Predicting Prolonged Intensive Care Unit Stay Among Patients With Sepsis-Induced Hypotension

We performed a retrospective secondary analysis of 2 prospective study cohorts. Both studies were approved by the institutional review board and adhered to the Strengthening the Reporting of Observational Studies in Epidemiology guidelines.¹⁰ 

The first study was conducted at a large, urban, academic ED in the northeastern United States with 55 000 annual visits. Researchers in that study prospectively enrolled all adult (age 18 years or older) ED patients with persistent hypotension, defined as a systolic blood pressure of less than 90 mm Hg after resuscitation with at least 1 L of intravenous fluids; a vasopressor requirement; or hypotension with intravenous fluid restriction because of a documented concern of fluid overload (ie, patients with congestive heart failure or dependent on hemodialysis). The study excluded patients with a documented baseline systolic blood pressure of less than 90 mm Hg, unless a 10 mm Hg decrease in systolic blood pressure occurred, and those who were discharged from the ED.¹¹  The second study was conducted in 2 urban, academic EDs in the US Pacific Northwest with 90 000 combined annual visits, and it retrospectively included all patients meeting severe sepsis criteria.¹² 

The current study leveraged the 2 original data sets to create a cohort of infected patients with at least 1 systolic blood pressure reading of less than 90 mm Hg after administration of at least 1 L of intravenous fluids documented during the ED stay. All patients with nonelective intubation, vasopressor administration, and/or death in the ED were excluded because these patients have obvious requirements for ICU-level care. Patients determined by investigator adjudication to have sepsis as the underlying cause of shock in the emergency department were included for analysis.

The screening and verification process has been described for both studies.^11,12  Medical history and the presence of altered mental status were manually abstracted from the hospital records by using a standard data collection form. Altered mental status included any mention in the ED record of altered mentation, confusion, or somnolence. Abstraction was performed by trained research assistants in the first study and by a senior emergency medicine resident in the second, all under the direct supervision of the principal investigator. Basic demographics, vital signs, length of stay, disposition, and laboratory values were obtained from the electronic health records. The primary outcome was ICU stay of 24 hours or longer or death occurring less than 24 hours after ICU admission.

We performed data analysis with statistics software (SAS 9.3, SAS Institute Inc). For continuous covariates, we used the Student t test and the Wilcoxon rank sum test. We created a multivariable logistic regression model to predict ICU stay of 24 hours or longer. The univariate relationships between demographics, vital signs, and laboratory study cutoffs between ICU length of stay groups were assessed with the χ² test for binary variables and the Student t test for continuous variables. We included covariates with P less than .10 in univariate analyses as candidate variables for the logistic regression model. We included age and serum lactate level as continuous variables and systemic inflammatory response syndrome criteria and severe sepsis organ dysfunctions as binary candidate variables a priori. We built a stepwise selection model with P less than .05 required for entry and P less than .10 required to stay in the model. The rule of 1 variable per 10 events determined the maximum number of model covariates. We assessed model discrimination with area under the curve (AUC) and model calibration with the Hosmer-Lemeshow test. We performed univariate logistic regressions for continuous qSOFA score and serum lactate level to predict ICU stay of 24 hours or longer. For the sensitivity analysis, we generated univariate logistic regression models predicting mortality using (1) any of the derived model’s predictors of ICU stay of less than 24 hours, (2) stratified lactate concentrations (normal, < 2.0 mmol/L; intermediate, ≥ 2.0 mmol/L and < 4.0mmol/L; high, ≥ 4.0 mmol/L [divide lactate values by 0.111 to convert to mg/dL]; missing lactate values were imputed as an independent covariate), and (3) qSOFA score of 2 or less.

The 2 original studies involved 437 patients with infectious causes of disease. We included 233 patients in our analysis (see Figure).

