Two Physiologic Latent Classes of Acute Hypoxemic Respiratory Failure in Sepsis Are Distinguished by Lung Mechanics and Gas Exchange

Abstract

OBJECTIVES:

Physiologic subtypes of acute hypoxemic respiratory failure (AHRF) may confer a differential response to treatments, particularly therapeutic strategies that are specific to pulmonary organ failure. We sought to identify physiologic latent classes of sepsis-associated AHRF defined by respiratory mechanics, oxygenation, ventilation, and radiographic patterns of lung injury, and to determine the association between class membership and 30-day mortality.

DESIGN:

We performed latent class analysis of patients with AHRF newly requiring mechanical ventilation enrolled in a prospective cohort of patients with sepsis from 2011 to 2020. We used logistic regression adjusted for Acute Physiology and Chronic Health Evaluation to determine the association between class membership and 30-day mortality and examined the distribution of patients classified as “hyperinflammatory” by previously described biomarker-based subphenotyping paradigms.

SETTING:

Philadelphia, Pennsylvania, United States.

PATIENTS:

Eight hundred eighty-two patients.

MEASUREMENTS AND MAIN RESULTS:

We identified two physiologic latent classes. Class 1 (n = 390) was characterized by low static compliance and impaired ventilation when compared with class 2 (n = 432). Mortality at 30 days was higher in the more physiologically severe class 1 when compared with class 2 (adjusted risk difference 0.12, p < 0.001) despite a similar severity of sepsis. Class 1 also contained a higher proportion of female patients and patients with obesity.

CONCLUSIONS:

We identified two physiologic latent classes of sepsis-associated AHRF. Relative to class 2, class 1 was distinguished by low compliance, impaired ventilation, and higher 30-day mortality independent of the severity of sepsis. The higher percentage of female patients and patients with obesity in class 1 suggests a potential role for body composition in class determination. Physiologic classes were not primarily determined by qualification for acute respiratory distress syndrome or previously described biomarker-based subphenotypes, suggesting a distinct physiologic “axis” of heterogeneity.

KEY POINTS

Question: Are physiologic subtypes of sepsis-associated acute hypoxemic respiratory failure (AHRF) associated with differential 30-day mortality?

Findings: This study details two latent class-derived physiologic subtypes of AHRF, distinguished by compliance and ventilatory efficiency, with a statistically significant differential 30-day mortality.

Meaning: Physiologic class membership provided prognostic information beyond the severity of sepsis. Physiologic classes were not primarily driven by qualification for the acute respiratory distress syndrome (ARDS) and were distinct from previously described biomarker-based subphenotypes of ARDS.

In the United States, over one million patients are hospitalized with acute ventilator-dependent respiratory failure annually (1). Short-term mortality among patients with acute hypoxemic respiratory failure (AHRF) is estimated between 30% and 40% (2, 3). Those who survive the acute phase of illness are at risk for late mortality following hospital discharge, cognitive impairment, mental health disorders, and functional limitations (2, 4). Patients with sepsis, including pneumonia, are at high risk for AHRF characterized by lung inflammation and resulting physiologic derangements (5, 6). There is a pressing need to better understand physiologic and biologic heterogeneity to facilitate a precision medicine approach to AHRF in sepsis. Personalization of systemic and lung-specific treatment strategies has the potential to reduce the unacceptably high morbidity and mortality associated with AHRF in critical illness.

Physiologic heterogeneity in AHRF has been recognized for decades. Thomas Petty and David Ashbaugh’s first description of the acute respiratory distress syndrome (ARDS) in 1971 detailed a severe subtype of AHRF characterized by a loss of pulmonary compliance responsive to increased positive end-expiratory pressure (PEEP) (7). Measures of oxygenation and the radiographic pattern of lung injury have since been incorporated into historical and contemporary definitions of ARDS and severity categories (8, 9). However, physiologic parameters related to compliance and dead space were considered but ultimately not included in the Berlin definition of ARDS based on failure to improve prognostic validity for mortality in a small subset of a retrospective cohort (10). Nevertheless, in subsequent work, both compliance and dead space have demonstrated a strong association with mortality in AHRF and ARDS (11–14), conferred a differential benefit to important treatment modalities (15–17), and have been proposed as strategies for clinical trial enrichment and stratification (18, 19). Furthermore, these parameters have been used to identify meaningful cross-sectional and longitudinal physiologic classes of ARDS via unsupervised learning techniques (20–23). Routinely available physiologic parameters that are not included in the syndrome of ARDS may result in heterogeneous treatment effects of therapeutic strategies and prove consequential in precision medicine for AHRF.

