Efficient and Automatic Subspace Relevance Determination via Multiple Kernel Learning for High-Dimensional Neuroimaging Data

摘要

Alzheimer's disease is a major cause of dementia. Its pathology induces complex spatial patterns of brain atrophy that evolve as the disease progresses. The diagnosis requires accurate biomarkers that are sensitive to disease stages. Probabilistic biomarkers naturally support the interpretation of decisions and evaluation of uncertainty associated with them. We obtain probabilistic biomarkers via Gaussian Processes, which also offer flexible means to accomplish Multiple Kernel Learning. Exploiting this flexibility, we propose a novel solution, Multiple Kernel Learning for Automatic Subspace Relevance Determination, to tackle the challenges of working with high-dimensional neuroimaging data. The proposed Gaussian Process models are competitive with or better than the well-known Support Vector Machine in terms of classification performance even in the cases of single kernel learning. Also, our method improves the capability of the Gaussian Process models and their inter-pretability in terms of the known anatomical correlates of the disease.