Comparison of signal sources for real-time fMRI neurofeedback

1958 

Abstract Submission 

Dong-Youl Kim¹, Jonathan Lisinski¹, Matthew Caton¹, Brooks Casas^1,2, Stephen LaConte^1,3, Pearl Chiu^1,2

¹Fralin Biomedical Research Institute at VTC, Virginia Tech, Roanoke, VA, ²Department of Psychology, Virginia Tech, Blacksburg, VA, ³Department of Biomedical Engineering and Mechanics, Virginia Tech, Blacksburg, VA

Dong-Youl Kim  
Fralin Biomedical Research Institute at VTC, Virginia Tech
Roanoke, VA

Jonathan Lisinski  
Fralin Biomedical Research Institute at VTC, Virginia Tech
Roanoke, VA

Matthew Caton  
Fralin Biomedical Research Institute at VTC, Virginia Tech
Roanoke, VA

Brooks Casas  
Fralin Biomedical Research Institute at VTC, Virginia Tech|Department of Psychology, Virginia Tech
Roanoke, VA|Blacksburg, VA

Stephen LaConte  
Fralin Biomedical Research Institute at VTC, Virginia Tech|Department of Biomedical Engineering and Mechanics, Virginia Tech
Roanoke, VA|Blacksburg, VA

Pearl Chiu  
Fralin Biomedical Research Institute at VTC, Virginia Tech|Department of Psychology, Virginia Tech
Roanoke, VA|Blacksburg, VA

Extant studies have provided evidence for the feasibility of voluntarily modulating neural features using real-time functional magnetic resonance imaging (rtfMRI)-based neurofeedback (NF) approaches to affect symptoms or control of cognitive processes [1,2,4-6]. Recent applications of rtfMRI-NF approaches have adopted a machine learning (ML)-based accuracy to estimate NF signals rather than using neural activation (NA) within regions-of-interest (ROIs) or ROIs-based functional connectivity (FC), but no study has directly compared NF performance based on measurements. In this study, we evaluated NF training performance for voluntarily regulating smoking craving across repeated rtfMRI-NF runs depending on these measurements.

Thirty-one smokers performed three rtfMRI-NF runs while viewing a series of smoking-related images and receiving instructions, based on task condition, to either 'crave' or 'don't crave'. The prior run trained the ML model to be tested in the following run (Figure 1). To classify task conditions, NF signal was reflecting the distance to the hyperplane of support vector machine (3d-svm with a linear kernel implemented in AFNI [3]). A slider-bar integrated the feedback signal within a block and the step size for each volume was set based on the block length. To evaluate the performance of ML-based accuracy, ROI-based NA, and ROIs-based FC NF measurements, individual classification accuracy with k-fold (k = 10) nested cross-validation scheme was estimated using the percentage of BOLD signals to compare task conditions within each run. NA was measured within each ROI related to smoking craving, such as the left and right anterior cingulate cortex, posterior cingulate cortex, and insula, through the comparison of beta values in task conditions. FC was estimated between pairs of ROIs using the BOLD signals by concatenating time-series for each task condition. Pearson's correlation coefficients were converted by Fisher's r-to-z transformation and z-scored FC was subtracted between two conditions. Post-hoc paired t-tests and a linear regression model were performed to compare the measurements between pairs of the three runs, and across the three runs, respectively.

Figure 2 presents ROI-based NA (A), ROIs-based FC (B), and individual classification accuracy (C), and the corresponding spatial patterns (D), where the weight features of classification were subjected to a one-sample t-test for group inference. From the ROI-based NA, only the anterior cingulate cortex showed marginal significance of changing activation across repeated runs and significant difference between a pair of runs was found in the left anterior cingulate cortex and bilateral posterior cingulate cortex. From the ROI-based FC, only connection between the left and right posterior cingulate cortex represented marginal significance of changes across repeated runs. From the individual classification, high accuracy (> 80%) and consistently increased performance across the repeated runs was observed for training and test sets with a statistical significance (p < 0.001) of linear regression. The left anterior insula and bilateral caudate predicted the conditions from the following runs compared to the first run and frontal regions showed decreased tendency across repeated runs. Overall spatial patterns of the insula and frontal areas showed shrinkage patterns from the third run compared to the second run.

Our findings demonstrate the feasibility of our machine learning based rtfMRI-NF method through the enhanced discrimination of smoking craving across repeated NF training. Future work is needed to examine the efficacy of our method against other measures of neural signal, such as network analysis. In addition, meta-analyses would be warranted to evaluate the homogeneity and heterogeneity of NF performance depending on the measurements in consideration of populations and cognitive processes.

Activation (eg. BOLD task-fMRI)

Classification and Predictive Modeling ²

Connectivity (eg. functional, effective, structural)

Multivariate Approaches ¹

Addictions

FUNCTIONAL MRI

Machine Learning

Other - Real-time fMRI, neurofeedback, smoking craving, cigarettes, support vector machine, individual classification, neuronal activation.