The ARDS construct initially allowed for clinical trial enrichment that led to breakthroughs in critical care, such as low-stretch ventilation and prone positioning (24, 25). However, promising treatment strategies have likely failed to demonstrate a uniform benefit in ARDS in randomized controlled trials due to unresolved physiologic and biologic syndrome heterogeneity (26–28). Recent advances in our understanding of biologic heterogeneity within ARDS include the identification of ARDS subphenotypes defined by biomarkers reflective of systemic inflammation (29). These subphenotypes confer heterogeneity of treatment effect to systemic therapies such as Simvastatin and IV fluids (30, 31). However, the relevance of these biomarker-based subphenotypes to lung-specific therapeutic strategies is less clear, as a differential treatment effect was observed in post hoc analyses of randomized controlled trials of PEEP but not neuromuscular blockade (29, 32). In addition, as ARDS remains under-recognized and subject to poor diagnostic interrater reliability (6, 33), subphenotyping paradigms within ARDS may exclude patients with severe respiratory impairment who would also benefit from therapeutic consideration, such as patients with unilateral lung injury. Physiologic subphenotyping of AHRF may better predict treatment response to certain treatments, particularly those that are specific to pulmonary organ failure.

Our objective was to identify physiologic latent classes of AHRF defined by respiratory mechanics, oxygenation, ventilation, and radiographic pattern of lung injury in a prospective cohort of patients with sepsis. We hypothesized that these physiologic classes would be prognostically significant and would differ from previously described biomarker-based subphenotypes of ARDS. Some of the results of this study have previously been reported in the form of an abstract (34).

MATERIALS AND METHODS

Study Population

This study was nested within the Molecular Epidemiology of SepsiS in the ICU (MESSI) prospective cohort, an ongoing cohort of critically ill patients with sepsis. Patients were eligible for the primary cohort if they were admitted to the ICU within the University of Pennsylvania Health System and met contemporary criteria for sepsis (35, 36), and were excluded if they were previously enrolled, chronically ventilator-dependent, or elected exclusively comfort-focused care at the time of ICU admission (37, 38). Patients were enrolled with a waiver of timely informed consent, and then patients or their surrogates ultimately provided consent for enrolled patients. The institutional review board at the University of Pennsylvania approved this protocol (MESSI, number 808542, approval date July 22, 2024), and all study protocols were followed in accordance with institutional standards for the responsible conduct of human research and with the Helsinki declaration of 10975.

We identified a subpopulation enrolled in MESSI between 2011 and 2020 who developed AHRF within 5 days of enrollment, defined by need for invasive mechanical ventilation and a Pao₂ to fraction of inhaled oxygen (Pao₂/Fio₂ ratio) of equal to or less than 300. Patients were considered to be enrolled in this subcohort at the time of initiation of invasive mechanical ventilation. We excluded patients if they were transferred from another hospital or if they did not have an arterial blood gas or respiratory mechanics recorded within 48 hours of initiation of mechanical ventilation.

Data Collection

Trained research personnel collected clinical data by direct review of the medical record using Research Electronic Data Capture (Vanderbilt University, Nashville, TN)-based electronic data capture forms and automated abstraction of the electronic medical record (39). Data collected included demographic information, detailed medical history, physiologic variables, and laboratory results.

