Clinical Implementation of Molecular Phenotypes of Critical Illness - ABSTRACT Acute respiratory distress syndrome (ARDS) and Sepsis are frequently encountered in critical care and lead to high morbidity and mortality. Despite decades of large clinical trials, other than antibiotics, no effective pharmacotherapies have been identified. Heterogeneity in both syndromes is increasingly recognized as a principal cause for these negative trials. To address this heterogeneity, we used latent class analysis (LCA) with clinical data and protein biomarkers to identify two distinct molecular phenotypes, called the Hyperinflammatory and Hypoinflammatory phenotypes. Independently, in multiple cohorts of ARDS and sepsis, the phenotypes have different outcomes and responses to therapies, offering a potential route to precision-based trials. Translating the phenotypes into the acute clinical setting has been challenging, as LCA models are incompatible for bedside use due to their complexity and protein biomarkers that discriminate the phenotypes don’t have rapid point-of-care (POC) assays. To address this, we have developed and validated a machine learning model, called the clinical classifier model (CCM), which can assign phenotypes using vital signs and laboratory values only. The primary objective of our proposal is to develop and validate an automated electronic health record (EHR)- embedded CCM system that classifies molecular phenotypes at the bedside, enabling future phenotype-enriched trials. Before such trials can be conducted, we need to broaden our understanding of the phenotypes and validate the CCM comprehensively. First, we will evaluate phenotype transition and the CCM’s performance longitudinally over time by using well validated biomarker-informed models to identify the molecular phenotypes in two previous ARDS trials (ROSE and FACTT) and one observational cohort of sepsis (MARS) at multiple timepoints (Aim 1A). We will then evaluate the performance of the CCM to identify these phenotypes at each timepoint (Aim 1B). Second, to enable bedside phenotype classification, we will develop an EHR-embedded CCM system that will automatically extract its component variables and feed it to the CCM to enable rapid classification of the molecular phenotypes (Aim 2A). We will silently run the EHR-embedded system in patients enrolled in the PRECCISE cohort and validate it prospectively against biologically derived phenotypes at multiple timepoints (Aim 2B). Third, comprehensive knowledge of the distinct epidemiology of phenotypes will enable more efficient future trial design. To do this, after using the CCM to classify molecular phenotypes in 78,000 critically ill patients with ARDS and Sepsis, we will use multistate modelling to map phenotype-specific hospital trajectories and outcomes (Aim 3A). We will also use this cohort to determine optimal trial design factors by conducting phenotype-specific target trials emulation of interventions previously associated with differential responses in the phenotypes (corticosteroids, fluids strategy, and PEEP strategy) (Aim 3B). Completion of our specific aims would be a game changer for precision medicine in critical care, as the EHR-embedded CCM system would harness the clinical potential of the molecular phenotypes, paving the way for phenotype-enriched trials.
