Abstract
Background
Survivors of ARDS experience significant morbidity and mortality after hospital discharge. A latent class analysis (LCA) proposed 5 sepsis survivor subtypes that may be of use for ARDS patients as well, but did not include an operationalization algorithm for assigning patients to subtypes.
Research Question
Can the Taylor sepsis survivor subtypes be operationalized in an ARDS population, and does such operationalizations distinguish patients at varying risk of post-discharge mortality and readmission?
Study Design and Methods
We conducted a secondary analysis of adults with acute respiratory distress syndromes requiring mechanical ventilation enrolled in the Reevaluation of Systemic Early Neuromuscular Blockade (ROSE) trial. We compared two methods of operationalizing subtype to develop an assignment rule: transported subtypes based on an algorithm developed using five variables from the previously published derivation cohort, and de novo subtypes derived from a new latent class analysis (LCA). Random-effect logit models were used to estimate the association between subtype and the primary outcome of mortality at 12 months.
Results
580 participants were assigned to five subtypes using the transported approach. In age-adjusted regression, transported subtype was significantly associated with mortality at 12 months (p=0.027) and readmission at 3 months (p=0.043). A de novo LCA identified 5 distinct subtypes as the optimal solution; subtypes were associated with mortality at 3, 6, and 12 months (p<0.001) and readmission at 3 months (p=0.008). Agreement between the two assignment methods was 48% (kappa 0.361), and concentrated in the lowest (134/145, 92.4%) and highest risk (44/44, 100%) groups.
Interpretation
Sepsis survivor subtypes can be transported and operationalized in ARDS patients. Different approaches to operationalization yield similar but not identical subgroups classifications. Both approaches assign individuals to groups that are significantly associated with patient-important outcomes among ARDS survivors. Optimal strategies to transport externally derived subtypes, versus internal rederivation, may depend on the specific use case.
MeSH keywords: critical care, patient reported outcome measures
Individuals who survive acute respiratory distress syndrome (ARDS) experience high rates of disability and mortality after hospital discharge.1 After one year, up to 40% are readmitted, 75% report difficulty with daily activities, and only 20% return to work.2,3 Survivors of ARDS are heterogenous by many measures: the nature of the inciting event, prior health status, severity of disease, and access to care, among others. The discovery of hypo- and hyperinflammatory subtypes that exhibit divergent inflammatory profiles and short-term response to therapeutic interventions has catalyzed interest in ARDS population heterogeneity.4,5 Identification of distinct subtypes associated with similar risk for adverse outcomes and need for services could enable precise and efficient post-discharge support.6
Among patients hospitalized with sepsis, Taylor and colleagues performed a latent class analysis (LCA) using inputs related to post-discharge complexity alongside routinely collected clinical data to identify a model with five subtypes.7 These sepsis survivor subtypes exhibited significantly different proportions of patients experiencing mortality and readmission at thirty days, but have not been operationalized in a clinical setting. The LCA as published did not include an algorithm that could be used to prospectively assign patients to survivor subtypes to test the generalizability (in sepsis) or transportability (to other conditions, specifically here ARDS as our subject of interest).
We sought to fill these gaps in two steps. First, we operationalized the Taylor sepsis survivor subtypes for use in ARDS using two distinct strategies (described below). Second, we tested whether these operationalizations in ARDS have predictive validity for patient-centered outcomes after discharge. Our central hypothesis is that sepsis survivor subtypes can be operationalized in a population with ARDS and that the odds of important clinical outcomes will differ significantly between subtypes. The scientific rationale is that the physiology and processes of care for both conditions, while not identical, have similarities.
Methods
Setting
We included participants who enrolled and survived until discharge in the Reevaluation of Systemic Early Neuromuscular Blockade (ROSE) trial, a multicenter randomized controlled trial conducted in the United States.2,8 Adult patients with moderate-to-severe ARDS (ratio of partial pressure of arterial oxygen to the fraction of inspired oxygen [PaO2/FiO2] of less than 150 mm Hg with a positive end-expiratory pressure of >=8 cm of water) requiring mechanical ventilation were randomized 1:1 to receive either usual care with light sedation targets or a 48-hour continuous infusion of cisatracurium. All ROSE participants who on day 28 were alive in the hospital or had been discharged were included for analysis. Participants with missing data for any of the four class-defining variables were excluded. This secondary analysis of anonymized data was deemed exempt by the corresponding author’s institutional review board.
Description of previously validated sepsis survivor subtypes
In a prior study, the Taylor sepsis survivor subtypes were derived using a latent class analysis of data from patients presenting to hospitals in the Southeast United States who met a clinical definition of sepsis. The specific names given to each subtype and their descriptions are (Figure 1):
Figure 1.
Overview of sepsis survivor subtypes derived by Taylor et al
Low risk, barriers to care. Younger patients with few comorbidities. Hereafter referred to as “low risk” for brevity.
Multimorbidity. Older patients with more comorbidities who are frequently hospitalized.
Poor functional status. Older patients with high prevalence of frailty at discharge and high functional needs who are often discharged to a facility instead of home.
Previously healthy with severe illness and complex needs after discharge, barriers to care. Younger patients with longer hospital stay and high prevalence of frailty at discharge. Hereafter referred to as “healthy baseline with severe illness.”
Existing poor health with severe illness and complex needs after discharge. Older patients with severe comorbidities, high functional needs, and prolonged hospital stay. Hereafter referred to as “Unhealthy baseline with severe illness.”7
Variables for developing sepsis survivor subtypes in ROSE
A broad overview of our approach can be found in Figure 2. Latent class analysis is a statistical technique that identifies subgroups within a population that share characteristic patterns.9 It is a descriptive technique. There is no standard method to operationalize a latent class analysis and its identified groups in one population into a decision tree with specific classification thresholds for use in a different population. Instead, by describing broad patterns in how characteristics are grouped, latent class analysis can generate hypotheses about classification rules that could be applied to prospectively sort participants from a new population into groups.10 For example, Sinha et al have published on using logistic regression to develop a classification approach for the Calfee et al ARDS inflammatory phenotypes that may be useful in the inpatient setting, but has previously been shown to not distinguish survivor outcomes.4,11,12
Figure 2.
Overview of how subtypes were recreated using two different methods
The Taylor derivation latent class analysis used 14 routinely collected variables across five conceptual domains that were abstracted from the electronic health record and dichotomized (e-Table 1). Of these, there were four variables with a high degree of between-group differences that could be mapped directly to variables collected in ROSE:
Prolonged admission, defined as greater than or equal to 7 days.
High comorbidity burden, defined in the Taylor paper using Charlson comorbidity index greater than or equal to five. This specific measure of comorbidity was not collected in ROSE, and the age and comorbidity component of the Acute Physiology, Age, Chronic Health Evaluation (APACHE) III score was used instead.
Frequent hospitalization, defined as greater than or equal to two admissions in prior six months. This was adapted to greater than or equal to five days of hospitalization in the prior month by report of patient or proxy, based on the data collected in ROSE.
Frailty at discharge, defined as Braden score greater than or equal to 16. This was adapted to the highest level of physical activity achieved on the last recorded study date before discharge. Participants whose highest level of physical activity was transferring from bed to chair or worse were considered frail.
Three variables from the derivation study (required organ support, required mobility intervention, time to antibiotics) were not appropriate for inclusion because all ROSE patients required organ support; all participants received mobility intervention during their hospitalization per ROSE protocol; and not all had a condition for which antibiotics was appropriate. The seven remaining variables either contributed little to distinguishing groups in the derivation study or did not have plausibly corresponding measurements in ROSE: immunosuppressed condition, discharge to a facility, new device or procedure, polypharmacy at discharge, delirium during admission, Medicaid insurance or self-pay, and area deprivation index.
Approach to Operationalizing Survivor Subtypes
To test our hypothesis, we operationalized the Taylor survivor subtype classification in two complementary ways in an ARDS population: (1) using clinical judgement to create an algorithm that transports the logic of the Taylor survivor subtypes; and (2) and empirically by conducting a de novo latent class analysis using variables constrained to be similar to those available in the Taylor survivor subtype derivation, and then using a classification and regression tree to categorize patients into a survivor subtype.
Developing a decision tree to assign transported survivor subtype
Two authors (RJF, TJI) who provide clinical care for patients with ARDS and sepsis developed a decision tree by mutual consensus. The heat map of variable frequency in the Taylor derivation study and clinical experience was used to generate the overall hierarchy using variables available in ROSE (Figure 3). Attention was paid to ensuring that, to the extent possible based on variable availability, the final tree assigned patients to subtypes whose characteristics approximate those in the derivation study. Preference was given to a simple structure that did not require any calculations and used whole-number thresholds at each node. Node thresholds were varied if final group assignment was less than 5% of the total population. Subtypes assigned using this method are referred to as transported survivor subtypes. These enable application of the results of the Taylor study without re-assembling a large population cohort to redo a latent class analysis.
Figure 3.
Decision tree developed to assign transported sepsis survivor subtypes
De novo latent class analysis
The transported survivor subtypes might not be an optimal classification even with the same input variables in a new population. We also conducted a latent class analysis on the ROSE data guided by best practice recommendations.13,14 The same four variables were used as inputs: comorbidity burden, number of recent hospitalizations, length of stay, and frailty at discharge. No post-hospitalization data were used. Exploratory analyses varied the number of these that were included and tested use of log-transformation and collapsing continuous variables into ordinal categories while varying the number of subtypes from two to ten (e-Table 2). Models and number of subtypes were assessed by Akaike information criterion (AIC), Bayesian information criterion (BIC), Vuong-Lo-Mendell-Rubin test statistic, size of the smallest class, and entropy. Each participant was assigned to the subtype with the highest absolute posterior probability. Subtypes assigned using this method are referred to as de novo subtypes. After subtypes were assigned, a classification and regression tree (CART) analysis was done to visualize the structure of an empirically optimized decision tree for comparison with the transported subtype assignment rules (e-Figure 1).