We selected class-defining variables to reflect oxygenation, ventilation, pulmonary mechanics, and the radiographic pattern of lung injury. Although the relationship between respiratory physiology and ventilator management is highly intertwined, we prioritized variables that were most likely to reflect underlying physiology rather than bedside clinician decision-making. We obtained arterial pH, Pao₂, and Paco₂ from arterial blood gas analysis. We obtained plateau pressure, driving pressure, and static compliance from clinical interrogation of ventilator mechanics by respiratory therapists using an inspiratory breath hold. In our medical system, respiratory therapists routinely perform these measurements and record them in the medical record immediately after intubation and every 4–8 hours of invasive ventilation. Patients in support modes of mechanical ventilation are temporarily switched to controlled modes for the purposes of interrogating mechanics per protocol. Measurement and recording of plateau pressure is omitted at the discretion of the respiratory therapist for patients with spontaneous respiration who are judged to have a high work of breathing. All Pao₂, fraction of inhaled oxygen, static compliance, driving pressure, and plateau pressure collected for clinical purposes during the first 24 hours of mechanical ventilation (i.e., the first 24 hr from enrollment in the subcohort) were averaged and operationalized as continuous variables. Physiologically implausible measurements (i.e., zero or negative driving pressure) were excluded from the analysis. Study investigators reviewed chest radiographs and cross-sectional imaging for the presence of bilateral infiltrates consistent with ARDS, with group-level adjudication for equivocal cases (8, 40). Radiographic pattern was represented using a binary variable for the presence of bilateral infiltrates consistent with ARDS (i.e., no infiltrate or unilateral infiltrate vs. bilateral infiltrates). We determined mortality at 30 days via review of the electronic health record, with an online obituary search for patients without a healthcare encounter documented at least 30 days after enrollment.

In addition, we sought to compare physiologic phenotypes with biomarker-based inflammatory subphenotypes using a subset of patients with plasma biomarker measurements available from the day of enrollment in the parent study (which was, in some cases, distinct from the first day of mechanical ventilation). These measurements were performed on blood drawn into citrate vacutainers for clinical purposes upon presentation to the emergency department or ICU and, per clinical laboratory protocol, centrifuged within 30 minutes and stored at 4°C for up to 48 hours. MESSI coordinators obtained aliquots of plasma from these samples and froze them at –80°C. Inflammatory biomarkers were assayed with electrochemiluminescence (Meso Scale Diagnostics, LLC, Rockville, MD) as previously described (41). We classified patients as hyperinflammatory or hypoinflammatory based on the parsimonious model described by Sinha and colleagues using interleukin-8 (IL-8), soluble tumor necrosis factor receptor-1 (sTNFR-1), and serum bicarbonate (42).

Statistical Analysis

Statistical analyses were conducted in STATA (STATA BE, Version 17.0; StataCorp, College Station, TX). Continuous class-defining variables underwent log-transformation and standardization to a z scale to approximate a normally distributed, uniformly scaled dataset. We used STATA’s gsem lclass function, assessing convergence over 1000 iterations and 100 random seeds. We examined Bayesian information criterion (BIC), entropy, class size, and clinical plausibility and relevance to determine the optimal number of classes (43). We assigned patients to their most likely class and determined differences in non-class-defining clinical characteristics across classes using Kruskal-Wallis or chi-square tests, as appropriate. Logistic regression was used to assess the relationship between class membership and 30-day mortality with adjustment for Acute Physiology and Chronic Health Evaluation III (APACHE III).

RESULTS

Of the 3112 patients enrolled in MESSI, 1186 had AHRF requiring invasive mechanical ventilation, and 882 met no exclusion criteria (Fig. 1). Included patients had a median age of 62 (interquartile range [IQR] 53 to 81); 61% were male, 60% White, and 32% Black or African American. There was a high prevalence of active malignancy (371, 45%) consistent with our medical system’s large oncology program, and the majority of patients met criteria for ARDS on the day of intubation (562, 68%).

Figure 1.

Selection of the study population. ABG = arterial blood gas, AHRF = acute hypoxemic respiratory failure, IMV = invasive mechanical ventilation, MESSI = Molecular Epidemiology of SepsiS in the ICU.

