Identification and validation of respiratory subphenotypes in patients with COVID-19 acute respiratory distress syndrome undergoing prone position

Abstract Background The heterogeneity of acute respiratory distress syndrome (ARDS) patients is a challenge for the development of effective treatments. This study aimed to identify and characterize novel respiratory subphenotypes of COVID-19 ARDS, with potential implications for targeted patient management. Methods Consecutive ventilated patients with PCR-confirmed COVID-19 infection, in which prone positioning was clinically indicated for moderate or severe ARDS, were included in a prospective cohort. The patients were assigned to development or validation cohorts based on a temporal split. The PaO2/FiO2 ratio, respiratory compliance, and ventilatory ratio were assessed longitudinally throughout the first prone session. The subphenotypes were derived and validated using machine learning techniques. A K-means clustering implementation designed for joint trajectory analysis was utilized for the unsupervised classification of the development cohort. A random forest model was trained on the labeled development cohort and used to validate the subphenotypes in the validation cohort. Results 718 patients were included in a prospective cohort analysis. Of those, 504 were assigned to the development cohort and 214 to the validation cohort. Two distinct subphenotypes, labeled A and B, were identified. Subphenotype B had a lower PaO2/FiO2 response during the prone session, higher ventilatory ratio, and lower compliance than subphenotype A. Subphenotype B had a higher proportion of females (p