Host-Response Subphenotypic Classication with A Parsimonious Model Offers Prognostic Information in Patients with Acute Respiratory Failure: A Prospective Cohort Study.

Background: Recent research in patients with ARDS has consistently shown the presence of two distinct subphenotypes of host-responses (hyper- and hypo-inammatory) with markedly different outcomes and responses to therapies. However, inherent uncertainty in reaching the diagnosis of ARDS creates considerable biological and clinical overlap with other broadly-dened syndromes of acute respiratory failure, such as patients with risk factors (e.g. sepsis or pneumonia) for ARDS (at-risk for ARDS [ARFA]) or patients with decompensated congestive heart failure (CHF). Limited data are available for the presence of subphenotypes in such broader critically-ill populations. Methods: We enrolled mechanically-ventilated patients with acute respiratory failure (ARDS, ARFA, and CHF) and measured 11 plasma biomarkers at baseline. We applied latent class analysis (LCA) methods to determine optimal subphenotypic classications in this inclusive patient cohort by considering clinical variables and biomarkers. We then derived a parsimonious logistic regression model for subphenotype predictions and compared clinical outcomes between subphenotypes. Results: We included 334 patients (123 [37%] ARDS, 177 [53%] ARFA, 34 [10%] CHF) in a derivation cohort and 36 patients in a temporally-independent validation cohort. A two-class LCA model was found to be optimal, classifying 29% of patients in the hyper-inammatory subphenotype, consistent with prior ndings. A 4-variable parsimonious model (angiopoietin-2, soluble tumor necrosis factor receptor-1, procalcitonin and bicarbonate) for subphenotype prediction offered excellent classication (area under the curve = 0.98) compared to LCA classications. For both LCA- and regression model classications, hyper-inammatory patients had higher severity of illness by Sequential Organ Failure Assessment scores, fewer ventilator-free days and higher 30- and 90-day mortality (all p<0.01) compared to the hypoinammatory group. Subphenotype predictions in the validation cohort revealed consistent trends for clinical outcomes and higher levels of inammatory biomarkers in the hyper-inammatory group (22%). Conclusions: Host-response subphenotypes are observable in broader and heterogeneous patient populations beyond just patients with ARDS, and subphenotypic classications offer prognostic information on clinical outcomes. Accurate subphenotyping is possible with the use of a simple predictive model to improve clinical applicability.

subphenotypes were initially discovered from independent, unsupervised examinations of randomized clinical trial data with latent class analysis (LCA) of both clinical and biomarker variables. However, LCA and other unsupervised clustering approaches are laborious and cannot be implemented clinically. For that reason, efforts have been made to derive simpler models for subphenotype prediction in patients with ARDS, involving either predictive models with biomarkers (6) or machine-learning classi ers based on clinical variables alone (7).
We recently discovered that the two distinct subphenotypes of host-responses are also present in patients with risk factors for ARDS (e.g. pneumonia or sepsis) who do not meet the clinical criteria of ARDS diagnosis (i.e. bilateral opacities on chest-radiographs and severe hypoxemia) (8). In such patients at-risk for ARDS (ARFA), classi cation to a hyper-in ammatory subphenotype by LCA was consistently associated with higher severity of illness, persistently elevated biomarkers of host injury and in ammation, and worse clinical outcomes. These observations suggest that common pathways of systemic in ammatory injury are likely present in broader critically-ill populations, beyond the subset of patients with ARDS (~ 10% of all ICU admissions) (2). Additionally, inherent subjectivity in making the diagnosis of ARDS, primarily due to clinicians' disagreement on radiographic criteria or exclusion of cardiac failure in the Berlin de nition, is likely to generate considerable overlap among patient subgroups diagnosed as ARDS, ARFA or with cardiogenic pulmonary edema from decompensated congestive heart failure (CHF) (9)(10)(11)(12). Therefore, subphenotyping efforts in a more inclusive diagnostic framework that considers broader patient populations with acute respiratory failure (ARDS, ARFA or CHF) may uncover distinct subgroups of patients that may bene t from targeted enrollment in clinical trials. We sought to determine whether hyper-and hypo-in ammatory subphenotypes could be detected in a heterogeneous critically-ill population with acute hypoxemic respiratory failure, and whether a parsimonious predictive model with a small number of relevant variables could be derived for future clinical or investigative application.

Methods:
Extensive methods are provided in Additional le 1.

Clinical cohort:
From October 2011 to August 2019 we prospectively enrolled a convenience sample of consecutively admitted mechanically-ventilated patients with acute respiratory failure in Medical Intensive Care Units (ICUs) at the University of Pittsburgh Medical Center to the Pittsburgh Acute Lung Injury Registry (ALIR) and Biospecimen Repository (8,(13)(14)(15). We excluded patients unable to provide informed consent or mechanical ventilation for > 72 hours prior to enrollment. Written informed consent was provided by all participants or their surrogates. The study was approved by the University of Pittsburgh Institutional Review Board (protocol STUDY19050099). We recorded baseline demographics, comorbidities, mechanical ventilation and laboratory variables, and calculated sequential organ failure assessment (SOFA) scores.