BIC decreased and entropy increased as the number of classes increased. We identified the clearest inflection points in the BIC elbow plot at two and four classes (Supplemental Fig. 1, https://links.lww.com/CCX/B551). We considered two-, three-, and four-class models. Although fit indices slightly improved with an increased number of classes, the additional classes were small, and on our evaluation, did not augment the clinical relevance of the models; thus, we selected the two-class model (Supplemental Results, https://links.lww.com/CCX/B551). The two-class model exhibited excellent classification, with a median (IQR) probability of membership to class 1 of 98.9% (89.2–100.0%) among patients assigned to class 1 and 0.7% (0.0–10.0%) among patients assigned to class 2 (Supplemental Fig. 2, https://links.lww.com/CCX/B551).

Compared with class 2 (n = 432, 52.3%), class 1 (n = 390, 47.5%) exhibited more severe respiratory physiology, characterized by worse respiratory mechanics and gas exchange (Table 1), with higher plateau and driving pressure, lower static compliance, higher Pco₂, lower pH, and lower Pao₂. Pulmonary mechanics and parameters related to ventilatory efficiency, including driving pressure, static compliance, and Paco₂, showed greater differences between classes than variables related to oxygenation (Fig. 2). All class-defining variables, with the exception of Fio₂, meaningfully contributed to the prediction of physiologic class (Supplemental Results, https://links.lww.com/CCX/B551). Bilateral opacities consistent with ARDS were more common in class 1 than class 2 but were prevalent in both classes (75.4% vs. 62.0%, p < 0.001).

TABLE 1.

Description of Latent Class Analysis-Derived Classes of Acute Hypoxemic Respiratory Failure

Latent Class Characteristic	Class 1, n = 390	Class 2, n = 432	p
Class-defining variables
Plateau pressure (cm H2O)	26.0 (23.0–29.0)	19.0 (16.0–21.0)	< 0.001
Compliance (mL/cm H2O)	23.2 (18.8–26.5)	38.5 (32.6–46.4)	< 0.001
Paco2 (mm Hg)	41.5 (34.4–47.7)	36.5 (31.3–42.4)	< 0.001
pH	7.3 (7.2–7.4)	7.3 (7.3–7.4)	< 0.001
Pao2 (mm Hg)	111.8 (88.8–142.2)	117.0 (95.4–151.8)	0.003
Fio2	0.9 (0.6–1.0)	0.8 (0.6–1.0)	0.038
Driving pressure (cm H2O)	18.0 (16.2–21.0)	12.0 (10.0–14.0)	< 0.001
Bilateral opacities on chest imaging	294 (75.4%)	268 (62.0%)	< 0.001
Demographics
Age (yr)	60.5 (51.0–69.0)	64.0 (53.5–72.0)	0.001
Male sex	205 (52.6%)	292 (67.6%)	< 0.001
Racea
American Indian or Alaska native	0 (0.0%)	4 (0.9%)
Asian	17 (4.4%)	15 (3.5%)
Black or African American	141 (36%)	123 (28%)
Native Hawaiian or other Pacific Islander	1 (0.3%)	0 (0.0%)
White	218 (56%)	274 (63%)
Multiple races or not reported	13 (3.3%)	16 (3.7%)
Hispanic ethnicity	13 (3.3%)	15 (3.5%)
Baseline comorbidities
Obesity (body mass index > 30 kg/m2)	160 (41.0%)	131 (30.3%)	0.001
Chronic obstructive pulmonary disease	42 (14.4%)	44 (12.9%)	0.57
Interstitial lung disease	28 (7.2%)	24 (5.6%)	0.34
Congestive heart failure	57 (14.6%)	55 (12.7%)	0.43
Chronic kidney disease	67 (17.2%)	67 (15.5%)	0.64
Active solid malignancy	71 (18.2%)	85 (19.7%)	0.55
Active hematologic malignancy	101 (25.9%)	127 (29.4%)	0.26
Clinical characteristics
Acute Physiology and Chronic Health Evaluation III	112.0 (86.0–134.0)	111.0 (88.5–137.0)	0.82
Pulmonary source of sepsis	236 (54.6%)	265 (67.9%)	< 0.001
Outcome
Mortality at 30 d	251 (64.5%)	230 (53.2%)	0.001
Relationship to biomarker-based subphenotypes	n = 250	n = 253
Membership in hyperinflammatory subphenotype	140 (48.1%)	151 (51.9%)	0.70

