Methods and Applications of Multivariate Pattern Analysis in Functional MRI Data Analysis.

Abstract

In spite of the tremendous advances in science and technology, the human brain and its functions are still not completely understood. Functional magnetic resonance imaging (fMRI) is an imaging modality that allows for non-invasive study of brain function and physiology. Thus, fMRI has found many applications in various fields involved in the study of cognition, psychology, psychiatry, neuroscience, etc. Machine learning techniques have gained tremendous interest in recent times for fMRI data analysis. These methods involve learning from numerous examples and then making predictions for new unseen examples. This work addresses the use of machine learning techniques to find and study multivariate patterns in the fMRI brain data.

The two main applications explored in this work include temporal brain-state prediction and subject categorization. The within-subject brain-state prediction setup has been used to compare and contrast three different acquisition techniques in a motor-visual activation study. It has also been implemented to highlight the differences in pain regulation networks in healthy controls and subjects with temporomandibular disorders. Lastly, regression has been used to predict graded fMRI activation on a continuous scale in a motor activation and craving study. The between-subject categorization setup has been used to distinguish between patients with Asperger's disorder and healthy controls.

A major contribution of our work involves a novel multi-subject machine learning framework. This technique helps to learn a model which is based on information acquired from multiple other subjects' data in addition to the subject's own data. This has been used to classify the craving and non-craving brain states of nicotine-dependent subjects, allowing examination of both population-wide as well as subject-specific neural correlates of nicotine craving. A real-time neurofeedback setup was implemented to provide feedback to a subject using their own brain activation data. Subjects can then be trained to self-regulate their own brain activation.