| Computational methods have been rigorously studied and deployed for many advances in natural sciences, in particular, for the study of biological processes, diseases, and disorders. One such domain that benefited from computational algorithms is the study of brain diseases like Alzheimer's Disease and other forms of dementia. A natural next step for several such studies is towards developing, and testing of targeted drugs via appropriately designed clinical trials. The context of this work is the interface of well-studied classical approaches to designing efficient clinical trials, and the remarkable success of data-driven machine learning and computer vision methods in modeling the underlying disease. The questions we pose include---What goes into setting up a clinical trial? And how can machine learning methods be leveraged into improving this setup? The hypothesis for this thesis derives from the fact that non-trivial structure in datasets---for instance, brain scans and cognitive tests acquired for studying clinical changes in Alzheimer's disease---should provide useful information for designing clinical trials.;This thesis presents new testing, learning and inference algorithms primarily aimed at (a) designing efficient outcomes for measuring clinical trial efficacy; (b) selecting optimal study population for the clinical trial; (c) harmonizing computational outcomes and enrichers across multiple trial sites; and (d) exploring hierarchical interactions of disease markers to improve trial design. Beyond clinical trials, the proposed algorithms are applicable for fast hypothesis testing, learning neural networks from small-sized datasets, selecting/ranking hierarchically interacting entities, interpreting learned representations and other problems in machine learning and computer vision. Open-source implementations accompany each algorithm. |
