Statistical analysis
We first selected a set of cytokines most closely related to the viral immune response based on our previous analyses and published literature. These included IFN- γ, IL-1β, IL-6, IL-8, IL-10, IL-12, MIP-1β, MCP-1, IP-10 (10, 11). Given the temporal variation for both the cytokine concentrations and semi-quantitative SARS-CoV-2 RT-PCR, we standardized and rescaled the values from 0 to 1 per week. The BAL SARS-CoV-2 RT-PCR Ct values and the prespecified set of cytokines in BAL and plasma were included in the clustering analyses. We used a hierarchical clustering on principal components approach. This was employed separately for plasma and BAL with complete linkage and Euclidean distances predefining the number of clusters as 2 and 3. The number of clusters was a priori specified given the known partitioning into 2 clusters of subjects with COVID-19 ARDS and our hypothesis on the role of semi-quantitative SARS-CoV-2 RT-PCR in differentiating a unique cluster. Hierarchical clustering is a common unsupervised machine learning technique that aims to group similar subjects by measuring the distance between them. The variables of interest (i.e. SARS-CoV-2 Ct values and cytokine concentrations) define a multidimensional space where the distances between subjects are measured. We analyzed the variables explaining the variance between the clusters and containing the most information in the data by using principal component analysis. This was important as the pro-inflammatory cytokines are correlated with each other. We retained the first 3 principal components and the explanatory variables were represented graphically as component loadings. The clinical, laboratory data, immune characteristics and outcomes of the clusters were listed. We compared the BAL and the plasma clusters by using the Dunn index. Finally, we juxtaposed the semi-quantitative SARS-CoV-2 RT-PCR and cytokine concentrations per week for both plasma and BAL. Analyses were performed in R 4.0.3 (R Core Team. A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria).