Biology-guided algorithms Improved cardiovascular risk prediction and biomarker discovery

Medical research has seen a stark increase in the amount of available data.
The sheer volume and complexity of measured variables challenge the use of traditional statistical methods and are beyond the ability of any human to comprehend. Solving this problem demands powerful models capable of capturing the variable interactions and how those are non-linearly related to the condition under study. In this thesis, we first use Machine Learning (ML) methods to achieve better cardiovascular risk prediction/disease biomarker identification and then describe novel bio-inspired algorithms to solve some of the challenges.
On the clinical side, we demonstrate how combining targeted plasma proteomics with ML models outperforms traditional clinical risk factors in predicting first-time acute myocardial infarction as well as recurrent atherosclerotic cardiovascular disease. We then shed some light on the pathophysiological pathways involved in heart failure development using a multi-domain ML model. To improve prediction, we develop a novel graph kernel that incorporates protein-protein interaction information, and suggest a manifold mixing algorithm to increase inter-domain information flow in multi-domain models.
Finally, we address global model interpretability to uncover the most important variables governing the prediction. Permutation importance is an intuitive and scalable method commonly used in practice, but it is biased in the presence of covariates. We propose a novel framework to disentangle the shared information between covariates, making permutation importance competitive against methodologies where all marginal contributions of a feature are considered, such as SHAP.