Outcomes
The primary outcome was 12-month mortality. Secondary outcomes were hospital admission, health-related quality of life (HRQOL), physical function, and cognitive impairment at 12 months. Primary and secondary outcomes were also examined at intermediate time points of three and six months available in ROSE. Outcomes were assessed at three, six, and 12 months by telephone interview with the participant or a caregiver. The EQ-5D-5L instrument was used to measure HRQOL.15 Physical function was measured using the Katz Activities of Daily Living and Lawton Instrumental Activities of Daily Living (I/ADLs) to determine the participant’s ability to perform ten I/ADLs.16 Cognitive impairment was assessed using the Montreal Cognitive Assessment if the patient was the respondent, otherwise the Alzheimer’s Disease 8 was used with proxies. A description of key measurements and rationales for their inclusion in ROSE are detailed elsewhere.8
Statistical analysis, including rationale for primacy of unadjusted analyses
The analysis was guided by a use case where survivor subtype could be estimated near the time of discharge by a clinician, with subtype alone summarizing an estimated risk of adverse events after discharge. In this use case, increasing the number of included variables reduces the clinical usefulness of the subtype designation by increasing clinician burden. Further, since the primary exposure, survivor subtype, is a function of four other exposure variables, there is increased risk of overfitting and collinearity with adjusted models. For these reasons, we prespecified that only unadjusted and age-adjusted models would be fit. Adjustment for age was done to estimate if subtype provided additive information distinct from age, given the marked heterogeneity by age in the original development cohort and its prognostic importance.
Baseline characteristics between subtypes were described at the patient level. Between group differences were compared using one-way analysis of variance (ANOVA) and chi-squared tests for continuous and categorical variables, respectively. Unadjusted and age-adjusted odds ratios were calculated using a random effects logit model for binary outcomes of mortality and hospital admission, with center as random effect. We used the Wald test to calculate group tests of significance for each model. We performed sensitivity analyses for persistence of associations in groups defined by varied according to presence/absence of sepsis, ROSE intervention arm, and type of lung injury. Analyses were done in Stata 17 (StataCorp) and R (version 4.4.2; R Foundation for Statistical Computing) using the ‘rpart’ package.17
Results
Of 1,006 participants enrolled in ROSE, 371 died before day 28 and 55 were missing data on class-defining variables. Compared to those who patients who survived, those who died before day 28 were older, had more I/ADL limitations, more frequently had active malignancy, and had greater illness severity per multiple measures (APACHE III, modified SOFA, and PaO2:FiO2). There were no significant differences between patients included in the final analysis and those excluded due to missing data on class-defining variables (e-Table 3).
Transported survivor subtypes
Survivor subtypes were assigned to 580 participants using the transported algorithm (Table 1). Most individuals were in the healthy baseline, severe illness (273, 47%) and low risk (162, 30%) subtypes, with the remaining assigned to multimorbidity (52, 9%), low functional status (49, 8%), and unhealthy baseline, severe illness (44, 8%). Baseline demographic and clinical characteristics were significantly different between groups. The low risk subtype was significantly younger than other groups. The healthy baseline with severe illness group tended to be older with higher EQ-5D-5L and APACHE III scores on enrollment, and experienced longer duration of their index admission. The unhealthy baseline with severe illness subtype less frequently lived at home, had the lowest baseline EQ-5D-5L scores, and experienced the longest duration of index admission compared to other groups.
Table 1.
Characteristics of participants according to transported subtype