^aRace and ethnicity were not operationalized as biological variables, and as such, differences between classes were not statistically examined.

Values are listed as median (interquartile range) or number (percentage).

Figure 2.

Continuous variables were log-transformed and scaled to a mean of 0 and sd of 1 (z scale). DP = driving pressure, Pplat = plateau pressure.

Mortality at 30 days was high in the overall population, and higher in the more physiologically severe class 1 than class 2 (64.5% vs. 53.2%, unadjusted risk difference 0.12; 95% CI, 0.05–0.18; p = 0.001). Outcomes differed by predicted class membership despite a similar severity of illness between class 1 and class 2 (APACHE III score 112.5 vs. 111.0). Class membership provided prognostic information beyond the severity of sepsis, as the association with 30-day mortality was robust to adjustment for APACHE III score (adjusted risk difference 0.12; 95% CI, 0.05–0.18; p < 0.001).

In addition to class-defining variables, classes also differed by characteristics not incorporated into the model. The more severe class 1 contained a higher proportion of female patients (47.4% vs. 32.4%, p < 0.001) compared with class 2. Given this finding, we performed a post hoc analysis comparing tidal volume adjusted for predicted body weight (PBW) in female vs. male patients in our cohort. Tidal volume adjusted for PBW was higher in female patients than male patients (7.02 vs. 6.78 cc/kg PBW, p = 0.03). Patients with membership to class 1 were also more likely to have obesity as defined by body mass index (BMI) (41.0 vs. 30.3%, p = 0.003) and to have a pulmonary source of sepsis (68.0% vs. 54.6%, p = 0.001) compared with class 2. We did not detect differences in other comorbidities by class.

Inflammatory plasma cytokines were available for 503 patients (57%). This subset was enriched for patients with baseline hematologic malignancy and with ARDS (Supplemental Table 4, https://links.lww.com/CCX/B551). Of these patients, 250 had membership in class 1, and 253 had membership in class 2. The proportion of patients classified as hyperinflammatory based on IL-8, sTNFR, and serum bicarbonate was similar between class 1 and class 2 (56% vs. 60%, p = 0.40, Table 2).

TABLE 2.

Comparison of Biomarkers and Prevalence of Parsimonious Biomarker-Based Subphenotypes Between Physiologic Classes of Acute Hypoxemic Respiratory Failure

Biomarker/Class Assignment	Class 2, n = 250	Class 2, n = 253	p
Interleukin-8 (pg/mL)	90 (27—641)	115 (39–493)	0.49
Soluble tumor necrosis factor receptor-1 (pg/mL)	6,854 (3,270–14987)	7,696 (3,931–13,350)	0.67
Bicarbonate	19 (14–23)	18.7 (15–22)	0.55
Hyperinflammatory	140 (56%)	151 (60%)	0.40
Hypoinflammatory	110 (44 %)	102 (40%)

“Hyperinflammatory” and “hypoinflammatory” refer to membership in the parsimonious, biomarker-based subphenotype.

DISCUSSION

Using latent class analysis, we derived two physiologic classes of AHRF requiring mechanical ventilation, primarily distinguished by static compliance and ventilatory abnormalities, within a prospective cohort of critically ill patients with sepsis. Physiologic class exhibited differential 30-day mortality despite similar severity of sepsis and proportion of patients classified as hyperinflammatory by plasma cytokines (42). Classes also differed significantly in terms of sex and obesity as defined by BMI. Specifically, the higher mortality class 1 had a higher proportion of women and a higher BMI relative to the lower mortality class 2.

