Predictors of hospital outcomes and clinical subphenotypes differ between COVID-19 and influenza pneumonia

ABSTRACT

Introduction: Pneumonia due to SARS-CoV-2 (Coronavirus Disease 2019, COVID-19) has frequently been compared to other viral pneumonias, including influenza. While some data suggest significant differences in biological responses, dissimilarities in the clinical course and characteristics between SARS-COV-2 and influenza pneumonia remain unknown. We evaluated differences in clinical predictors of outcomes and early clinical subphenotypes in COVID-19 and influenza pneumonia. Methods: We performed a retrospective cohort study of all patients hospitalized for > 24 hours, requiring oxygen support, at Barnes-Jewish Hospital with COVID-19 (March-July 2020) or influenza (Jan 2012-Dec 2018). In-hospital mortality or hospice discharge was the primary outcome. First, supervised machine learning classifier models (XGBoost) were trained using bootstrap replications of each viral cohort to predict the primary outcome. 28 candidate predictor variables among the most extreme vital signs and laboratory values within 24 hours of hospitalization were preselected, excluding highly correlated variables. We compared each model's internal discrimination to its performance in the alternate cohort and evaluated differences in variable importance between the two viral pneumonia models. Next, we evaluated differences in clinical subphenotypes in two ways 1) a previously-validated algorithm to group patients into four distinct subphenotypes based on temperature trajectories within 72 hours of hospitalization;2) latent class analysis (LCA) to identify unmeasured subgroups within each viral cohort based on the predictor variables described above. In both analyses, we compared frequency of subphenotype membership and each subphenotype's primary outcome between viral cohorts. Results: We evaluated 321 unique hospitalizations with COVID-19 and 535 with influenza. The primary outcome was experienced in 23% and 9.5% of patients, respectively. Influenza predictor model discriminated outcomes worse in COVID-19 than on internal evaluation (Panel A), suggesting prognostic variables differ between the viral pneumonias. Only one of the top five contributory variables was shared between the two models (Panel B). Prevalences of temperature trajectory subphenotype also differed significantly between viral pneumonias. All COVID-19 temperature trajectory subphenotypes experienced the primary outcome more frequently than their influenza counterparts (Panel C). LCA identified two distinct classes in each cohort, with each viral pneumonia's minority class experiencing worse outcomes than the majority class. Of each model's top 5 classdefining variables, only 2 were shared (Panel D). Conclusions: COVID-19 and influenza pneumonia differ markedly in predictors of outcome and in clinical subphenotypes. These findings emphasize observable pathogen-specific differential responses in viral pneumonias and suggest that distinct management approaches should be investigated for these diseases. (Table Presented).