| Patient Subtype (labels as per Taylor et al 2022) | All (n = 580) |
p | |||||
|---|---|---|---|---|---|---|---|
| Low Risk (n = 162) |
Healthy baseline, severe illness (n = 273) |
Multimorbidity (n = 52) |
Low functional status (n = 49) |
Unhealthy baseline, severe illness (n = 44) |
|||
| Age—median (IQR) | 34 (28–40) | 61 (54–67) | 57 (51–65) | 64 (55–75) | 58 (51–64) | 55 (40 – 64) | <0.001 |
| Sex—N (%) | |||||||
| Female | 74 (46) | 120 (44) | 28 (54) | 19 (39) | 25 (57) | 266 (46) | 0.303 |
| Male | 88 (54) | 153 (56) | 24 (46) | 30 (61) | 19 (43) | 314 (54) | |
| Race—N (%) | |||||||
| White | 119 (73) | 191 (70) | 39 (75) | 35 (71) | 26 (59) | 410 (71) | 0.104 |
| Black | 19 (12) | 44 (16) | 9 (17) | 10 (20) | 6 (14) | 88 (15) | |
| Other | 6 (4) | 6 (2) | 0 () | 1 (2) | 0 () | 13 (2) | |
| Missing | 18 (11) | 32 (12) | 4 (8) | 3 (6) | 12 (27) | 69 (12) | |
| Ethnicity—N (%) | |||||||
| Hispanic or Latino | 21 (13) | 32 (12) | 4 (8) | 5 (10) | 8 (18) | 70 (12) | 0.678 |
| Not Hispanic or Latino | 129 (80) | 226 (83) | 46 (88) | 43 (88) | 33 (75) | 477 (82) | |
| Missing | 12 (7) | 15 (5) | 2 (4) | 1 (2) | 3 (7) | 33 (6) | |
| Residence before admission—N (%) | |||||||
| Home independently | 128 (79) | 208 (76) | 40 (77) | 30 (61) | 22 (50) | 428 (74) | <0.001 |
| Home with help | 22 (14) | 52 (19) | 8 (15) | 12 (24) | 15 (34) | 109 (19) | |
| Rehab or nursing facility | 1 (1) | 5 (2) | 2 (4) | 5 (10) | 5 (11) | 18 (3) | |
| Other | 8 (5) | 8 (3) | 1 (2) | 0 (0) | 2 (5) | 19 (3) | |
| Not answered | 3 (2) | 0 (0) | 1 (2) | 2 (4) | 0 (0) | 6 (1) | |
| Baseline EQ-5D-5L—Median (IQR) | 0.877 (0.603–1.00) | 0.817 (0.572–0.943) | 0.815 (0.463–0.940) | 0.667 (0.287–0.904) | 0.419 (0.114–0.844) | 0.817 (0.507–0.943) | <0.001 |
| Limitations in I/ADLs—Median (IQR) | 0 (0–2) | 1 (0–3) | 1 (0–3) | 4 (1–7) | 4 (1–7) | 1 (0–4) | <0.001 |
| Baseline comorbidities—N (%) | |||||||
| Active malignancy | 0 (0) | 20 (8) | 2 (4) | 7 (15) | 8 (19) | 37 (7) | <0.001 |
| Diabetes mellitus | 25 (16) | 89 (34) | 11 (22) | 23 (48) | 19 (44) | 167 (30) | <0.001 |
| Hypertension | 39 (26) | 153 (58) | 31 (62) | 34 (71) | 24 (56) | 281 (50) | <0.001 |
| Heart failure | 3 (2) | 23 (9) | 6 (12) | 7 (15) | 1 (2) | 40 (7) | 0.006 |
| Chronic pulmonary disease | 18 (12) | 48 (18) | 11 (22) | 11 (23) | 7 (16) | 95 (17) | 0.269 |
| End stage renal disease | 3 (2) | 12 (5) | 0 (0) | 3 (6) | 1 (2) | 19 (3) | 0.281 |
| Primary cause of lung injury—N (%) | |||||||
| Pneumonia | 80 (49) | 160 (59) | 28 (54) | 37 (76) | 26 (59) | 331 (57) | 0.001 |
| Aspiration | 27 (17) | 40 (15) | 12 (23) | 8 (16) | 9 (20) | 96 (17) | |
| Sepsis | 15 (9) | 41 (15) | 4 (8) | 2 (4) | 8 (18) | 70 (12) | |
| Trauma | 19 (12) | 11 (4) | 0 (0) | 1 (2) | 0 (0) | 31 (5) | |
| Transfusion | 4 (2) | 6 (2) | 3 (6) | 0 (0) | 1 (2) | 14 (2) | |
| Other | 17 (10) | 15 (5) | 5 (10) | 1 (2) | 0 (0) | 38 (7) | |
| Admission duration, days—Median (IQR) | 19 (11–32) | 26 (19–39) | 10 (8–11) | 11 (10–13) | 38 (22–49) | 21 (13–35) | <0.001 |
| Baseline APACHE III score—Median (IQR) | 86 (70–105) | 100 (86–124) | 86 (72–107) | 102 (90–121) | 105 (86–116) | 96 (78–117) | <0.001 |
| Max modified SOFA—Median (IQR) | 6 (3–8) | 6 (4–9) | 4 (3–6) | 5 (4–7) | 7 (5–11) | 6 (4–8) | <0.001 |
| Nadir PaO2:FiO2—Median (IQR) | 107 (87–130) | 113 (83–135) | 116 (91–145) | 114 (80–131) | 112 (80–138) | 112 (85–135) | 0.042 |
Cumulative all-cause mortality differed by survivor subtype at all time points through 12 months (e-Figure 2). In unadjusted logistic regression using the low risk subtype as the referent, mortality was significantly associated with subtype at all time points (group tests of significance p=0.003 at three, p=0.001 at six, and p<0.001 at twelve months; Table 2). The association with mortality was most different from the low risk patients among patients in the unhealthy baseline with severe illness group, who experienced a 4.9-fold higher risk of death at 12 months (interquartile range [IQR] 2.1–11.5). Including age as a covariate attenuated the association between subtype and mortality such that only 12-month mortality was significantly associated with subtype (p=0.027).
Table 2.
