Imbalanced fMRI Classification: An Ensemble Approach

Abstract

Functional Magnetic Resonance Imaging (fMRI) has recently come to dominate the brain mapping field due to its reliable, noninvasive ability to gather massive amounts of data about task-specific brain activity. The high-dimensional, often noisy datasets recorded by these studies, however, pose significant statistical challenges that must be overcome before the data is ready for neurological interpretation. One such challenge is that of imbalanced data, where instances of one class far outnumber the other, which makes accurate classification very difficult by traditional methods. This thesis studies a series of imbalanced datasets from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and develops a new three-phase ensemble model that combines majority class undersampling through unsupervised K-Medoids clustering, an iterative l(_{2}) logistic loss algorithm for feature selection, and a divide and conquer classification strategy to produce reliably stronger performance than the field’s current frontrunners.