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 < 0.001) and lung disease (p = 0.005), higher baseline SAPS III (p = 0.002) and SOFA (p < 0.001) scores, and lower body mass index (p = 0.05). Subphenotype B had also higher levels of the pro-inflammatory biomarker IL-6 (p = 0.017). Subphenotype B was independently associated with an increased risk of 60-day mortality (OR 1.89, 95% CI 1.51-2.36). Additionally, Subphenotype B was associated with a lower number of ventilator-free days on day 28 (p < 0.001) and a lower hospital length of stay (p < 0.001). The subphenotypes were reproducible in the validation cohort.

Conclusion: Our study successfully identified and validated two distinct subphenotypes of COVID-19 ARDS based on key respiratory parameters. The findings suggest potential implications for better patient stratification, risk assessment, and treatment personalization. Future research is warranted to explore the utility of these novel subphenotypes for guiding targeted therapeutic strategies in COVID-19 ARDS.

Keywords: ARDS; Acute respiratory distress syndrome; Covid-19; Machine learning; Mechanical ventilation; Subphenotypes.

Fig. 1

Trajectories of the clustering variables in subphenotypes A (green) and B (blue) in the development cohort. a PaO₂/FiO₂; b Respiratory compliance; c Ventilatory ratio. The lines represent the mean, and the bands represent the 95% confidence interval

Fig. 2

Differences in the respiratory variables between the subphenotypes. The heatmap represents the differences (subphenotype A minus B) of respiratory parameters (rows) at each of the measured time points during the prone session (columns). All variables were standardized so that the means correspond to 0 and the standard deviation correspond to 1

Fig. 3

Comparison of thromboinflammation blood markers between the subphenotypes. The p-values were calculated using the Wilcoxon test

Fig. 4

Kaplan–Meier plot of the probability of survival until day 60. The bands represent the 95% confidence intervals for point estimates of the survival curves. The p-values were calculated using the log-rank test