Surface Based Decoding of Fusiform Face Area Reveals Relationship Between SNR and Accuracy in Support Vector Regression

Abstract

The objective of this study was to expand on a method previously established in the lab for predicting subcortical structures using functional magnetic resonance imaging (fMRI) data restricted to the cortical surface. Our goal is to enhance the utility of low cost, portable imaging modalities, such as functional near infrared spectroscopy (fNIRS), which is limited in signal penetration depth. Previous work in the lab successfully employed functional connectivity to predict ten resting state networks and six anatomically de fined structures from the outer 10 mm layer of cortex using resting state fMRI data. The novelty of this study was two-fold: we chose to predict the functionally de fined region fusiform face area (FFA), and we utilized the functional connectivity of both resting state and task activation. Right FFA was identi ed for 27 subjects using a general linear model of a functional localizer tasks, and the average time series were extracted from right FFA and used as training and testing labels in support vector regression (SVR) models. Both resting state and task data decoded activity in right FFA above chance, both within and between run types. Our method is not specific to resting state, potentially broadening the scope of research questions depth-limited techniques can address. We observed a similarity in our accuracy cross-validation to previous work in the lab. We characterized this relationship between prediction accuracy and spatial signal-to-noise (SNR). We found that this relationship varied between resting state and task, as well as the functionality of features included in SVR modeling.