Unadjusted and age-adjusted odds of mortality and readmission for transported subtype (Panel A) and de novo subtype (Panel B) estimated using logit
| Subtype | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Low Risk | Healthy baseline, severe illness | Multimorbidity | Low functional status | Unhealthy baseline, severe illness | Grouptest of significance | ||||||
| Unadjusted | Age-Adjusted | Unadjusted | Age-Adjusted | Unadjusted | Age-Adjusted | Unadjusted | Age-Adjusted | Unadjustec Age-Adjusted | |||
| A. Transported subtype | |||||||||||
| Mortality | |||||||||||
| 3 months | Referent | 3.3 (1.5–7.2) | 1.2 (0.4–3.6) | 0.4 (0.1–3.1) | 0.2 (0.0–1.4) | 4.3 (1.6–11.9) | 1.4 (0.4–5.3) | 5.0 (1.8–14.2) | 2.1 (0.6–7.2) | 0.0026 | 0.1713 |
| 6 months | 2.7 (1.4–5.2) | 1.2 (0.5–3.2) | 0.5 (0.1–2.3) | 0.2 (0.1–1.3) | 4.1 (1.7–10.0) | 1.7 (0.5–5.4) | 4.5 (1.8–11.2) | 2.3 (0.8–6.7) | 0.0007 | 0.0638 | |
| 12 months | 3.2 (1.7–5.9) | 2.0 (0.9–4.8) | 0.9 (0.3–2.8) | 0.6 (0.2–2.1) | 4.1 (1.8–9.6) | 2.5 (0.8–7.4) | 4.9 (2.1–11.5) | 3.3 (1.2–9.0) | 0.0001 | 0.0266 | |
| Readmission | |||||||||||
| 3 months | Referent | 2.2 (1.3–3.7) | 2.4 (1.1–5.1) | 0.9 (0.4–2.0) | 0.9 (0.3–2.4) | 1.3 (0.6–3.1) | 1.4 (0.5–3.9) | 2.0 (0.8–4.8) | 2.1 (0.8–5.9) | 0.0104 | 0.0426 |
| 6 months | 1.9 (1.2–3.0) | 1.8 (0.8–3.7) | 0.7 (0.3–1.5) | 0.7 (0.3–1.7) | 1.6 (0.7–3.7) | 1.5 (0.6–4.1) | 2.0 (0.8–4.6) | 1.9 (0.7–4.9) | 0.0192 | 0.0811 | |
| 12 months | 1.8 (1.1–2.8) | 2.2 (1.0–4.7) | 1.6 (0.8–3.2) | 1.9 (0.8–4.8) | 2.2 (1.0–4.9) | 2.7 (1.0–7.7) | 1.1 (0.5–2.7) | 1.4 (0.5–3.9) | 0.1148 | 0.2796 | |
| B. De novo subtype | |||||||||||
| Mortality | |||||||||||
| 3 months | Referent | 0.6 (0.1–3.0) | 0.4 (0.1–2.2) | 2.6 (0.9–7.6) | 1.3 (0.3–5.2) | 9.3 (3.4–25.3) | 5.1 (1.5–17.8) | 6.6 (2.2–19.2) | 3.9 (1.1–13.9) | <0.0001 | <0.0001 |
| 6 months | 0.8 (0.2–2.9) | 0.6 (0.2–2.3) | 2.4 (0.9–6.0) | 1.4 (0.4–4.7) | 9.1 (3.8–21.9) | 5.7 (1.9–17.2) | 6.4 (2.5–16.4) | 4.2 (1.4–12.9) | <0.0001 | <0.0001 | |
| 12 months | 1.0 (0.3–2.9) | 0.9 (0.3–2.9) | 2.9 (1.3–6.5) | 2.6 (0.9–7.6) | 9.3 (4.3–20.4) | 8.5 (3.1–23.1) | 6.3 (2.7–14.8) | 5.8 (2.2–15.9) | <0.0001 | <0.0001 | |
| Readmission | |||||||||||
| 3 months | Referent | 1.2 (0.6–2.4) | 1.5 (0.7–3.4) | 2.4 (1.3–4.4) | 3.5 (1.4–8.6) | 2.7 (1.4–5.2) | 3.8 (1.5–9.4) | 3.1 (1.4–6.6) | 4.1 (1.7–10.2) | 0.0034 | 0.0082 |
| 6 months | 1.2 (0.6–2.2) | 1.3 (0.6–2.8) | 2.1 (1.2–3.7) | 2.5 (1.0–6.1) | 1.9 (1.0–3.7) | 2.2 (0.9–5.4) | 2.5 (1.2–5.2) | 2.8 (1.2–6.8) | 0.0265 | 0.1108 | |
| 12 months | 1.7 (0.9–3.2) | 2.1 (1.0–4.6) | 1.9 (1.1–3.3) | 2.6 (1.0–6.8) | 1.7 (0.9–3.3) | 2.3 (0.8–6.1) | 2.0 (1.0–4.3) | 2.6 (1.0–6.6) | 0.1259 | 0.3071 | |
Measures are odds ratio (95% confidence interval). P-values calculated by Wald test.
Among secondary endpoints, self-reported readmission was significantly associated with survivor subtype at three (p=0.010) and six (p=0.019) months, but not at 12 months. Adjusting for age attenuated this association such that only 3-month admission showed a significant association (p=0.043). Survivor subtype was significantly associated with I/ADL limitations and EQ-5D-5L scores at three months, but the effect diminished by twelve and six months, respectively (Table 3). Sensitivity analyses did not reveal any new patterns in the association between subtype and mortality when stratifying by presence of sepsis or pneumonia, or by treatment arm (e-Table 4).
Table 3.
