Abstract Rationale Identification of acute respiratory failure (ARF) subphenotypes through clustering of a broad panel of biomarker measurements from respiratory samples could have important prognostic and treatment implications. Methods We applied latent class analysis (LCA) to 25 protein biomarkers measured using electrochemiluminescent immunoassays from bronchoalveolar lavage fluid (BALF) collected from patients with ARF to identify alveolar subphenotypes (A-SPs) (Derivation Cohort: n = 466, 56% with ARDS). We derived a parsimonious classification model and applied it to an external cohort with biomarker measurements made with a similar platform (Validation Cohort: n = 86, 100% with ARDS). We then applied our model to a third cohort of patients with ARF who had BALF biomarkers measured using a different measurement platform (bead-based immunoassays) (Validation Cohort Two: n = 94, 89% with ARDS). We tested for associations between A-SPs and mortality using log-rank tests. Results A four-subphenotype LCA model had the best fit in the Derivation Cohort (VLMR p-value = 0.018) (Panel A). Mortality 28-days following the bronchoscopy was significantly different between the four subphenotypes (Panel B). A-SP2 was characterized by elevated PD-L1 and T cell chemokines (e.g. CCL13, CCL17, CXCL10), A-SP3 had relatively low levels of all 25 mediators, and A-SP4 was distinguished by high pyrogen and neutrophil-associated biomarker levels (Panel C). We used a gradient-boosted machine algorithm to develop a parsimonious classifier model and interpreted the estimates using Shapley additive explanation (SHAP) values. We carried forward the biomarkers with the highest SHAP values for each cluster to the Validation Cohort (A-SP1: IL-6; A-SP2: PD-L1; A-SP3: CSF3 (G-CSF); A-SP4: IL-8) to predict A-SP assignment. The distribution of P/F (or S/F) ratios and BALF %neutrophils between the four clusters was similar between the Derivation and Validation Cohorts. 28-day mortality was not statistically different between the four subphenotypes in the Validation Cohort (log-rank p = 0.21) (Panel B) or in Validation Cohort Two. Conclusions Patients with ARF that have similar degrees of hypoxemia and illness severity can exhibit highly variable alveolar molecular signatures that have distinct associations with outcomes. Globally reduced alveolar immune activation (A-SP3) and higher levels of neutrophil-associated (A-SP4) mediators are associated with worse outcomes, whereas checkpoint proteins and T cell chemokines (A-SP2) are associated with better outcomes. We speculate associations between these subphenotypes and mortality did not replicate in much smaller validation cohorts due to power and batch effects. These organ-specific molecular subphenotypes likely provide complementary biologic and clinical risk-stratification information to existing critical illness subphenotypes. This abstract is funded by: NIH
Morrell et al. (Fri,) studied this question.