Comparison of neural signal sources for discriminating ‘crave’ and ‘don’t crave’ task conditions: Implications for fMRI neurofeedback

Numerous prior studies have shown that neural features can be voluntarily modulated through real-time functional magnetic resonance imaging (rtfMRI)-based neurofeedback (NF), and that such modulation can influence cognitive and affective states, as well as improve psychiatric and neurological symptoms. To generate neurofeedback, these studies have leveraged information from a variety of signal sources, including neural activity within regions-of-interest (ROIs) and machine learning (ML) based multivariate networks, however little work has evaluated the relative performance of differing sources. In this study, based on the premise that improved differentiation between task conditions ought to improve neurofeedback, and to provide a foundation for future neurofeedback studies, we evaluated the performance of four different neural signal sources for discriminating among ‘crave’ and ‘don’t crave’ task conditions using our previously published data in which participants regulated cigarette craving across rtfMRI-NF runs. Thirty-one smokers performed three runs including an initial run for training a baseline ML model and subsequent run for testing the baseline model, as well as an updated model from the second run being tested during the third run. The ML NF signal reflected the distance to the hyperplane of a support vector machine classifying ‘crave’ and ‘don’t crave’ conditions. To compare the performance of different neural signal sources for discriminating task conditions over repeated runs, we evaluated within-ROI neural activity, between-ROIs functional connectivity, activity-based classification accuracy, and connectivity-based classification accuracy for differentiating ‘crave’ and ‘don’t crave’ task conditions over runs. ML-based sources showed better performance for discriminating conditions than non-ML-based sources, with connectivity-based ML sources exhibiting the best performance among signal sources. Taken together, we demonstrate that multivariate machine learning based analyses enhance discriminability of neural measures associated with smoking cravings across repeated neurofeedback training and suggest the potential for seed-based functional connectivity to further enhance discriminability.