| Earlier work has suggested that various neurological or neuropsychiatric disorders may result in characteristic spatial patterns in brain activation, potentially allowing their detection from maps of brain activity under different conditions. However, standard techniques for analysis of brain activity have traditionally focused on finding the most significant areas of brain activations for a given function, and this approach does not apply well to detecting disorders in individuals.; In this work, we test several simple and well-understood computational approaches to pattern recognition to attempt to detect patients with a single kind of disorder (early Alzheimer's disease, schizophrenia, or mild traumatic brain injury) from among healthy controls. The primary method uses a linear approach based on principal components analysis (PCA) and the Fisher linear discriminant (FLD) to cast the problem as one of pattern recognition within a small subspace of the very large space of possible brain activity. In addition to its simplicity and elegance, this method has the additional advantage of allowing for the integration of multiple maps of brain activity for each subject, and can correctly separate patients from controls with approximately 70--85% accuracy for each of three different disorders.; This work also considers extensions to the basic method: nonlinear dimensionality reduction (multidimensional scaling), a nonlinear version of the linear classifier (Kernel Fisher Discriminant), and incorporation of anatomic data, all of which can support useful classification when appropriately parameterized. Nearest neighbor methods, including K nearest neighbors (KNN) and nearest mean classification (NMC) are also tested and found to classify adequately in many cases. In addition, a data sharing framework for allowing indexing and exchange of sensitive data is presented with the aim of allowing integration of activation maps from multiple sites into larger studies. |
