Utilization of domain-knowledge for simplicity and comprehensibility in predictive modeling of Alzheimer's disease

Positron Emission Tomography scans are a promising source of information for early diagnosis of Alzheimer's disease. However, such neuroimaging procedures usually generate high-dimensional data. This complicates statistical analysis and modeling, resulting in high computational complexity and typically more complicated models. However, the utilization of domain-knowledge can reduce the complexity and promote simpler models. In this study, we investigate Gaussian processes, which may incorporate domain-knowledge, for predictive modeling of Alzheimer's disease. This study uses features extracted from PET imagery by 3D Stereotactic Surface Projection. Since the number of features can be high even after applying prior knowledge, we examine the benefits of a correlation-based feature selection method. Feature selection is desirable as it enables the detection of metabolic abnormalities that only span certain portions of the anatomical regions. Our proposed utilization of Gaussian processes is superior to the alternative (Automatic Relevance Determination), resulting in more accurate diagnosis with less computational effort.