Clinical group classi cations:
A consensus committee retrospectively reviewed all available clinical and radiographic data without knowledge of biomarkers values and classi ed subjects into distinct clinical categories of acute respiratory failure: a) ARDS per Berlin criteria (10), b) (ARFA) based on presence of an identi able lung injury risk factor upon enrollment but not ful lling ARDS criteria, c) cardiogenic pulmonary edema from CHF, d) acute on chronic respiratory failure (e.g. acute exacerbation of chronic obstructive pulmonary disease [COPD]), e) intubation for airway protection, and f) "other" category, including cases for which the committee could not reach consensus for clinical classi cation into any of the categories above.

Outcomes:
Primary outcomes included ventilator-free days (VFD) and 30-and 90-day mortality. Patients were also followed prospectively for incidence of shock within the rst week of enrollment (de ned as need for vasopressor agents), acute kidney injury (AKI), time-to-liberation from mechanical ventilation, and ICU length of stay.

Subphenotypic classi cations and statistical analyses
For subphenotypic classi cation with LCA models rst and then derivation of a parsimonious predictive model, we considered patients from the ARDS, ARFA, and CHF groups in our primary analyses. Our selection of these three primary clinical groups was made on the basis of biological similarity (similar distribution of subphenotypes between ARDS and ARFA patients in our previous study) (8), as well as due to the higher clinical relevance of risk strati cation and treatment selection in these critically-ill patients, compared to patients intubated for airway protection or acute exacerbation of COPD for example.
For our primary analysis, we divided our cohort of patients with ARDS, ARFA and CHF into two temporally independent datasets: a derivation dataset of 334 patients enrolled up to February 2019 and a validation dataset including 36 patients enrolled from March-August 2019. In secondary analyses, we also considered the 110 patients with acute on chronic repiratory failure, intubation for airway protection and "other" category. Data for 235/334 (70%) of the patients in the derivation dataset had been previously utilized for application of LCA models separately in patients with ARDS and in patients ARFA (8).
We performed subphenotypic classi cations by applying LCA in the derivation dataset. First, we estimated the optimal number of classes that best t our patient cohort, as subphenotyping analysis has not yet been performed in such an inclusive patient population with acute respiratory failure. We considered a total of 35 baseline clinical and biomarker variables similar to those in LCA models previously used in ARDS subphenotyping trials (Additional File 2), without consideration of the clinical outcomes. The continuous variables were graphically examined by plotting their standardized values to a common z-scale (mean of 0, standard deviation of 1). We selected only variables that were found to be discriminatory (p≥0.1) for consideration in development of LCA and parsimonious models. Categorical variables were also compared graphically via Fisher exact tests with non-discriminatory variables being removed. We also applied the LCA to ARDS, ARFA, and CHF groups separately to con rm best tness of the model in each individual subgroup.
We then developed a parsimonious logistic regression model based on a best subsets generalized linear model approach using Bayesian Information Criteria (BIC) (30). We subsequently applied subphenotype classi cations provided from the parsimonious model in a) the derivation cohort and b) validation cohort of patients with ARDS, ARFA and CHF, and c) in the secondary analysis of patients with other forms of respiratory failure. Comparisons between hyper-in ammatory and hypo-in ammatory subphenotypes were obtained from Wilcoxon test for continuous variables and Fisher's test for categorical variables. Kaplan-Meier curves and Cox-proportional hazard models were created for survival and time-to-liberation. We examined p-values for bootstrapped parametric likelihood ratio tests to select the nal number of classes. We tested the proportional hazard assumption in all models. We performed LCA in Mplus 8.3 and all other analyses in R v.3.5.1 (31,32 (Fig. 1). ARDS patients had the highest frequency of pneumonia, higher peak inspiratory pressures, and worse hypoxemia, whereas ARFA patients had higher incidence of aspiration and extra-pulmonary sepsis compared to the other groups (p < 0.01) (Additional File 3).

