Real-Time Functional MRI Classification of Brain States Using Markov-SVM Hybrid Models: Peering Inside the rt-fMRI Black Box

Real-time functional MRI (rt-fMRI) methods provide the ability to predict and detect online changes in cognitive states. Applications require appropriate selection of features, preprocessing routines, and efficient computational models in order to be both practical to implement and deliver interpretable results. We predict video activity in nicotine-addicted subjects using both regional spatial averages and pre-constructed independent component spatial maps we refer to as an ”IC dictionary.” We found that this dictionary predicted better than the anatomical summaries and was less sensitive to preprocessing steps. When prior state information was incorporated using hybrid SVM-Markov models, the online models were able to predict even more accurately in real-time whether an individual was viewing a video while either resisting or indulging in nicotine cravings. Collectively, this work proposes and evaluates models that could be used for biofeedback. The IC dictionary offered an interpretable feature set proposing functional networks responsible for cognitive activity. We explore what is inside the black box of real-time fMRI, and examine both the advantages and shortcomings when machine learning methods are applied to predict and interpret cognitive states in the real-time context.