Background: The identification of factors associated with Intensive Care Unit (ICU) mortality and derived clinical phenotypes in COVID-19 patients could help for a more tailored approach to clinical decision-making that improves prognostic outcomes.
Methods: Prospective, multicenter, observational study of critically ill patients with confirmed COVID-19 disease and acute respiratory failure admitted from 63 Intensive Care Units(ICU) in Spain. The objective was to analyze patient’s factors associated with mortality risk and utilize a Machine Learning(ML) to derive clinical COVID-19 phenotypes. Patient features including demographics and clinical data at ICU admission were analyzed. Generalized linear models were used to determine ICU morality risk factors. An unsupervised clustering analysis was applied to determine presence of phenotypes. The prognostic models were validated and their performance was measured using accuracy test, sensitivity, specificity and ROC curves.
Results: The database included a total of 2,022 patients (mean age 64[IQR5-71] years, 1423(70.4%) male, median APACHE II score (13[IQR10-17]) and SOFA score (5[IQR3-7]) points. The ICU mortality rate was 32.6%. Of the 3 derived phenotypes, the C(severe) phenotype was the most common (857;42.5%) and was characterized by the interplay of older age (>65 years), high severity of illness and a higher likelihood of development shock. The A(mild) phenotype (537;26.7%) included older age (>65 years), fewer abnormal laboratory values and less development of complications and B (moderate) phenotype (623,30.8%) had similar characteristics of A phenotype but were more likely to present shock. Crude ICU mortality was 45.4%, 25.0% and 20.3% for the C, B and A phenotype respectively. The ICU mortality risk factors and model performance differed between whole population and phenotype classifications.
Conclusion: The presented ML model identified three clinical phenotypes that significantly correlated with host-response patterns and ICU mortality. Different risk factors across the whole population and clinical phenotypes were observed which may limit the application of a “one-size-fits-all” model in practice.