Subphenotypic classi cations and baseline variables:
Utilizing all clinical and biomarker variables (n = 35), a two-class LCA model offered optimal t (p < 0.001). Analysis of the LCA after exclusion of non-discriminatory variables (n = 11) improved the likelihood ratio of the model and also demonstrated excellent agreement with the model prior to elimination of these variables (Fig. 2). Twenty-nine percent of patients were assigned to the hyper- Table 1 Comparisons of baseline variables between hyper-in ammatory and hypo-in ammatory patients by latent class analysis model.  (Fig. 3). To con rm a two-class model offered best t within each clinical group, the LCA was applied to the ARDS, ARFA, and CHF subgroups individually and was con rmed to offer optimal t. Thus, LCA models con rmed the presence of hyper-and hypo-in ammatory subphenotypes in ARDS and ARFA groups and established subphenotypic presence in patients with acute respiratory failure secondary to CHF.
In derivation of the parsimonious model, 11 variables were found to be the most discriminatory between the two subphenotypes ( Fig. 2) baseline clinical variables and biomarkers between subphenotypes were very similar to those of the LCA models (Fig. 4, Additional le 4).
Secondary analysis in the remaining combined clinical groups (acute on chronic respiratory failure, airway protection, and "other"; n = 110) by the parsimonious predictive model demonstrated a very low prevalence of the hyper-in ammatory subphenotype (n = 8, 7%), which was signi cantly lower compared to the ARDS, ARFA and CHF groups combined (p < 0.01; Fig. 3).
Severity of illness and clinical outcomes by subphenotypes: For LCA-derived subphenotypes, hyper-in ammatory patients had higher SOFA scores, vasopressor usage, AKI, 30-and 90-day mortality and fewer VFDs (p < 0.01) ( Table 2). Similar associations with severity of illness and outcomes were observed for the parsimonious model predicted subphenotypes. Hyperin ammatory patients had worse 30-day survival and longer times to liberation from mechanical ventilation for both LCA and parsimonious model classi cations (Fig. 5). Discussion: In a prospective, observational cohort of mechanically-ventilated patients with acute respiratory failure, we demonstrated that the subphenotypic distinction described for patients with ARDS applies in a broader and more heterogeneous critically-ill population (5,33). While previous subphenotyping models have mainly been derived from secondary analyses of clinical trial populations in patients with ARDS, we revealed the presence of such subphenotypes in an observational, inclusive cohort study. We generated a simpli ed LCA model utilizing only discriminatory variables at baseline, and then de ned a parsimonious 4-variable regression model consisting of well-described biomarkers in ARDS and other critical illness syndromes. Consistent with prior studies, we found that hyper-in ammatory patients at baseline exhibited higher severity of illness, elevated plasma biomarkers beyond the ones used for the subphenotypic predictions, and worse clinical outcomes. Our model encompasses a broad range of patients presenting with acute respiratory failure and various clinical manifestations, ndings that support the use of biomarker-based subphenotyping in broader critically ill patient populations.
Subphenotyping research has gained wide popularity in recent years, not only within ARDS but within other critical care syndromes such as sepsis and AKI. Amongst these syndromes, attempts to subdivide patients solely based on clinical criteria have been largely unsuccessful when compared to biomarkerbased modeling. In ARDS, subgroups receiving differential ventilator management strategies based on CT morphology showed no difference in mortality, and this strategy even proved to be harmful if misclassi ed into the wrong group (34). In AKI, similar to ARDS, no effective therapeutic interventions exist other than supportive care and renal replacement therapy. However, with biomarker-based subphenotyping, two distinct subphenotypes of AKI emerged with differential responses to therapy such as vasopressin (35). Parsimonious modeling revealed that the distinguishing biomarkers in AKI are extremely similar to those of our cohort, including bicarbonate, angiopoietin-2, and TNFR1, ndings which raise the possibility of common pathways of systemic in ammatory responses between heterogeneous populations of critically-ill patients.
An important challenge in critical care pertains to the subjective and non-speci c nature of diagnostic criteria for critical illness syndromes. ARDS is systematically underrecognized or under-reported as a diagnosis in clinical practice; and even among expert providers, there are signi cant rates of diagnostic disagreement (11,36). ARDS recognition is straightforward in cases with typical presentations of diffuse, bilateral in ltrates on imaging with an obvious risk factor, such as pneumonia or sepsis. However less classic radiographic presentations are a source of uncertainty and diagnostic discordance. In a retrospective analysis where expert panels identi ed ARDS by electronic medical record review, ARDS diagnosis had been documented in the clinical chart only 12.4% of the time, at least in part due to underrecognition of the syndrome at the time of clinical encounter (12). Further investigation into underrecognition of ARDS demonstrated that interobserver agreement of ARDS diagnosis under Berlin criteria has only been moderate, with lack of consensus on chest radiograph interpretation accounting for most differences (11). Inter-rate agreement in ARDS radiographic identi cation remained low even after educational interventions (9). Given such inherent diagnostic uncertainty in ARDS, there is compelling need to understand the underlying biological mechanisms that drive different outcomes in these conditions. Our parsimonious model is not dependent on the requirement of a de nite clinical diagnosis (i.e. ARDS vs. CHF) and allows for objective classi cation into two distinctly behaving subphenotypes.
