Abstract Rationale The Hypoinflammatory phenotype of critical illness, beyond its contrast to the Hyperinflammatory phenotype, remains poorly characterized, despite comprising 60-75% of ARDS and sepsis patients. We aimed to identify underlying heterogeneity in the Hypoinflammatory phenotype applying unsupervised clustering and modeling methods to multi-modal data incorporating either protein biomarkers and clinical variables or RNA sequencing (RNA-Seq). Methods We analyzed three cohorts of critically ill adult sepsis patients with Hypoinflammatory phenotype previously assigned by latent class analysis (LCA): Early Assessment of Renal and Lung Injury (EARLI; Hypo=530), the Crystalloid Liberal or Vasopressor Early Resuscitation in Sepsis trial (CLOVERS; Hypo=1195), and Molecular Diagnosis and Risk Stratification of Sepsis (MARS; Hypo=895). In ascending order, these cohorts had increasing protein biomarker availability. Analyses were performed independently in each cohort using a combination of clinical variables (laboratory values and vital signs) and protein biomarkers. Data pre-processing involved excluding variables with either high missingness (30%) or a pair of high correlated variable (r 0.5). First, we performed principal component analysis (PCA) to visualize potential subgroups/variance with distinct features in lower dimensions. Next, LCA was used to identify clusters within the Hypoinflammatory group. We used a model based unsupervised clustering method, FSCseq (PMCID: PMC8386505), on whole blood RNA-Seq data from 170 Hypoinflammatory patients in EARLI. Results Our data processing pipeline identified: 17 features (4 protein biomarkers, and 13 clinical features) in EARLI, 23 features in CLOVERS (9 biomarkers and 14 clinical variables), and 25 features in MARS (13 biomarkers and 12 clinical variables). In all three cohorts, PCA using the top two principal components identified low variance within the data (Panel A, B, C). LCA models in CLOVERS and MARS showed decreasing Bayesian Information Criterion (BIC) values with increasing K (number of clusters), but low entropy indicated poorly differentiated classes (Panel B, C). In EARLI LCA, a two-cluster solution had entropy above 0.85; however, this was a small manifest class of patients with the highest creatinine values. RNA-Seq clustering in EARLI suggested a possible three-cluster solution as best fitting, but the clusters showed limited stability, with low Jaccard index scores over 100 random samples (Median (IQR)=0.46 (0.4-0.49)). Conclusions Our multi-modal approach investigating subsets in the Hypoinflammatory phenotype failed to uncover consistent heterogeneity, suggesting that the Hypoinflammatory phenotype lacks distinctive features in the dimensions that we investigated. Broader studies employing additional data dimensions such as proteomics and metabolomics with larger patient populations may be required to better understand this population. This abstract is funded by: R35HL177135 (CSC), R01HL173531, R35GM142992 (PS)
Tripathi et al. (Fri,) studied this question.