Phenotyping in ARDS: State of the art and clinical implications

Abstract

Purpose of the review:

Decades of research in ARDS have led to few interventions that impact clinical outcomes. The pandemic of patients with ARDS due to the novel SARS-CoV-2 infection has stressed the need for more effective therapies in ARDS. Phenotyping may enable successful trials and precision therapeutics in this patient population.

Recent findings:

Clinical phenotypes that group patients by shared etiology, time-course, or radiographic presentation are of prognostic value but their use is limited by misclassification. Physiological phenotypes including the P/F ratio, ventilatory ratio, and dead space fraction predict poor outcomes but can rapidly change, making them unstable over time. Biologic phenotypes have prognostic value with composite clinical and biomarker sub-phenotypes additionally impacting treatment response but are yet to be prospectively validated.

Summary:

Though much progress has been made in ARDS phenotyping, implementation of precision medicine practices will depend on conducting phenotype-aware trials using rapid point of care assays or machine learning algorithms. Omics studies will enhance our understanding of biologic determinants of clinical outcomes in ARDS sub-phenotypes. Whether biologic ARDS sub-phenotypes are specific to this syndrome or rather more broadly identify endotypes of critical illness remains to be determined.

Introduction:

Acute Respiratory Distress Syndrome (ARDS) has long suffered from a paucity of practice-changing discoveries despite rigorous research, owing to its wide range of triggers(1), broad definition(2), and variable outcomes. The surge of patients with ARDS beginning in 2020 from the novel SARS-CoV-2 infection overwhelmed healthcare systems and highlighted the need for more effective therapies in ARDS. Phenotyping this heterogeneous syndrome into more homogeneous subgroups may lead to more success in identifying effective therapies.

Since ARDS was defined in 1967, researchers have endeavored to understand its heterogeneity, fueled of late by advancements in phenotyping in other fields. In oncology, molecular phenotyping of melanoma led to the introduction of checkpoint inhibitors and therapies targeting BRAF V600 mutation that have significantly prolonged survival(3). Breast cancer treatment approaches and outcomes are vastly different based on hormone receptor and gene mutation phenotyping(4). In pulmonology, biomarker-based phenotyping has led to targeted therapies for patients with Th2 dependent inflammation and eosinophilic asthma(5–9). These successes have spurred the search for treatable phenotypes in syndromes of critical illness, namely sepsis and ARDS.

The COVID-19 pandemic led to large numbers of patients with a uniform trigger for ARDS. Clinical trials in patients with COVID-19-related ARDS (CARDS) met with more success in identifying effective therapies than decades of large, well-designed randomized trials had in “classical” ARDS. For example, numerous studies on the role of steroids in ARDS arrived at varying conclusions(10–12), yet steroids in patients with CARDS have more consistently demonstrated a mortality benefit(13, 14). The contrast between these findings in COVID-19 and classical ARDS studies suggest that a steroid-responsive subgroup likely exists within groups of patients with classical ARDS and needs characterization. Even within CARDS, randomized trials studying the same drugs have met with mixed results, likely in part due to heterogeneous biologic response to ARDS and/or differing management strategies for patients with complex critical illness(14–23). Nonetheless, CARDS trials demonstrate that we are more likely to find successful ARDS therapies by selecting sub-groups of patients more likely to respond to a given treatment (predictive enrichment) and those at higher risk of poor outcomes (prognostic enrichment).

In this paper, we aim to briefly review current concepts in phenotyping ARDS, highlight some inherent challenges to phenotyping, and identify key directions toward which the field is headed in the coming decade.

Phenotyping: The Present

Clinical Phenotyping:

Clinical phenotyping in ARDS subdivides patients based on either a shared etiology, time-course, or radiographic presentation of ARDS (Table 1). Evidence suggests that etiologic sub-phenotypes carry different prognoses and, in some cases, different treatment responses. The most prominent recent example of phenotyping based on a shared etiology is CARDS. Outside of CARDS, however, ascertaining an ARDS trigger can be difficult, and sometimes multiple etiologies are at play. Nevertheless, data suggests that prognosis and attributable mortality of ARDS differ based on etiology(24, 25). For example, ARDS induced by trauma carries a lower mortality rate than non-trauma related ARDS(26). ARDS resulting from direct injury to the lungs portends a better prognosis than ARDS due to indirect insults (e.g. non-pulmonary sepsis)(27). As of yet, aside from COVID-19, there is no compelling evidence that ARDS due to different etiologies responds differently to therapies.

Table 1.

Overview of selected ARDS phenotypes, clinical utility, and main limitations.

