Development of a biomarker mortality risk model in acute respiratory distress syndrome

Acute respiratory distress syndrome (ARDS) is a devastatingly-intense inflammatory lung disorder characterized by severe respiratory failure requiring mechanical ventilation with a high mortality rate of 30–40% [1]. ARDS exhibits clinical and biological heterogeneity with respiratory and multi-organ system failure in response to diverse inciting stimuli. For example, in the USA, sepsis and pneumonia are the major inducers of ARDS, whereas in India, ARDS occurs in response to various tropical infections (malaria, miliary tuberculosis, dengue infections, etc.) [2]. The clinical and radiographic diagnostic criteria for ARDS are relatively imprecise [3]. Diagnostic uncertainty in ARDS further exacerbates disease heterogeneity and is a potential source of bias in conducting clinical trials. Consequently, the development of novel therapies in ARDS has been extremely challenging and contributes to the abysmal track record of Phase II/III ARDS clinical trials [4]. Scoring systems in critically ill patients such as the APACHE II score [5] or the lung injury severity score [6] successfully link to patient outcomes but fail to provide consistent and accurate predictive estimates of the risk of death in patient populations with a specific disease process. General severity scores lack specificity by failing to distinguish sepsis, ARDS, or acute kidney injury and are not significantly different between ARDS survivors and non-survivors [7]. Furthermore, general severity scores do not have pathophysiologic input and, therefore, are unlikely to guide personalized therapies. Attempts to characterize predictors of death in ARDS by developing a prognostic index [8] remain controversial and have not yet been replicated or validated. Thus, there is a compelling unmet medical need to identify clinical and/or biochemical disease-specific parameters that risk-stratify patients for both accurate prognostication and clinical trial purposes. Stratification of ARDS patients with reliable biomarkers predictive of mortality would optimize participant selection for clinical trial enrollment by focusing on those most likely to benefit from new clinical interventions [9, 10].

We and others have used “omic” approaches and translational systems biology methods to identify a panel of plasma protein biomarkers with diverse biological mechanisms that suggest a possible association with ARDS mortality [11‐22]. The objective of the current study was to utilize plasma protein biomarkers to derive a prognostic model for predicting ARDS patients least likely to survive to 28 days. We used latent class analysis and predictive analytics on eight plasma biomarkers collected from a well-phenotyped combined ARDS cohort. Biomarkers included cytokine-chemokines (interleukin-6 [IL-6], interleukin-8 [IL-8], interleukin-1 beta [IL-1B], and interleukin-1 receptor antagonist [IL-1RA]), dual-functioning cytozymes, i.e., cytokine/intracellular enzymes (macrophage migration inhibitory factor [MIF], nicotinamide phosphoribosyltransferase [eNAMPT]) and vascular injury markers (sphingosine 1-phosphate receptor 3 [S1PR3], angiopoietin-2 [Ang-2]). These analyses utilizing complementary analytical approaches, latent class analysis, classification and regression tree (CART) analysis, and rank aggregation were consistent and allowed the identification of six biomarkers that were most predictive of ARDS mortality.

We initially included eight ARDS-relevant plasma biomarkers derived from specific pathobiologic groups including: (i) inflammatory cytokine-chemokines (IL-6, IL-8, IL-1B, and IL-1RA), (ii) dual-functioning cytozymes, i.e., an intracellular enzyme that also functions as a cytokine when secreted (macrophage migration inhibitory factor, eNAMPT), and iii) vascular injury markers (S1PR3, Ang-2). Our findings indicate that a panel of six biomarkers potentially predict ARDS survival. In the latent class modeling we performed, a biomarker-based phenotype with higher plasma levels of Ang-2, MIF, IL-8, IL-1RA, IL-6, and eNAMPT was associated with significantly higher mortality (Fig. 1). The complementary CART analysis revealed that the maximum accuracy for predicting high mortality was achieved by four out of these six biomarkers (MIF, IL-8, IL-6, and eNAMPT) (Fig. 2). Finally, rank aggregation identified Ang-2, MIF, IL-8, IL-1RA, IL-6, and eNAMPT as the most important in contributing to mortality (Additional file 2: Table S6).

