CCEP Inpainting: A Computer Vision-Based Method to Predict Cortico-Cortical Evoked Potentials
Poster No:
1521
Submission Type:
Abstract Submission
Authors:
Rachel June Smith1, Arie Nakhmani1, Kyle Evans-Lee2, Helen Brinyark1, Joshua LaRocque3, Erin Conrad3
Institutions:
1University of Alabama at Birmingham, Birmingham, AL, 2Systems and Technology Research, Woburn, MA, 3University of Pennsylvania, Philadelphia, PA
First Author:
Co-Author(s):
Introduction:
Single-pulse electrical stimulation (SPES) elicits cortico-cortical evoked potentials (CCEPs) that reflect effective connectivity in the human brain. Though SPES is gaining popularity in the epilepsy research community, SPES is not part of routine clinical care because the stimulation procedure is time-consuming and can be tedious to perform. Computational methods to infer the full effective connectivity network from a sampled subset would reduce the need to stimulate all brain areas during testing and lower the barrier to routine use in clinical epilepsy investigations. Using inspiration from computer vision, we present "CCEP inpainting", where we predict the effective connections of missing network nodes by minimizing the Dirichlet energy over the missing data domain.
Methods:
We gathered intracranial EEG data collected during SPES from ten epilepsy patients treated for drug-resistant epilepsy at the University of Pennsylvania Hospital. Brief biphasic pulses of electrical stimulation were administered in adjacent pairs of contacts at 1 Hz for 30 trials, and responses were captured in all other contacts. The root-mean-square values of the trial-averaged waveforms in the fast transient (10-50 ms after stimulus) time interval were computed and organized as a matrix with stimulation channels as columns and response channels as rows (Figure 1A). We hid varying percentages of columns (simulating the effective network if specific sites were never stimulated) and then predicted those hidden columns with the inpainting algorithm (Figure 1B). We computed the root-mean-squared error (RMSE) of the difference between the predicted and the original matrix. We also randomized the selection of channels and selected blocked subsets of channels to measure the robustness of the algorithm.
Results:
We used the harmonic image inpainting algorithm to impute the hidden columns of the CCEP matrix for all ten patients. A mean RMSE value of 35.4 +/- 12.2 uV resulted when every other column (50% of the full matrix) was hidden. We found that the RMSE increased on average by 0.94 with the addition of one hidden column in the CCEP matrix. The blocked and randomized subsets of hidden channels were more difficult for the algorithm to predict than the regularly spaced hidden channels.
Conclusions:
We proposed a novel application of image inpainting to CCEP matrices computed from epilepsy patients. The technique demonstrates that some of the effective connections that would have been directly measured from stimulating a specific site in the brain can be accurately predicted from the stimulation of surrounding areas. These results may indicate that only a subset of sites needs to be stimulated to accurately estimate the full true effective connectivity network as defined by SPES. Increased efficiency of the SPES procedure may invite its adoption into routine clinical procedures for epilepsy investigations.
Brain Stimulation:
Invasive Stimulation Methods Other 2
Modeling and Analysis Methods:
Connectivity (eg. functional, effective, structural) 1
Keywords:
Computational Neuroscience
Data analysis
Modeling
1|2Indicates the priority used for review
