Resting State fMRI and Machine Learning Applications in Healthy and Patient Populations

Resting state functional magnetic resonance imaging (rs-fMRI) has emerged as a way to characterize brain networks without task difficulty associated with task fMRI. It allows noninvasive examination of neural activity and offers clinically valuable functional information about normal and patient populations and individuals. Advanced neuroimaging measures from fMRI and diffusion tensor imaging (DTI) along with clinical variables acquired during standard imaging protocols offer predictive value for patients when used in machine learning applications. Machine learning methods are well matched in a clinical fMRI setting and serve purposes of prediction and prognostication, diagnosis, insight into normal and patient populations and aid in automation of clinical tasks. In this dissertation, image analysis and machine learning are applied for classification and prediction tasks of: discriminating normal aging subjects and stroke patients, extracting and labeling resting state networks in epilepsy patients and predicting brain tumor patient outcomes. Consequent model interpretation provides insight into the respective underlying processes and reveals influential measures. High accuracy performance of 80-90% accuracy is achieved in binary classification for discriminating age, stroke disease, and prediction of brain tumor outcomes. 80--90% accuracy is also seen for multi-class classification of resting state networks in epilepsy patients. fMRI allows the investigation of neural activity and is well matched in machine learning applications. The studies in this work show that fMRI and DTI provide a rich source of structural and functional information for patient representation in machine learning methods of classification and prediction in normal aging subjects, as well as stroke, epilepsy and brain tumor patients.