The pathogenesis of ARDS includes a combination of endothelial injury, epithelial injury, an intense inflammatory cascade, dysregulated coagulation, fibrosis, and apoptosis in response to diverse stimuli. Studies exploring the byproducts of acute dysregulation of various cellular pathways have generated more than 45 potential biomarkers [30, 31]. However, no single biomarker or clinical variable has demonstrated adequate prognostic or predictive ability to identify sub-phenotypes of ARDS [9, 10, 32, 33]. The ability to stratify ARDS patients by pathobiology and likelihood of treatment response would greatly enrich future clinical trials and enhance the ability to detect a treatment effect [34]. Sub-phenotyping/endotyping has been successfully accomplished in airways diseases such as asthma and COPD with important therapeutic implications [35, 36] and may exist within severe sepsis [4, 37, 38]. However, there is a paucity of data elucidating ARDS sub-phenotypes/endotypes. Recent studies [9, 39], utilizing two ARDSnet cohorts, identified two ARDS sub-phenotypes that markedly differed in natural history, clinical and biological characteristics, biomarker profiles, response to positive end-expiratory pressure (PEEP), and ventilator- and organ failure free days and in mortality. The hyperinflammatory ARDS sub-phenotype is characterized by a higher prevalence of sepsis and severe shock, high plasma levels of inflammatory biomarkers (IL-6, IL-8, etc.), greater vasopressor use, and metabolic acidosis. In contrast, the low inflammatory ARDS sub-phenotype exhibited less severe inflammation and shock. Surprisingly, the level of ARDS severity (PaO2/FiO2 ratio), renal or hepatic injury severity, or leukocytosis level failed to distinguish these two phenotypes [9]. The hyperinflammatory phenotype was associated with higher mortality, fewer ventilator-free and organ failure-free days, and altered responses to ventilator strategies when compared to the low inflammatory phenotype [40]. Importantly, no single clinical or biological variable was sufficient to identify the sub-phenotype including the severity of ARDS APACHE scores (PaO2/FiO2 ratio), severity of renal or hepatic failure, or leukocytosis, suggesting that phenotype membership was not merely a reflection of severity of illness as measured by traditional indices [40].

Our findings are strengthened by the use of three complementary approaches, including predictive analytics, to identify a common biomarker combination that is important in predicting ARDS survival. We also used a well-phenotyped cohort of ARDS patients, the majority of whom were enrolled as part of a large multicenter trial. Therefore, the participants represent a demographically diverse cohort of patients with ARDS strengthening the generalizability of our findings.

Our study has several limitations including the relative low observed mortality of 22% in the examined cohort compared to the expected ARDS mortality of 30–40%. This is likely due to the selection criteria we utilized where only patients surviving to day 7 were included in order to analyze both days 0 and 7 samples, i.e., patients who died before day 7 were not included. A second limitation of our work is that we had limited statistical power to detect significant differences in the predictive modeling algorithms. Finally, we recognize the lack of an external validation of our ARDS biomarker-based risk model in an independent cohort. Clearly, despite our utilization of a random sample of 80% of ARDS patients at D0 for the derivation and resampling technique on the remaining 20% of patients to demonstrate good test characteristics (Table 3), we plan to next evaluate our mortality panel an independent cohort to determine the reproducibility of these six biomarkers in predicting ARDS survival.

We have used complementary analytic approaches to identify six biomarkers (Ang-2, MIF, IL-8, IL-1RA, IL-6, eNAMPT) that show promise in predicting survival in ARDS. These biomarkers provide an important proof-of-concept that a combination of ARDS biomarkers can improve sub-phenotyping and enhance predictive and prognostic ability at the time of diagnosis. This information could be useful for guiding enrollment in clinical trials and represents an important step towards better patient stratification at the time of presentation to enhance the likelihood of a positive clinical trial in this vexing disorder.

Not applicable.