Interpretable machine learning models for investigating and pre-screening cardiomyopathies with high-dimensional genomic and imaging data

Abstract

Cardiomyopathies have long been considered as monogenic, Mendelian diseases with rare variant implications; however, as more studies look at the common variation in these low prevalence diseases, additional discoveries of the polygenic interplay are being revealed. Part of these discoveries also associate genomics with phenotype, specifically derived from cardiac magnetic resonance (CMR) imaging, and with the use of machine learning (ML), insights to the complex diversity of these diseases are revealed. The overall purpose of this thesis is to leverage the information hidden in genomic and imaging datasets with the use of ML to either 1) develop new methodologies for handling such complex datatypes and 2) create ML models that can potentially predict or generate risk scores for cardiomyopathies, based on genomic, clinical and imaging data. This is achieved by developing LAVASET, a novel ensemble algorithm that enhances the Random Forest algorithm and is able to mitigate variable correlation issues. In addition, traditional and interpretable ML algorithms are used to pre-screen and stratify cases of cardiomyopathy, while also displaying increased model performance when using LAVASET-derived imaging features. Finally, a deep-learning method for calculating Polygenic Risk Scores (PRS) is described, showing promising associations between high PRS and disease risk as well as high PRS and known cardiomyopathy outcomes and traits.