Composite Kernels for Automatic Relevance Determination in Computerized Diagnosis of Alzheimer's Disease

Voxel-based analysis of neuroimagery provides a promising source of information for early diagnosis of Alzheimer's disease. However, neuroimaging procedures usually generate high-dimensional data. This complicates statistical analysis and modeling, resulting in high computational complexity and typically more complicated models. This study uses the features extracted from Positron Emission Tomography imagery by 3D Stereotactic Surface Projection. Using a taxonomy of features that complies with Talairach-Tourneau atlas, we investigate composite kernel functions for predictive modeling of Alzheimer's disease. The composite kernels, compared with standard kernel functions (i.e. a simple Gaussian-shaped function), better capture the characteristic patterns of the disease. As a result, we can automatically determine the anatomical regions of relevance for diagnosis. This improves the interpretability of models in terms of known neural correlates of the disease. Furthermore, the composite kernels significantly improve the discrimination of MCI from Normal, which is encouraging for early diagnosis.