Alterations in pulmonary mechanics and gas exchange are central to the pathophysiology of ARDS, in which the development of a severe inflammatory lung injury and subsequent decreased alveolar volume clinically manifests as low static compliance and increased dead space (22). Our subtypes were principally driven by parameters related to static compliance and ventilatory efficiency, and provided prognostic information beyond the severity of sepsis. Notably, as the relationship between respiratory physiology and ventilator management is often bidirectional, subtypes may be driven both by intrinsic differences in respiratory physiology and by systematic differences in ventilatory management and critical care delivery. Our work complements the latent classes described by Wendel Garcia et al (23), which were derived from clinical variables, respiratory mechanics, and granular CT-derived variables at standardized PEEP levels after administration of neuromuscular blockade among patients with ARDS. Phenotypes were distinguished by compliance, dead space, ICU mortality, and recruitment of non-aerated lung tissue at 45 cm H₂O of PEEP. This further substantiates the importance of static compliance and dead space in outcomes from AHRF and suggests a differential response to ventilator management strategies. Further research is required to determine whether paralysis and CT are necessary to predict recruitability, as clinical trial enrichment using a subtyping paradigm reliant on routinely available data in a broader population, such as ours, may enhance feasibility and enrollment.

Our subtypes were not principally driven by the presence of bilateral opacities on chest imaging, which qualifies patients for the Berlin definition of ARDS within our cohort. Notably, respiratory mechanics were available for less than 10% of the cohort used to validate the Berlin definition, and this subset may have been underpowered to detect the importance of compliance (8). Our work suggests that respiratory physiology beyond the ARDS construct may be relevant to clinical practice and to enrichment of clinical trials investigating therapeutics for AHRF. This is supported by a recent post hoc analysis of the Reevaluation Of Systemic Early neuromuscular blockade (ROSE) trial demonstrated heterogeneity of treatment effect by baseline compliance (17). Further efforts to understand the predictive validity of physiologic parameters are already underway; for example, an ongoing trial investigating driving pressure-based ventilatory strategies is stratified by static compliance (18). Additional research is needed to assess whether physiologic phenotypes confer a differential response to treatment modalities, particularly therapeutics specific to the respiratory system, such as PEEP, neuromuscular blockade, and prone positioning.

Among patients with measured inflammatory cytokines, the proportions of patients characterized as hyperinflammatory and hypoinflammatory were similarly distributed between physiologic classes. This suggests that our derived physiologic classes do not merely recapitulate previous biomarker-based ARDS subphenotypes. ARDS/sepsis subphenotypes previously described by Calfee et al (44) have demonstrated relevance to systemic therapies, but may be less predictive of response to therapeutic strategies that are specific to respiratory failure, particularly given that systemic biomarkers of inflammation do not fully capture the inflammatory milieu of the lung compartment . Furthermore, physiologic derangements in AHRF are not exclusively mediated by acute inflammation, and physiologic subtyping may capture differences in anatomy, body composition, chronic respiratory physiology, noninflammatory pulmonary edema, and fibrosis that have relevance to therapeutic candidates for AHRF. Physiologic heterogeneity may thus be complementary to established biomarker-based subphenotypes of ARDS in predicting response to supportive strategies that are specific to pulmonary organ failure, such as neuromuscular blockade, which exhibited heterogeneity of treatment effect by compliance but not by inflammatory subphenotype in post hoc analyses of a randomized controlled trial (17, 32).