Our LCA and parsimonious models were created after combining ARDS, ARFA, and CHF patients. While the sample sizes of each group within our study differ, we con rmed that the same hyper-and hypoin ammatory subphenotypes exist in all 3 subgroups independently, mitigating concern that any one group may be the dominating factor driving subphenotypic results. The remaining respiratory failure groups demonstrated a very low prevalence of hyper-in ammatory patients (7%), con rming that we included only clinically pertinent groups in our primary analysis. Patients intubated for airway protection or those with acute exacerbation for COPD appear to have extremely low prevalence of a hyperin ammatory subphenotype, and thus this classi cation framework would not be relevant in these forms of acute respiratory failure.
CHF is a clinical group with paucity of data on subphenotyping, although it has already been proposed that treatment of systolic heart failure should be shifted away from strategies that simply improve cardiac function towards interventions that modulate the systemic responses to cardiac dysfunction instead (37). Interestingly, our results demonstrate that hyper-and hypo-in ammatory subphenotypes exist in CHF patients in very similar proportion to ARDS and ARFA patients, further suggesting biological commonalities amongst critical illness syndromes, regardless of clinical diagnosis. Exclusion of cardiac edema is a major source of disagreement in the clinical decision to diagnose ARDS, and as seen in the FACTT trial, a signi cant amount of patients diagnosed with ARDS have elevated pulmonary capillary wedge pressures (38). Consequently, it is probable that a certain unknown proportion of patients with ARDS are misclassi ed as cardiogenic pulmonary edema, and vice versa. A recent study evaluated the longitudinal evolution of radiographic pulmonary edema in ARDS and found that baseline radiographic pulmonary edema was not associated with ARDS severity, clinical outcomes or hyper-in ammatory subphenotype (13). These ndings further highlight the disadvantage in using clinical criteria rather than objective biological markers to subdivide such populations.
Our 4-variable parsimonious model includes two markers that have been associated in cohorts much larger than our current study (bicarbonate and TNFR1) (8,20), one well-validated marker in sepsis and ARDS (angiopoeitin-2) (23,24), and one already widely clinically available and utilized test (procalcitonin) (27,28). Angiopoietin-2 may be a possible causal factor in development of ARDS in septic patients (24), in addition to an independent predictor of increased mortality in ARDS (25,39). Angiopoietin-2 has been included in prior parsimonious models (25,35,40) for ARDS and AKI subphenotypes. Similarly, bicarbonate and TNFR1 have been identi ed as key predictors in prior parsimonious models for ARDS (6,41). Notably, these analyses did not include the other two biomarkers in our parsimonious model (angiopoietin-2, procalcitonin) as possible classi er variables. Further veri cation in larger data sets is therefore required. While a parsimonious model allows for easier clinical applicability than complex LCA, models that depend on biomarker variables will eventually require point-of-care or rapid turnaround tests to ensure timely acquisition of results and application to model predictions.
Our study utilized a simpli ed approach to variable selection for the LCA models. Prior studies have utilized large amounts of variables without any selectivity. In our approach, we removed nondiscriminatory variables prior to analysis and found excellent agreement between models before and after removal (Fig. 2). These ndings demonstrate the ability to perform LCA with a smaller set of variables and similar or improved statistical performance of the model.
Our study has several limitations. While our study prospectively enrolled consecutive patients with acute respiratory failure, it is limited by sample size and single center design. We also only examined baseline data (within 48 hours intubation), therefore it is unclear whether patients transition between subphenotypes over time. In one analysis, from day 0 to day 3 the majority (> 94%) of patients remained in the same subphenotype (42), though preliminary data within our own cohort demonstrate the possibility of higher rates of transition by days 3-6 (43). Further examination of the stability of subphenotypes will be important, as transition from one group to another will affect the ability to target clinical interventions. While we had an independent data set for validation demonstrating similar trends with outcomes biomarkers, this validation dataset is small, illustrating the need to validate our model in larger data sets. Additionally, we have not yet investigated any differential therapeutic responses between groups, which will be required for clinical applications. Larger, prospective trials of heterogeneous critically ill populations will be necessary to verify subphenotypic models and and explore treatment effects.

Conclusions:
The     Hyper-in ammatory subjects have higher mortality and longer mechanical ventilation duration in LCA and parsimonious models. Kaplan Meier curves for 30-d survival (left panels) and time-to-liberation from mechanical ventilation (right panels) for each subphenotype as derived by LCA (top row) and the parsimonious 4-variable model (bottom row). P-values for differences between subphenotypes were obtained with a log-rank test. Adjusted hazard ratios (aHR) with 95% con dence intervals are displayed for the effects of the hyper-in ammatory subphenotype and were derived from multivariate Coxproportional hazards models. All models were adjusted for age. 90-day survival data were very similar to 30-day and are not shown.

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download. AdditionalFile4.docx AdditionalFile5.docx AdditionalFile2.docx