Age, sex, most comorbidities, and do-not-resuscitate status were similar between the 2 patient groups (those with ICU stays < 24 hours and those with ICU stays ≥ 24 hours). Heart failure was more common among patients with ICU stays of 24 hours or longer, and chemotherapy was more common among patients with ICU stays of less than 24 hours (Table 1). Other than respiratory rate, initial vital signs were similar in the 2 groups (Table 2). Altered mental status was present in 22 of 125 patients with ICU stays of less than 24 hours (17.6%; 95% CI, 11.4%-25.4%) and in 30 of 108 patients with ICU stays of 24 or more hours (27.8%; 95% CI, 19.6%-37.2%). The mean (95% CI) ICU stays were 0.3 (0.2-0.4) days for patients with ICU stays of less than 24 hours and 3.4 (2.9-3.9) days for those with ICU stays of 24 hours or longer.

Twenty-four in-hospital deaths occurred (10.3% of patients; 95% CI, 7%-15%), all in the cohort of patients with ICU stays of 24 hours or longer. Vasopressor support was required in 6 of 125 patients with ICU stays of less than 24 hours (4.8%; 95% CI, 2%-10%) and in 25 of 108 patients with ICU stays of 24 hours or longer (23.1%; 95% CI, 16%-32%). Nonelective intubation occurred in 3 patients with ICU stays of less than 24 hours (2.4%; 95% CI, 1%-7%) and in 10 patients with ICU stays of 24 hours or more (9.3%; 95% CI, 5%-16%).

Of the 233 included patients, 108 (46.4%) had ICU stays of 24 hours or longer. Multivariable logistic regression identified heart failure history, bicarbonate level less than 20 mEq/L, respiratory rate greater than 20 breaths per minute, and creatinine level greater than 2.0 mg/dL as independent positive predictors of ICU stay of 24 hours or longer (Table 3). Creatinine level greater than 2.0 mg/dL remained an independent predictor (P < .01) when the covariate chronic kidney disease (chronic kidney disease or end-stage renal disease) was included in the model (Table 4), although a history of chronic kidney disease or end-stage renal disease was not an independent predictor (P = .97). No covariates were predictors of ICU stay of less than 24 hours. Of 108 patients with ICU stays of 24 hours or longer, 89 (sensitivity, 82.4%; 95% CI, 73.9%-89.1%) exhibited at least 1 of the 4 model predictors (positive likelihood ratio = 1.64). Of 125 patients without ICU stays of 24 hours or longer, 62 (specificity, 49.6%; 95% CI, 40.5%-58.7%) exhibited none of the model predictors (negative likelihood ratio = 0.35). Conversely, qSOFA score of 2 or greater had a sensitivity of 52.8% and a specificity of 65.6% for identifying patients with ICU stays of 24 hours or longer.

We compared the derived predictive factors for ICU stay of 24 hours or longer with the predictive performance of initial lactate concentration and qSOFA score. Each 1.0 mmol/L increase in serum lactate had an odds ratio of 1.2 (95% CI, 1.0-1.4) among patients with ICU stays of 24 hours or longer (AUC, 0.62; 95% CI, 0.55-0.69). Each 1-point increase in qSOFA had an odds ratio of 1.9 (95% CI, 1.3-3.0) for the primary outcome (AUC, 0.60; 95% CI, 0.54-0.66). Compared with qSOFA score and lactate level as continuous variables, the study model improved the prediction of ICU stay of 24 hours or longer (AUC, 0.74; 95% CI, 0.66-0.79). In the sensitivity analysis, the presence of a single model covariate (AUC, 0.66; 95% CI, 0.60-0.72) performed similarly to stratified lactate level (AUC, 0.60; 95% CI, 0.54-0.67) and qSOFA score of 2 or greater (AUC, 0.59; 95% CI, 0.52-0.66).

This study sought clinical factors to predict ICU stay of 24 hours or longer among infected ED patients with hypotension. Our model, calibrated to ICU length of stay, outperformed both continuous serum lactate level and qSOFA score, suggesting that these widely known tools may not provide the most informative risk stratification when considering hospital resource needs. When applied clinically, the presence of a single model predictor demonstrated improved sensitivity for ICU stay of 24 hours or longer as compared with 2 or more qSOFA criteria. When predictors are present, early admission may be appropriate because prolonged ED stays (> 5 hours) are associated with higher rates of ICU complications.¹³  Yet in this exploratory study neither our derived model nor existing tools adequately identified patients who required ICU stays of 24 hours or longer, suggesting that further investigation is needed to assist clinicians with disposition among this subset of patients with sepsis.

