Nonlinear dimensionality reduction combining MR imaging with non-imaging information

We propose a framework for the extraction of biomarkers from low-dimensional manifolds representing inter-subject brain variation. Manifold coordinates of each image capture information about structural shape and appearance and, when a phenotype exists, about the subject's clinical state. Our framework incorporates subject meta-information into the manifold learning step. Apart from gender and age, information such as genotype or a derived biomarker is often available in clinical studies and can inform the classification of a query subject. Such information, whether discrete or continuous, is used as an additional input to manifold learning, extending the Laplacian Eigenmap objective function and enriching a similarity measure derived from pairwise image similarities. The biomarkers identified with the proposed method are data-driven in contrast to a priori defined biomarkers derived from, e.g., manual or automated segmentations. They form a unified representation of both the imaging and non-imaging measurements, providing a natural use for data analysis and visualization. We test the method to classify subjects with Alzheimer's Disease (AD), mild cognitive impairment (MCI) and healthy controls enrolled in the ADNI study. Non-imaging metadata used are ApoE genotype, a risk factor associated with AD, and the CSF-concentration of Aβ _1-42, an established biomarker for AD. In addition, we use hippocampal volume as a derived imaging-biomarker to enrich the learned manifold. Our classification results compare favorably to what has been reported in a recent meta-analysis using established neuroimaging methods on the same database.