| Alzheimer's disease is a major cause of dementia. The disease pathology induces spatially complex patterns in the patient's brain that evolve as the disease progresses. The diagnosis requires accurate biomarkers that are sensitive to disease stages. Neuroimaging biomarkers derived from PET and MRI scans have been widely used to this end. However, the visual inspection of neuroimagery is susceptible to errors due to human limitations. Given the neuroimaging data, computerized methods can be more accurate than human experts. In this study, we regard the creation of a model for classification as designing a computational biomarker for disease staging. Analogously, the creation of a probabilistic classifier results in a probabilistic biomarker. We are in favor of probabilistic biomarkers due to their built-in support for interpretation of decisions. We obtain the probabilistic biomarkers via Gaussian Processes. In comparison with the well-known Support Vector Machine, the Gaussian Process models allow us to make better use of the scarce neuroimaging data. In addition, they offer more flexible means to carry out the Multiple Kernel Learning and tackle the challenges of high dimensionality. Our research results also demonstrate that the Gaussian Process models are interpretable in terms of the neural correlates of Alzheimer's disease. In conclusion, the Gaussian Process models proposed as probabilistic biomarkers in this dissertation emphasize the most prominent anatomical regions for accurate staging of Alzheimer's disease. Furthermore, they are competitive with or better than the Support Vector Machine, which has been the workhorse of the multivariate predictive analysis of brain data. |