Patient-reported outcomes after discharge according to transported subtype
| Patient Subtype | All | p | |||||
|---|---|---|---|---|---|---|---|
| Low Risk | Healthy baseline, severe illness | Multimorbidity | Low functional status | Unhealthy baseline, severe illness | |||
| Limitations in I/ADLs | |||||||
| 3 months | 3 (1, 5) | 5 (2, 8) | 2 (0, 5) | 4 (1, 9) | 6 (3, 9) | 4 (1, 7) | <0.001 |
| 6 months | 2 (0, 6) | 4 (1, 6) | 1 (0, 6) | 3 (1, 8) | 3 (1, 5) | 3 (1, 6) | 0.006 |
| 12 months | 2 (0, 5) | 4 (1, 6) | 3 (0, 6) | 2 (1, 8) | 3 (2, 5) | 3 (1, 6) | 0.342 |
| Change in I/ADL limitations from baseline | |||||||
| 3 months | 1 (0, 3) | 3 (1, 5) | 1 (0, 2) | 1 (0, 2) | 2 (−1, 4) | 2 (0, 4) | <0.001 |
| 6 months | 1 (−1, 2) | 1 (0, 4) | 0 (−1, 1) | 1 (−1,3) | 0 (−3, 2) | 1 (−1, 3) | 0.023 |
| 12 months | 0 (0, 3) | 1 (0, 4) | 1 (0, 3) | 0 (−1, 1) | 0 (−3, 2) | 1 (0, 3) | 0.017 |
| EQ-5D-5L index value | |||||||
| 3 months | 0.69 (0.39, 0.88) | 0.52 (0.13, 0.75) | 0.80 (0.47, 0.93) | 0.71 (0.11, 0.90) | 0.42 (−0.07, 0.71) | 0.60 (0.23, 0.84) | <0.001 |
| 6 months | 0.76 (0.49, 0.89) | 0.62 (0.31, 0.84) | 0.80 (0.44, 0.92) | 0.71 (0.38, 0.88) | 0.65 (0.30, 0.84) | 0.69 (0.39, 0.87) | 0.113 |
| 12 months | 0.72 (0.42, 0.92) | 0.64 (0.35, 0.86) | 0.61 (0.29, 0.88) | 0.70 (0.23, 0.94) | 0.70 (0.28, 0.84) | 0.68 (0.35, 0.88) | 0.662 |
| Presence of cognitive impairment—N (%) | |||||||
| 3 months | 48 (52) | 113 (73) | 30 (67) | 21 (84) | 10 (67) | 222 (67) | 0.004 |
| 6 months | 38 (45) | 104 (66) | 24 (56) | 16 (80) | 13 (65) | 195 (60) | 0.007 |
| 12 months | 49 (59) | 101 (70) | 26 (68) | 16 (73) | 15 (79) | 207 (67) | 0.350 |
All measures are median (IQR) unless stated otherwise
De novo latent class analysis to identify survivor subtypes
In exploratory analyses, the best-performing model incorporated all four continuous variables after log-transformation due to non-normal distributions and identified five survivor subtypes. 33 patients were excluded because they were missing data on prior hospital days in the last month (these were not excluded in the decision tree because they had already been assigned to a different group and thus did not reach a node that required this data). Varying the number of survivor subtypes from two to ten, we observed that both AIC and BIC decreased consistently, with a greater step-off observed in both criteria going from four to five subtypes as compared to other variations. Five classes also yielded the highest Vuong-Lo-Mendell-Rubin test statistic while maintaining acceptable entropy and group sizes (e-Table 2). Increasing the number of groups to six resulted in the generation of a group with less than 5% of the population (26/547, 4.7%) with only marginal improvements in AIC and BIC and a significant reduction in the Vuong-Lo-Mendell-Rubin test statistic. The final 5-class model had an AIC of 4,771, BIC of 4,892, and entropy of 0.806.
We conducted a CART analysis to develop an operationalized assignment algorithm that reproduced these classifications, and could be used prospectively afterwards; assignments based in the LCA and from the CART classification agreed (kappa = 0.947, p<0.001).
Association of de novo subtypes with outcomes
In unadjusted and age-adjusted analysis, the directionality and effect size of the associations between key outcomes and de novo subtype were similar to those estimated using transported subtypes (e-Figure 3). Two of the four subtypes demonstrated significantly increased odds of mortality at all time points through 12 months (Table 4). In the highest risk group compared to the low risk group, the odds ratio for mortality after adjusting for age were 5.1 at three months, 5.7 at six months, and 8.5 at 12 months (group test of significance p<0.001 for all). Significant but less extreme associations were estimated between subtype and self-reported post-ARDS hospital admission after age-adjustment at three months only (group tests of significance p=0.008 for three, p=0.111 for six, and p=0.307 for twelve months). The highest risk group compared to the low risk group had an odds ratio for self-reported post-ARDS hospital admission of 4.1 at three months
Table 4.
Model performance for outcomes of death and readmission through 12 months, according to method of subtype assignment (transported vs de novo)
| Transported | De novo | |||
|---|---|---|---|---|
| p-value | AIC | p-value | AIC | |
| Death | ||||
| 3 month | 0.1713 | 395.6 | <0.0001 | 357.9 |
| 6 month | 0.0638 | 468.5 | <0.0001 | 420.9 |
| 12 month | 0.0266 | 533.4 | <0.0001 | 508.6 |
| Readmission | ||||
| 3 month | 0.0426 | 550.8 | 0.0082 | 549.5 |
| 6 month | 0.0811 | 575.8 | 0.1108 | 575.1 |
| 12 month | 0.2796 | 560.4 | 0.3071 | 556.8 |
Models are logit models with facility level random effects, adjusted for age
LCA: latent class analysis; AIC: Akaike information criterion
Comparing de novo to transported subtypes
Spider plots comparing variable distribution according to subtype assignment method are found in Figure 4. Agreement between subtype according to the method of assignment (transported vs de novo) was greater than would be expected from chance alone (kappa=0.361, p<0.001; e-Table 5). Agreement was highest in the low risk (134/145, 92.4%) and unhealthy baseline, severe illness (44/44, 100%) groups. Models fit using de novo subtype outperformed transported subtype for mortality at all time points through 12 months, as measured by Wald group test of significance and AIC (Table 4, e-Figure 3).
Figure 4.
Spider plot showing distribution of the four deterministic variables among transported (left) and de novo subtypes (right). All measures z-scaled.