We observed a higher proportion of patients with obesity in the more severe physiologic class 1 relative to class 2. This may be attributable to mechanical factors related to decreased chest wall compliance, increased atelectasis, and increased V/Q mismatch associated with abdominal and chest adiposity (45). Patients with obesity exhibit increased pleural pressures, resulting in low lung compliance that is out of proportion to the degree of lung injury when compared with patients without obesity. Alternatively, physiologic class membership may be influenced by systematic differences in medical management according to BMI (46, 47), as well as the metabolic and immunomodulatory effects of adipose tissue in critical illness (48–50). Further research is needed to determine whether physiologic classes mirror the “obesity paradox” that has been described in ARDS (51, 52), in which patients with obesity are more likely to have membership in the more physiologically severe class 1 but are less likely to die than patients without obesity in class 1. The present study is unfortunately underpowered to detect effect modification by obesity on the relationship between class assignment and mortality, and we did not have access to esophageal manometry to distinguish between decreased compliance of the respiratory system attributable to acute lung injury vs. the effect of thoracic and abdominal adiposity on chest wall compliance.

The higher mortality class 1 was more likely to be female compared with class 2, despite prior literature demonstrating a higher mortality among men compared with women in ARDS (53). This is consistent with a recent study detailing a COVID-19-related ARDS subphenotype with a higher proportion of female patients that was also associated with low compliance, high ventilatory ratio, and higher mortality (21). There are several potential explanations for these findings. On average, female patients received a higher tidal volume adjusted for PBW than male patients in our cohort, which is consistent with prior studies documenting differential adherence to lung-protective ventilation by sex (54, 55). This may have led to worse injury and, therefore, an increased proportion of female patients in class 1. Alternatively, differences in body composition by sex may lead to a greater degree of ventilator-induced lung injury with appropriate lung-protective ventilation, beyond differences accounted for in the formula for PBW. This is supported by literature documenting female sex as an independent risk factor for mortality among patients with severe ARDS (55). Finally, biologic differences attributable to the immunomodulatory effect of sex hormones may account for sex differences in outcomes (56).

Our study has several limitations. Establishing generalizability of the identified physiologic classes is critical, as subtypes may be influenced by differences inpatient populations, ventilator management protocols between institutions, and precipitating events related to AHRF (i.e., beyond sepsis). Our cohort was conducted over 10 years at a large referral center with a diverse patient population, but generalizability may be limited due to the very high mortality attributable to enrichment with patients with an active solid malignancy, hematologic malignancy, and cirrhosis. Second, our sample size may have been too small to demonstrate the relevance of three- and four-class models, which may confer additional prognostic and predictive validity to physiologic subtyping paradigms. Third, pulmonary mechanics played a disproportionately large role in the determination of class membership. Our population may be spread along a continuum of severity of indicator variables related to pulmonary mechanics, rather than true categorical latent classes. This phenomenon is termed the “salsa effect,” in which individuals are “forced” into separate classes when they are in fact distributed along a spectrum of a single variable (43). However, variables related to gas exchange, radiographic findings, and oxygenation enhanced the prediction of class membership, suggesting that classes were not exclusively driven by pulmonary mechanics. Further research should examine the prognostic and predictive validity of simpler, univariate physiologic models (such as respiratory compliance and ventilatory efficiency) in addition to more complex subtypes that integrate multiple physiologic domains. Fourth, our study was underpowered to determine effect modification by sex and obesity on the relationship between class membership and mortality. Finally, inflammatory cytokines were measured in a subset of patients that represents a convenience sample in our cohort, and were often measured before the initiation of mechanical ventilation. However, the characteristics of this subset did not substantially differ from the study population as a whole, and previously described biomarker-based subphenotypes of ARDS have demonstrated stability over several days (57).

CONCLUSIONS

We have identified two latent classes of sepsis-associated AHRF based on physiologic and radiographic variables. Relative to class 2, class 1 exhibits poor static compliance and impaired ventilation and has higher 30-day mortality. These classes differ from previously described biomarker-based subphenotypes of ARDS, suggesting a novel physiologic “axis” of heterogeneity in AHRF. The higher percentage of female patients and patients with obesity in class 1 suggests a potential role for body composition in class determination. Further research should focus on the impact of treatments specific to respiratory failure, such as neuromuscular blockade, on outcomes by class membership, as well as the relationship between lung-protective ventilation, sex, and body composition in AHRF.

Supplementary Material

cc9-7-e1314-s001.pdf ^{(815.9KB, pdf)}