A hybrid approach based on dynamic trajectories to predict mortality in COVID-19 patients upon steroids administration

Abstract

Since the World Health Organization (WHO) has declared Artificial Intelligence (AI) as a powerful tool in the fight against COVID-19, multiple studies have been launched aiming to shed light into risk factors for ICU admission and mortality. None of the existing studies, however, have captured the dynamic trajectories of hospitalized COVID-19 patients who receive steroids nor have explored trajectory-based mortality indicators. In this work, we present a novel, hybrid approach to address this need. Latent Growth Mixture Modelling (LGMM) was used to analyze the trajectories of patients who received steroids. The patients were then grouped into clusters based on the similarity of their dynamic trajectories. State-of-the art machine learning classifiers are trained on the original dataset with and without dynamic trajectories to assess whether their inclusion can enhance the prediction of mortality. Our results highlight the importance of trajectories for predicting mortality in patients who receive steroids yielding 4% and 5% increase in the sensitivity (0.84) and specificity (0.85). The FiO2 and percentage of neutrophils at day 5, along with the percentage of lymphocytes at day 7, were identified as the main causes for mortality in patients who receive steroids, where the SatO2 levels showed significant alterations in the dynamic trajectories.