Application of multiblock techniques in metabolomic and clinical data for the prediction of ventilatory therapies in Covid-19 patients

Abstract

Multiblock techniques, as the name suggests, are analytical methods that handle datasets organized into multiple blocks, each of them representing a set of variables with similar nature. These methods aim to model relationships across these blocks, delving into the internal structure of the data to uncover patterns that might not be apparent when analysing them individually. In this thesis, regularised CCA, PLS and multiblock-PLS-DA (DIABLO) have been applied to Covid-19 patients’ data to study the behaviour and relationships between two sets of variables, one containing metabolomic measures and the other evaluating different clinical features. Additionally, two different approaches have been proposed and discussed to incorporate multiblock analysis into the prediction of the failure of non-invasive respiratory support in Covid-19 patients. This has allowed not only the exploration of the relationship net between blocks but also the identification of similarities and correlations among the relevant variables in a predictive context. At the end, multiblock techniques have shown their potential to extract valuable information from the dataset and to optimize and maximize the classification procedure, leading to the identification of three variables (anthranilic acid, octanoic acid and malate/succinate ratio) that allow to successfully predict the failure of non-invasive respiratory support in Covid-19 patients.