Phenotypes	Sub-phenotypes	Established clinical use		Main limitations
		Prognostic	Therapeutic
Clinical
	Etiology	X	X (COVID-19 only)	High risk of misclassification
	Sepsis
	Trauma
	Direct lung injury
	Time course	X
	Early vs late
	Rapidly resolving
	Radiographic	X
	RALE score
	Murray lung injury score
	Focal vs non-focal
Physiological
	P/F ratio	X	X	Rapidly changing variables
	Dead space fraction	X
	Ventilatory ratio	X
	Driving pressure	X
Biological
	Protein biomarkers	X		Identifying interventions is complicated by partial understanding of complex biology
	Epithelial injury
	Endothelial injury
	Proinflammatory cytokines
	Disordered coagulation
	Reactive vs uninflamed
	Composite clinical/protein	X	X
	Hyperinflammatory vs
	Hypoinflammatory
	Metabolomics	X
	High vs low metabolite pulmonary edema fluid

The time-course of ARDS also has prognostic value. Studies have shown that late onset of ARDS (more than 48 hours after ICU admission) is associated with higher mortality rates than earlier onset(28, 29). Among those who present with ARDS, one study found that 63% with moderate to severe disease based on P/F ratio paradoxically have a rapidly resolving phenotype and a better prognosis(30). Thus, trials recruiting patients with moderate to severe ARDS upon hospital presentation may inadvertently be enriched with this rapidly resolving clinical sub-phenotype.

Radiographic findings provide another means for prognostic and predictive trial enrichment in ARDS. The RALE score, which systematically quantifies the extent and density of alveolar infiltrates on plain films, predicts 28-day mortality with an AUC of 0.82(31). Similarly, the Murray Lung Score, which incorporates radiographic findings, was used as one approach to prognostic enrichment in the CESAR trial for ECMO(32, 33). Radiographic findings were used for predictive enrichment in the LIVE trial, an innovative study across 20 ICUs in France(34). The investigators randomized patients to receive either standard lung-protective ventilation, or a personalized mechanical ventilation strategy based on the presence of focal or non-focal radiographic findings. In the personalized arm, the patients with focal ARDS received higher tidal volumes (8 mL/kg) and low PEEP, while those with non-focal disease received a lower tidal volume (6 mL/kg), recruitment maneuvers, and high PEEP. The trial found no significant differences in 90-day mortality, its primary outcome of interest. However, a post-hoc review led to the discovery that 21% of patients were radiographically misclassified. Accounting for this misclassification, the investigators found that a ventilator strategy misaligned with radiographic findings significantly increased 90-day mortality. The LIVE trial highlights both the perils of “one size fits all” therapies in ARDS and misclassification inherent in clinical phenotyping.

Physiological phenotyping:

Physiologic phenotyping separates groups of patients based on severity of lung impairment (Table 1). The Berlin Criteria introduced the most commonly applied physiologic sub-phenotypes of mild, moderate and severe ARDS, defined using the ratio of partial pressure of oxygen in arterial blood (PaO2) to the fraction of inspired oxygen (FiO2), the P/F ratio. These three categories of disease severity were associated with escalating mortality rates, and many recent trials have used P/F sub-phenotypes for prognostic enrichment by enrolling only those with moderate to severe ARDS (P/F <150). Despite the prevalence of its use, the P/F ratio is a mediocre predictor of mortality with an AUC of only 0.577 in one analysis(2). Other physiologic sub-phenotypes that predict poor outcomes include dead space fraction, ventilatory ratio, and driving pressure(35–37). The main limitation of physiological phenotyping is that variables can rapidly change, creating unstable sub-phenotypes that in some cases may be challenging to study in trial settings.

Biological phenotyping:

Drugs targeting biologic processes to reverse lung injury or enhance lung repair in ARDS have not lowered mortality, likely in part due to heterogeneity of the host response to ARDS. Biological phenotyping seeks to identify subgroups with a similar host response to ARDS to elucidate its pathophysiology and allow for prognostic and predictive trial enrichment (Table 1).

The largest body of work in this arena involves analyses of plasma protein biomarkers, several of which have demonstrated diagnostic and prognostic value in ARDS. These include biomarkers of alveolar epithelial injury such as soluble receptor for advanced glycation end products (sRAGE) and surfactant protein-D (SP-D)(38–47); endothelial injury such as angiopoietin-2 (Ang-2), von Willebrand factor (vWF), and intercellular adhesion molecule (CAM)-1(48–51); proinflammatory cytokines such as soluble tumor necrosis factor receptor I (sTNFr-1), interleukin (IL)-6, and IL-8(49, 50, 52–56); and disordered coagulation such as plasminogen activator inhibitor-1 (PAI-1) and protein C(57). Combinations of these biomarkers perform better in diagnosing and risk stratifying ARDS than each biomarker alone(44, 52, 58, 59).

