The intestinal microbiota predicts COVID-19 severity and fatality regardless of hospital feeding method

Abstract

SARS-CoV-2-positive patients exhibit gut and oral microbiome dysbiosis, which is associated with various aspects of COVID-19 disease (1-4). Here, we aim to identify gut and oral microbiome markers that predict COVID-19 severity in hospitalized patients, specifically severely ill patients compared to moderately ill ones. Moreover, we investigate whether hospital feeding (solid versus enteral), an important cofounder, influences the microbial composition of hospitalized COVID-19 patients. We used random forest classification machine learning models with interpretable secondary analyses. The gut, but not the oral microbiota, was a robust predictor of both COVID-19-related fatality and severity of hospitalized patients, with a higher predictive value than most clinical variables. In addition, perturbations of the gut microbiota due to enteral feeding did not associate with species that were predictive of COVID-19 severity. IMPORTANCE SARS-CoV-2 infection leads to wide-ranging, systemic symptoms with sometimes unpredictable morbidity and mortality. It is increasingly clear that the human microbiome plays an important role in how individuals respond to viral infections. Our study adds to important literature about the associations of gut microbiota and severe COVID-19 illness during the early phase of the pandemic before the availability of vaccines. Increased understanding of the interplay between microbiota and SARS-CoV-2 may lead to innovations in diagnostics, therapies, and clinical predictions.

Keywords: COVID-19; diet; enteral feeding; fatality; microbiota; predictors; random forest classification; severity.

Fig 1

Stool microbiome variables contribute to predictions of COVID-19 fatality. (A) Random forest classification modeling using clinical covariates alone identifies severity as the top predictor of fatality. Only clinical variables defined as significant by the Boruta algorithm were included in the model. (B) When combined with stool microbiome variables, combined modeling with microbiome abundances classified at the species level improved F1-score compared to clinical variables alone. Boxplots show mean, first, and third quartile scores from 11 modeling runs using the leave-one-out cross-validation method. (C) Feature ranking of combined models (clinical plus microbiome variables) identifies severity as the top predictor of fatality regardless of the microbiome abundance classification method. Stool microbiota species outrank most other clinical variables.

Fig 2

Stool microbiota can help predict COVID-19 severity and are independent of hospital feeding methods. (A) Combined random forest classification modeling of clinical variables and stool microbiome abundances classified at the ASV level improved the prediction of COVID-19 severity in our cohort. Boxplots show mean, first, and third quartile scores from 11 modeling runs using the leave-one-out cross-validation method. (B) Feature ranking of the combined model using ASV classification shows that stool microbiota species outrank most clinical variables in predictive importance. (C) SIRUS identified abundance changes in certain stool species that were predictive of severe or moderate disease. (D) Analysis of species correlated with enteral feeding or solid hospital food diet identified only two species that were among the top bacteria predictors for severity.