Experiments of machine learning for neurodiagnosis

Resumen

The application of Machine Learning algorithms to Neuroscience data has two main goals in this thesis. First, the construction of Computer Aided Diagnostic (CAD) systems to help alleviate the burden of increasing amounts of data for diagnosis. Second, the identification of image biomarkers corresponding to anatomical locations of the features selected for classification. This thesis is an empirical exploration of these ideas in the case of three neurological diseases. We have developed sound methodological frameworks, avoiding circularity effects in the validation process. For feature selection we have been working with supervised methods which produce good classification results and the ability to determine the spatial location of the features in the brain, among them an evolutionary wrapper selection method based on Extreme Learning Machines. The thesis covers experiments with a wide spectrum of classifiers for comparison. Finally, the approach has been tested on different MRI modalities showing its general applicability. A critical issue for multivariate modalities is the definition of appropriate scalar measures that may be useful for feature selection and extraction. The thesis tests the most appropriate for each modality: FA and MD for diffusion data, functional activity measures such as ReHo for functional data and measures from deformation maps resulting from non-linear registration of anatomical data. The results obtained show that the approach is useful for CAD systems in a variety of neurological diseases.