Resumen de Multi-feature machine learning analysis for an improved characterization of the cardiac mechanics
This thesis focuses on the development of machine learning tools to better characterize the cardiac anatomy and function in the context of heart failure, and in particular their extension to consider multiple parameters that help identifying the pathophysiological aspects underlying disease. This advanced and personalized characterization may eventually allow assigning patients to clinically-meaningful phenogroups with a uniform treatment response and/or disease prognosis. Specifically, the thesis copes with the technical difficulties that multivariate analyses imply, paying special attention to properly combine different descriptors that might be of different nature (e.g., patterns, continuous, or categorical variables) and to reduce the complexity of large amounts of data up to a meaningful representation. To this end, we implemented an unsupervised dimensionality reduction technique (Multiple Kernel Learning), which highlights the main characteristics of complex, high-dimensional data into fewer dimensions. For our computational analysis to be useful for the clinical community, it should remain fully interpretable. We made special emphasis in allowing the user to be aware of how the input to the learning process models the obtained output, through the use of multi-scale kernel regression techniques among others.