Discussion
Operationalizing sepsis survivor subtypes in an ARDS population is possible and yields non-unique assignment algorithms which nonetheless display predictive validity. Both (a) transporting from the published latent class analysis, and (b) de novo development of a new latent class analysis using the same variables yielded subtypes that had associations with patient-important outcomes through 12 months of follow-up. Greater agreement between operationalization strategies was seen in classification to the two survivor subtypes at either extreme end of risk, with significant disagreement between the two approaches in the intermediate risk groups.
The association between transported subtype and mortality suggests that operationalizing the survivor subtypes developed by Taylor in a population with sepsis also allowed identification of groups of ARDS patients who might benefit from more intensive support after discharge, and who could be readily identified with a few simple data elements. That a de novo latent class analysis was optimally fit with the same number of classes, demonstrated agreement in the lowest and highest risk groups, and yielded similar patterns of association supports the reality of latent phenotypes that persist across different but related critical care conditions.
Compared to Taylor’s derivation study in a large real-world population, our transported subtypes in a recruited and consented RCT population had differing prevalences of groups defined by a high burden of comorbidities, specifically the multimorbidity (8% vs 28%), low function status, (8% vs 14%), and unhealthy baseline with severe illness groups (8% vs 21%). Using de novo subtypes led to better alignment, however the unhealthy baseline with severe illness group still had a notably lower proportion of patients compared to Taylor (13% vs 21%). This may be because our study used data from a clinical trial that is less likely to enroll participants with a higher number of comorbidities, whereas Taylor’s work used all-hospitalization data18—although it may also be because our subtype operationalizations did not perfectly align with Taylor’s subtypes.
The clinical implication of these results is one of transportability. Our findings raise the possibility that the Taylor subtypes are not strictly bound by the specific syndrome and dataset they were developed in and could have useful applications in broader critical care syndromes. A simple algorithm with straightforward inputs would allow rapid stratification of a patient belonging to a particularly high risk group and provide an opportunity to arrange more robust support before the patient is discharged if the health system faces scarce resources and cannot provide maximal support to all. An important advantage is that assignment to a subtype can prompt services tailored to the characteristics and risk factors that define each subtype. For example, a testable hypothesis is that the poor functional status subtype (characterized by older age and frequent discharge to a facility) could benefit from geriatrics referral and frequent follow-up by phone in anticipation of eventually transitioning back to their home, whereas the previously healthy with severe illness subtype could preferentially benefit from post-intensive care unit clinics or intensive rehabilitation to achieve their prior level of functioning. Depending on the resources of the health system, the nature of support will vary. At a minimum, subtype assignment can help inform families of the likelihood and type of enduring needs to help them plan.
However, these findings must be interpreted in context of important limitations. First and most importantly, there is no simple, universally accepted methodology to translate a descriptive LCA and apply it to a new patient population. We used two approaches, one couched in our team’s collective clinical expertise and the other relying on empirical data and statistical concepts but requiring more up-front data on the population to which the operationalized algorithm would be applied. The imperfect concordance between these two approaches to assigning subtype is an important area for future research, and our data alone cannot distinguish whether this limitation stems from the conceptual model, the limited overlap in the data concepts available in the derivation data versus ROSE, or differences stemming from either of our two strategies to operationalize the Taylor survivor subtypes.
While promising, a further limitation of our results is that they cannot confirm nor refute that the groups we identified are the same as those identified by Taylor; latent classes are by definition latent and no perfect gold standard for classification exists. This means that other operationalizations of the Taylor survivor subtypes are possible, particularly when different variables are available. To the extent that our operationalizations—transported or de novo—misclassify survivors relative to some other more optimal classification, that might be expected to bias towards a null association with outcomes, unlike what we saw. While high fidelity reproduction of each subtype may be ideal, the fact that subtypes diverged significantly in their association with patient-important outcomes despite differences in data suggests the Taylor survivor subtypes offer useful information for post-discharge planning. We also note that when considering patient-centered outcomes, the quite high competing risk of mortality—different across different survivor subtypes—may blunt measurable differences. Finally, the statistical approach for our LCA did not adjust for bias due to classification error.19
In summary, we used two different approaches to operationalizing the Taylor sepsis survivor subtypes in a population of patients with ARDS, both of which assigned individuals to groups that differed significantly in the odds of 12-month mortality along with other patient-important outcomes. Our findings provide evidence of the clinical utility and transportability of the Taylor sepsis survivor subtypes, and demonstrate approaches to translating insights garnered from an LCA into clinical practice.