Using latent class analysis (LCA), a retrospective study of clinical and protein biomarker data from the landmark ARMA and ALVEOLI trials of ARDS identified two sub-phenotypes, designated as “hyperinflammatory” and “hypoinflammatory”, associated with distinct clinical outcomes(60–62). Retrospective analyses of data from five ARDS trials (ARMA(61), ALVEOLI(62), FACTT(63), SAILS(64), HARP-2(65)) as well as analyses of ARDS patients from two prospective observational cohorts, altogether comprising over 4000 patients, consistently demonstrate that patients with the “hyperinflammatory” sub-phenotype experience higher mortality rates than those with the “hypoinflammatory” sub-phenotype(66–69). Beyond prognostic utility, these sub-phenotypes additionally seem to have different responses to therapies such as PEEP, fluid strategy, and simvastatin(60, 66, 67) in secondary analyses of completed trials. A separate study using cluster analysis of protein biomarker data from a large cohort of patients across two ICUs in the Netherlands identified the presence of two molecular sub-phenotypes of ARDS designated as “reactive” and “uninflamed”(70). Specifically, the “reactive” sub-phenotype had higher levels of IL-6, Ang-1 and 2, PAI-1, and interferon-gamma levels and experienced worse outcomes than the “uninflamed” sub-phenotype.

Other types of biomarkers including RNA, metabolites, lipids, and extracellular vesicles hold potential for further untangling the complex biological phenotypes within ARDS. Transcriptomic analyses of peripheral blood leukocytes from the Netherland ICU cohort of patients identified upregulation of pathways of oxidative phosphorylation and mitochondrial dysfunction in the “reactive” sub-phenotype of ARDS(71). More recently in patients with CARDS, transcriptomic analyses of tracheal aspirates demonstrated reduced pro-inflammatory gene expression compared to classical ARDS(72). The dysregulated host response in patients with CARDS was characterized by expression of genes associated with non-canonical roles in inflammation, potentially explaining why these patients benefit from steroids. Using metabolomic analyses, a subgroup of patients from a small cohort with ARDS were found to have a distinct metabolic profile in pulmonary edema fluid samples associated with higher mortality(73).

Phenotyping: The Challenges

While phenotyping holds great promise for future research trials in ARDS, the field faces several considerable challenges that have limited clinical implementation. One noteworthy challenge is disagreement over what constitutes sufficient data to recommend a change in clinical practice(74). Here, the COVID-19 pandemic serves as an illustrative example. Early in the pandemic, researchers published observations of two apparent sub-phenotypes of CARDS with divergent lung compliance, radiographic, and physiological features(75). The “L” sub-phenotype was defined by having low lung elastance (high compliance) and was postulated to predominate in CARDS. The less common “H” sub-phenotype was defined by having high lung elastance and dense airspace filling on CT. The authors theorized that the high mortality rates early on in CARDS may have been partially related to inappropriate use of lung protective ventilation, one of the two interventions ever demonstrated to have a mortality benefit in ARDS(61), and recommended that the L sub-phenotype be treated with a different ventilator management strategy. Published in a prominent journal and cited over 600 times, the article reached a wide audience. However, numerous subsequent studies failed to identify evidence in support of the “L” and “H” sub-phenotype model, instead indicating that lung compliance in CARDS follows a normal distribution and is generally low, similar to its non-COVID counterpart(76, 77). Ideally, phenotyping should be data-driven and externally validated before clinical use.

Even phenotypes derived from rigorous, data-rich, and large studies still have limitations and must be interpreted with caution. For instance, as mentioned with the LIVE trial, clinical phenotypes suffer from high rates of misclassification(34). Despite retrospective reproducibility across multiple cohorts of ARDS, latent biologic phenotypes still require prospective validation before they can be of clinical utility. Lastly, detecting which phenotypes amongst the many that have and will be identified are clinically relevant and impact treatment response remains a major challenge. Adding to this complexity is our limited ability to identify successful interventions for complex biologic phenotypes. For instance, a retrospective subgroup analysis of a randomized controlled trial on the use of recombinant IL-1 receptor antagonist in patients with sepsis showed a paradoxical treatment benefit in the subset of patients with higher baseline levels of IL-1 receptor antagonists(78). This study illustrates that our understanding of pathobiology in critical illness remains rudimentary and stresses the need for more studies in preclinical models.

Phenotyping: The Potential

Advances in phenotyping and the COVID-19 pandemic have escalated the pace of phenotyping in ARDS. Implementation of precision medicine practices in ARDS will depend upon the research community conducting phenotype-aware trials, elucidating pathophysiologic pathways of lung injury in various forms of ARDS, and translating such discoveries to personalized therapies (Figure 1).

