Unsupervised Multiple Kernel Learning to characterize Heart Failure patients considering myocardial mechanics and hemodynamics

Abstract

Heart Failure with preserved Ejection Fraction has proven to be a suitable syndrome to be studied with machine learning approaches, as its complexity is not fully captured in clinical guidelines. For this reason, in this work we present a pipeline to characterize patients from Heart Failure with Preserved Ejection Fraction cohorts. This comprises from the creation of the databases to the development of a computational platform in Python to process the images and extract the descriptors. Lastly, we implemented machine learning and dimensionality reduction techniques to explore the data and clustering and kernel regression to obtain physiological insights on the population. We validated the Echocardiographic Image Analysis platform with clinical data from two clinical trials, succeeded at creating meaningful clusters to classify healthy and diseased patients and obtained an output space from Multiple Kernel Learning which encoded the principal modes of cardiac dysfunction.