com/UNCDEPENdLab/fmri.pipeline
- Bush, K., & Cisler, J. (2013). Decoding neural events from fMRI BOLD signal: A comparison of existing approaches and development of a new algorithm. Magnetic Resonance Imaging, 31(6), 976–989. Doi: https://doi.org/10.1016/j.mri.2013.03.015
- Dombrovski, A. Y., Luna, B., & Hallquist, M. N. (2020). Differential reinforcement encoding along the hippocampal long axis helps resolve the explore–exploit dilemma. Nature Communications, 11(1), 5407. Doi: https://doi.org/10.1038/s41467-020-18864-0
- Hallquist, M. N., & Dombrovski, A. (2020). Reinforcement Learning Approaches to Computational Clinical Neuroscience. In A. G. C. Wright & M. N. Hallquist (Eds.), Handbook of Research Methods in Clinical Psychology (pp. 168–189).
- Cambridge University Press. https://www.cambridge.org/core/books/cambridge-handbook-of-research-methods-in-clinical-psychology/clinical-computational-neuroscience/4027551F73028B522F4E88682F08999A
- Hoffman, L., & Rovine, M. J. (2007). Multilevel models for the experimental psychologist: Foundations and illustrative examples. Behavior Research Methods, 39(1), 101–117. https://doi.org/10.3758/BF03192848 Doi:https://doi.org/10.3758/BF03192848
Datasets: The MEDuSA approach can be used with most task-related fMRI datasets, but will be of greatest value to paradigms where 1) there is a topographical gradient or structure to the regions being analyzed; 2) there is specific interest in between-trial variability in neural activity; and/or 3) there is specific interest in decoding when in time (within a trial) certain computations occur. Moreover, datasets with longer trials and/or longer inter-trial intervals may be especially amenable to this approach.