SOLIDS: Statistical mOdeLs for COVID-19 Severity

COVID-19 is a rapidly expanding pandemic caused by the virus SARS-COV-2. As a new disease, its poor understanding can result in delayed identification and treatment. Hence, during the COVID-19 pandemic, is of crucial importance to (1) understand the factors that may influence the disease trajectory and increase the probability of the disease aggravation for hospitalized patients and (2) develop tools to assess the disease severity and predict the development of the infection, namely at an early stage, to shorten the disease course and reduce mortality. In addition, both aspects are also of the utmost importance for planning the allocation of medical resources. Many existing studies suggest that several patients’ characteristics may influence the COVID-19 length of stay (LOS) in hospitals, but often no conclusion is drawn because LOS is typically reported as a secondary outcome without being further analyzed. Up to now there is a very limited number of studies regarding the factors affecting the hospital and ICU length of stay. Also, the few studies are population specific and some mention important limitations namely the exclusion of some vital factors, such as previous patients’ diagnosed diseases. In addition, no consensus on particular risk factors has been reached which reinforces the need for further research. Lung ultrasound is a long-recognized highly accurate tool for the diagnosis of pulmonary diseases. For COVID-19 patients, a lung ultrasound score (LUS) is typically composed using nominal/ordinal scales to measure a set of sub-indicators evaluated directly on the images. Up to now this global score is given by the simple summation of the sub-indicators. In this approach all variables are given the same weight as no mathematical grounds exist for choosing a different scheme. This equal weighting scheme may imply the recognition of an equal status for all sub-indicators, which may not be necessarily true, potentially leading to a misrepresentative, biased and/or highly uncertain indicator. Therefore, careful attention needs to be given to the construction of this composite indicator to ensure a precise and unbiased estimator of the underlying trait (disease severity). Moreover, some studies suggest that LUS may act as a good predictor of death, ICU admission, and endotracheal intubation in COVID-19. Therefore, main goals of this study include 1) to describe the distribution of the clinical characteristics of hospitalized COVID-19 patients; 2) to explore the risk factors for LOS variation including patients’ demographic characteristics and clinical indicators regarding inflammation and renal, liver, heart and lung functions using classic and Bayesian survival analysis, (3) to derive a lung ultrasound COVID-19 index optimizing its relationship with the proxy variable for disease severity defined by the hospital length of stay, the ICU length of stay or the survival time depending on the considered patient outcome, (4) to evaluate the potential of this new indicator as a predictor of death, ICU admission, and endotracheal intubation in COVID-19 patients using ML models, and (5) to compare the derived composite indicator with the classic LUS.

Team: Regina Bispo, Isabel Natário and Inês Sequeira.