Several studies have attempted to use biomarkers and adapt illness severity scores, which were derived to predict mortality, to predict ICU length of stay or ICU admission. A recent attempt to repurpose mortality-derived illness severity scores for modeling predicted ICU length of stay showed promising results.¹⁴  However, the scoring systems used require variables not routinely captured in the ED, and the study populations were not specific to sepsis. More commonly, illness severity scores have been tested to predict ICU admission among patients without clear ICU needs (eg, those receiving vasopressors or mechanical ventilation), but the performance of scores more amenable to ED use is highly variable. The Pneumonia Severity Index and the CURB/CURB-65 (confusion, blood urea nitrogen, respiratory rate, blood pressure, age) criteria poorly predict ICU admission.¹⁵  Among elderly patients with sepsis, the SOFA score (AUC, 0.93) and abbreviated Mortality in Emergency Department Sepsis score (AUC, 0.95) were strongly predictive of ICU admission.¹⁶  However, other research suggests a more modest effect of SOFA score (AUC, 0.73) in predicting ICU transfer within 48 hours.¹⁷ 

Biomarkers in isolation perform marginally and are insufficient to predict ICU needs. For example, procalcitonin level had an AUC of 0.69 for predicting the need for vasopressor or mechanical ventilation within 72 hours of presentation¹⁸  and is modestly outperformed by American Thoracic Society and Infectious Diseases Society of America guidelines.¹⁹  Although a serum lactate level of greater than 4 mmol/L in patients with sepsis is a known risk factor for progression to shock and is associated with ICU admission, lactate level alone is not associated with ICU length of stay.²⁰  We report a similar inadequacy of serum lactate level for predicting ICU stay of 24 hours or longer. Although recent attempts to combine comorbidities and bio-markers have augmented mortality prognostication for patients with sepsis,²¹  highly predictive models for ICU admission or length of stay remain elusive.

We found that qSOFA score had inadequate sensitivity and specificity for ICU stay of 24 hours or longer in this population. Although qSOFA was validated with a combination of mortality and ICU stay of longer than 3 days,²²  studies specifically regarding short ICU stay and admission disposition have not been conducted previously. The shortcomings of illness severity models, like qSOFA, are most likely related to their origin; ICU length of stay is an unintended end point for scoring systems derived to predict mortality. Calibrating predictive models for additional clinically relevant end points such as ICU stay of 24 hours or longer may be more helpful than repurposing other tools in risk-stratifying patients with sepsis.

This report serves as a proof of principle after leveraging data sets created for a different purpose. Therefore, this secondary analysis is missing covariates that may be important for development of a clinical decision rule ready for clinical use, such as unmeasured patient, provider, and hospital-level factors that interact with ICU length of stay. We were not able to account for bed availability, although demand elasticity can affect admission disposition.⁴ 

Intensive care unit length of stay is an important, albeit imperfect, outcome of interest in estimating illness severity from the standpoint of ED disposition and hospital resource management. Other measures, including ICU length of stay assessed in fractions of a day, may yield more useful results, although that level of detail was not available in the data sets used. Nevertheless, using a threshold of 24 hours provides a clinically relevant outcome that clinicians may readily apply to predict inpatient resource requirements. Several implicit confounders can obscure interpretation because medical, social, psychological, and institutional factors affect ICU length of stay. For example, the delay between bed request and bed availability could be misleading because the prolongation of ICU stay reflects hospital limitations, not clinical instability.

In summary, a history of heart failure, serum bicarbonate level less than 20 mEq/L, tachypnea, or creatinine level greater than 2 mg/dL modestly increases the likelihood of ICU stay of 24 hours or longer in hypotensive ED patients with sepsis. Our results may have some clinical utility for patients with the right pretest probability, but neither our model nor established tools for risk stratification currently provide adequate guidance regarding the disposition of patients who have hypotension in the context of sepsis. Our predictors represent a preliminary framework deserving further evaluation because ED disposition for sepsis is a particularly relevant outcome of interest for emergency medicine providers.