Figure 1.

Schema for real-time phenotyping in ARDS.

EHR: electronic health record. POC: point of care.

Phenotype-aware trials and cohort studies enrolling patients with predictive and prognostic enrichment in mind are necessary to prospectively validate sub-phenotypes of ARDS. However, lack of point of care assays for rapid biologic phenotyping presents a significant barrier to conducting such studies. To circumvent this issue, Sinha and colleagues proposed a machine learning algorithm tool that uses readily available laboratory and clinical data to phenotype patients into “hyperinflammatory” and “hypoinflammatory” ARDS on admission to the ICU and correlates well with the gold standard LCA based biomarker sub-phenotypes(79). Furthermore, this algorithm might be incorporated into existing electronic health records for ease of use. Investigators have also found that a parsimonious model using three plasma biomarkers can classify patients with ARDS into the two inflammatory sub-phenotypes with high accuracy(80). Based on these findings, a point of care assay for rapid analysis of plasma IL-6 and soluble TNFr1 levels has been developed and is being studied in the PHIND trial, the first prospective cohort study of ARDS patients undergoing biologic phenotyping upon study entry(81). The PHIND trial also aims to prospectively validate these biologic sub-phenotypes and test their stability over time. The future holds promise that “real time” phenotyping in patients with ARDS is imminent.

More studies on genomic, transcriptomic, and metabolomic phenotyping in ARDS are underway and will enhance our understanding of the biologic determinants of clinical outcomes in sub-phenotypes of ARDS. A recent multi-omics study by Overmyer and colleagues in patients admitted with moderate to severe respiratory issues with and without COVID-19 found 219 molecules strongly associated with COVID-19 status and severity and pointed to dysregulation of biologic processes involving lipid transport, coagulation, endotheliopathy, and neutrophil degranulation(82). Studies on the local versus systemic host response in ARDS using tracheal aspirates, bronchoalveolar lavage fluid, and fluid from heat-moisture exchange filters will advance our understanding of ARDS pathophysiology and offer novel therapeutic targets. Genome wide association studies hold the promise of identifying novel ARDS biology, though are challenged by the syndromic definition of ARDS and the difficulties of identifying genetic control groups. Ultimately, the field of phenotyping in ARDS is moving towards deep phenotyping wherein multiple types of data using a variety of technologies lead to whole-body physiological profiling, furthering our understanding of host response mechanisms, and enabling personalized therapies in ARDS(83, 84).

The question remains as to whether ARDS phenotypes are specific to this syndrome or rather more broadly identify endotypes of critical illness. A recent study applied classifiers for cluster and LCA-derived biologic sub-phenotypes of ARDS to a population of mechanically ventilated patients without ARDS and found that the “reactive” and “hyperinflammatory” sub-phenotypes were associated with higher probability of mortality even in patients without ARDS(85). Another study sub-phenotyping patients at risk for developing ARDS found that a distinct LCA defined baseline “hyperinflammatory” sub-phenotype was associated with higher mortality and prolonged mechanical ventilation(86). Similar studies on ICU patients with and without ARDS are necessary to determine sub-phenotype specificity. One can imagine a future in which our approach to treatment of critically ill patients revolves around biologically identified treatable traits rather than the current syndrome-based paradigm.

Conclusions:

Phenotyping ARDS has identified subgroups of patients with distinct outcomes. The composite LCA defined “hyperinflammatory” and “hypoinflammatory” sub-phenotypes have additionally demonstrated differential responses to ARDS therapies, albeit in post-hoc analyses. Progress in the field is challenged by insufficient data, lack of prospective validation, difficulties identifying clinically relevant sub-phenotypes that impact treatment outcomes, and translating discoveries into effective therapies. Real-time sub-phenotyping using novel assays and machine learning algorithms will enable phenotype-aware trials that hold the promise of identifying successful ARDS therapies. Phenotyping ARDS and critical illness more broadly may lead to a paradigm shift away from syndrome-based definitions towards treatable traits.

Key points:

• Phenotyping ARDS using clinical, physiologic, and biologic data has identified subgroups of patients with distinct clinical outcomes with the hyperinflammatory and hypoinflammatory biologic subphenotypes demonstrating differential treatment response in retrospective analyses of randomized controlled trials and cohort studies.

• Researchers and clinicians must recognize the limitations of current phenotypes, including phenotypes derived from insufficient data, misclassification of clinical phenotypes, instability of physiologic phenotypes over time, and lack of prospective validation of biologic phenotypes.

• Precision medicine practices in ARDS depends upon the research community conducting phenotype-aware trials, elucidating pathophysiologic pathways of lung injury in various forms of ARDS, and translating such discoveries to personalized therapies.