Unsupervised Machine Learning Unveil Easily Identifiable Subphenotypes of COVID-19 With Differing Disease Trajectories

Background Given the clinical heterogeneity of COVID-19 infection, we hypothesize the existence of subphenotypes based on early inflammatory responses that are associated with mortality and additional complications. Methods For this cross-sectional study, we extracted electronic health data from adults hospitalized patients between March 1, 2020 and May 5, 2021, with confirmed primary diagnosis of COVID-19 across five Johns Hopkins Hospitals. We obtained all electronic health records from the first 24h of the patient’s hospitalization. Mortality was the primary endpoint explored while myocardial infarction (MI), pulmonary embolism (PE), deep vein thrombosis (DVT), stroke, delirium, length of stay (LOS), ICU admission and intubation status were secondary outcomes of interest. First, we employed clustering analysis to identify COVID-19 subphenotypes on admission with only biomarker data and assigned each patient to a subphenotype. We then performed Chi-Squared and Mann-Whitney-U tests to examine associations between COVID-19 subphenotype assignment and outcomes. In addition, correlations between subphenotype and pre-existing comorbidities were measured using Chi-Squared analysis. Results A total of 7076 patients were included. Analysis revealed three distinct subgroups by level of inflammation: hypoinflammatory, intermediate, and hyperinflammatory subphenotypes. More than 25% of patients in the hyperinflammatory subphenotype died compared to less than 3% hypoinflammatory subphenotype (p<0.05). Additional analysis found statistically significant increases in the rate of MI, DVT, PE, stroke, delirium and ICU admission as well as LOS in the hyperinflammatory subphenotype. Conclusion We identify three distinct inflammatory subphenotypes that predict a range of outcomes, including mortality, MI, DVT, PE, stroke, delirium, ICU admission and LOS. The three subphenotypes are easily identifiable and may aid in clinical decision making.