Multi-task Models for Predicting the Progression of Alzheimer's from MRI

Abstract

Alzheimer's disease (AD) is a severe neurodegenerative disorder a ecting over 5 million in the U.S. today. By 2050, the total is estimated to reach 13.8 million. Researchers are scrambling to nd methods for accurate and early prediction of AD progression, since its irreversible symptoms are particularly destructive and no treatments have been found. Previous studies have shown that changes in certain regions of the brain, as measured using magnetic resonance imaging (MRI), have been associated with the onset of AD. Clinical diagnosis of AD often relies on cognitive measures such as the Mini Mental State Examination (MMSE) and the Alzheimer's Disease Assessment Scale cognitive subscale (ADAS-Cog). In this thesis, we address the problem of predicting MMSE and ADAS-Cog scores at future time points from high-dimensional MRI data collected at a baseline initial visit. We structure the problem as a multi-task regression problem, where each task is the prediction at a certain time point. We propose three novel calibrated multi-task formulations and conduct experiments to evaluate their performance, using two separate data sets from the Alzheimer's Disease Neuroimaging Initiative (ADNI), and compare the results to those of their non-calibrated counterparts, as well as those of the single-task ridge and Lasso regressions. Multi-task methods consistently outperform single-task methods, and of those, the calibrated formulations achieve near-identical performance to those that are non-calibrated. Overall, the best model for the prediction of AD progression is not always intuitive but dependent on the input data.