Supplementary Material
Funding information:
R.J.F.: NHLBI T32HL007534
T.S.V.: NHLBI R01HL157361 (PRECIPICE)
M.A.-H.: NHLBI R01HL157361 (PRECIPICE)
T.J.I.: NHLBI R01HL132232 (PRIMROSE) and R01HL169533 (EXPERT)
Abbreviation list:
- ARDS
acute respiratory distress syndrome
- ROSE
Reevaluation of Systemic Early Neuromuscular Blockade
- I/ADLs
Instrumental Activities of Daily Living
- SOFA
Sequential Organ Failure Assessment
- APACHE
Acute Physiology, Age, Chronic Health Evaluation
Footnotes
Conflict of interest statement: no conflicts of interest for all authors
Prior presentation: preliminary data was presented at the American Thoracic Society 2024 International Conference, San Diego, USA, May 17–22 2024
References
- 1.Needham DM, Colantuoni E, Mendez-Tellez PA, et al. Lung protective mechanical ventilation and two year survival in patients with acute lung injury: prospective cohort study. BMJ. 2012;344:e2124. doi: 10.1136/bmj.e2124 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.The National Heart, Lung, and Blood Institute PETAL Clinical Trials Network. Early Neuromuscular Blockade in the Acute Respiratory Distress Syndrome. N Engl J Med. 2019;380(21):1997–2008. doi: 10.1056/NEJMoa1901686 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Ruhl AP, Huang M, Colantuoni E, et al. Healthcare utilization and costs in ARDS survivors: a 1-year longitudinal national US multicenter study. Intensive Care Med. 2017;43(7):980–991. doi: 10.1007/s00134-017-4827-8 [DOI] [PubMed] [Google Scholar]
- 4.Calfee CS, Delucchi K, Parsons PE, et al. Subphenotypes in acute respiratory distress syndrome: latent class analysis of data from two randomised controlled trials. Lancet Respir Med. 2014;2(8):611–620. doi: 10.1016/S2213-2600(14)70097-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Prescott HC, Calfee CS, Thompson BT, Angus DC, Liu VX. Toward Smarter Lumping and Smarter Splitting: Rethinking Strategies for Sepsis and Acute Respiratory Distress Syndrome Clinical Trial Design. Am J Respir Crit Care Med. 2016;194(2):147–155. doi: 10.1164/rccm.201512-2544CP [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Taylor SP, Rios A, Kowalkowski MA. Translating Postsepsis Care to Post-COVID-19 Care. The Case for a Virtual Recovery Program. Ann Am Thorac Soc. 2021;18(6):938–941. doi: 10.1513/AnnalsATS.202006-649IP [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Taylor SP, Bray BC, Chou SH, Burns R, Kowalkowski MA. Clinical Subtypes of Sepsis Survivors Predict Readmission and Mortality after Hospital Discharge. Ann Am Thorac Soc. 2022;19(8):1355–1363. doi: 10.1513/AnnalsATS.202109-1088OC [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Huang DT, Angus DC, Moss M, et al. Design and Rationale of the Reevaluation of Systemic Early Neuromuscular Blockade Trial for Acute Respiratory Distress Syndrome. Ann Am Thorac Soc. 2017;14(1):124–133. doi: 10.1513/AnnalsATS.201608-629OT [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Lazarsfeld PF, Henry NW. Latent Structure Analysis. Houghton, Mifflin; 1968. [Google Scholar]
- 10.Formann AK, Kohlmann T. Latent class analysis in medical research. Stat Methods Med Res. 1996;5(2):179–211. doi: 10.1177/096228029600500205 [DOI] [PubMed] [Google Scholar]
- 11.Sinha P, Delucchi KL, McAuley DF, O’Kane CM, Matthay MA, Calfee CS. Development and validation of parsimonious algorithms to classify acute respiratory distress syndrome phenotypes: a secondary analysis of randomised controlled trials. Lancet Respir Med. 2020;8(3):247–257. doi: 10.1016/S2213-2600(19)30369-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Hashem MD, Hopkins RO, Colantuoni E, et al. Six-month and 12-month patient outcomes based on inflammatory subphenotypes in sepsis-associated ARDS: secondary analysis of SAILS-ALTOS trial. Thorax. 2022;77(1):22–30. doi: 10.1136/thoraxjnl-2020-216613 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Weller BE, Bowen NK, Faubert SJ. Latent Class Analysis: A Guide to Best Practice. J Black Psychol. 2020;46(4):287–311. doi: 10.1177/0095798420930932 [DOI] [Google Scholar]
- 14.Sinha P, Calfee CS, Delucchi KL. Practitioner’s Guide to Latent Class Analysis: Methodological Considerations and Common Pitfalls. Crit Care Med. 2021;49(1):e63–e79. doi: 10.1097/CCM.0000000000004710 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Herdman M, Gudex C, Lloyd A, et al. Development and preliminary testing of the new five-level version of EQ-5D (EQ-5D-5L). Qual Life Res Int J Qual Life Asp Treat Care Rehabil. 2011;20(10):1727–1736. doi: 10.1007/s11136-011-9903-x [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Gerrard P The hierarchy of the activities of daily living in the Katz index in residents of skilled nursing facilities. J Geriatr Phys Ther 2001. 2013;36(2):87–91. doi: 10.1519/JPT.0b013e318268da23 [DOI] [PubMed] [Google Scholar]
- 17.Therneau T, Atkinson B. rpart: Recursive Partitioning and Regression Trees. Published online April 8, 1999:4.1.24. doi: 10.32614/CRAN.package.rpart [DOI] [Google Scholar]
- 18.Rogers JR, Liu C, Hripcsak G, Cheung YK, Weng C. Comparison of Clinical Characteristics Between Clinical Trial Participants and Nonparticipants Using Electronic Health Record Data. JAMA Netw Open. 2021;4(4):e214732. doi: 10.1001/jamanetworkopen.2021.4732 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Bolck A, Croon M, Hagenaars J. Estimating Latent Structure Models with Categorical Variables: One-Step Versus Three-Step Estimators. Polit Anal. 2004;12(1):3–27. doi: 10.1093/pan/mph001 [DOI] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.