Data sources

We utilized a large electronic healthcare database, eICU (https://eicu-crd.mit.edu/), including ≥139,000 unique patients admitted to different Intensive Care Units (ICUs) at 208 hospitals distributed across the US from 2014 to 2015 [22]. The eICU database includes comprehensive demographic and clinical data throughout patient’s hospital stays including, admission and discharge dates to the hospital and the ICU, vital signs, laboratory tests, prescribed and administered medications, and medical diagnoses with past medical history during admissions. These data allow us to study this research question because sepsis-3 definitions [16] require patients’ vital signs and laboratory data to calculate SOFA scores. The eICU search strategies are shown in the S1–S10 Tables.

Study design and patient population

In order to compare the effectiveness of corticosteroids in patients with sepsis defined by different definitions, we applied two algorithms to identify patients with sepsis. First, we explicitly defined sepsis based on sepsis ICD-9 (995.92; 785.52) or ICD-10 (R65.20; R65.21) diagnoses documented by clinicians during the ICU stay, because a validation study showed these ICD-9 codes had a PPV of 100% for identifying patients with severe sepsis [23]. To our knowledge, no validation study is yet available for the ICD-10 codes; however, these codes largely mirror the ICD-9 codes in eICU. We defined the time of sepsis diagnosis as the onset of sepsis. Because Sepsis-3 was published in 2016, and eICU only included patients admitted to the hospitals from 2014–2015 [22], we assumed that all sepsis-related diagnoses in eICU were based on the old definition [13, 14].

Second, we used explicit sepsis ICD codes combined with a total SOFA score ≥2 prior to the onset of sepsis as a proxy of the Sepsis-3 definition [16]. In this definition, all patients with suspected infection and an acute change in their total SOFA score ≥2 would be identified as having sepsis. However, the lack of complete body fluid cultures in the eICU data affected our ability to accurately define the onset of infection. We therefore used the time of sepsis diagnosis as a surrogate because patients with sepsis must have had an onset of infection prior to the onset of sepsis; the highest total SOFA score was assessed within 48 hours before sepsis onset (S1 Fig). To capture acute changes in total SOFA score, some subcategories were adjusted accordingly for patients who had a medical history of liver, renal, or respiratory failure. We included all patients with total SOFA score ≥2 within that 48-hour assessment window.

The index time was defined as the onset of sepsis (the time of sepsis diagnosis) for the two cohorts (S1 Fig). All patients were followed until discharge from the hospital (dead or alive). For all definitions, we excluded patients who: were <18 years, received any intravenous corticosteroids or etomidate before the onset of sepsis, had a history of long-term corticosteroids or other immunosuppressive drug use, or with HIV/AIDS, autoimmune disease, receiving cancer therapy, hematological or metastatic cancer, or organ transplantation. These criteria mimic enrollment criteria for the prior RCTs [11, 12] and are precautions or contraindications to intravenous corticosteroids or indications of previous receipt [24, 25].

Exposures

In order to mimic the corticosteroid treatments used in RCTs [4, 11, 26, 27], patients with sepsis who met the following two exposure criteria were classified as treated: (1) received any intravenous corticosteroids ≥24 hours continuously within 72 hours after the onset of sepsis and (2) had a treatment duration of ≤11 days. Those patients who did not meet the exposure criteria in the same exposure window were identified as controls. While an active comparator would be ideal, there is no real-world active comparator, so to mitigate this, we did sensitivity analyses comparing patients with different steroid doses. We handled immortal time bias [28] by excluding all patients who died within 96 hours after onset of sepsis (S1 Fig) and conducted several sensitivity analyses varying this exclusion.

Outcomes

The outcome of interest was in-hospital survival (which is the survival probability, and we used the term “survival” in the whole manuscript) up to 50 days beginning with the index time. Those patients with longer lengths of hospitalization had their follow-up administratively censored at day 50 after index. We chose 50 days of follow-up not only because the mean length of hospital stay among patients with sepsis from the RCTs [11, 26, 29–31] ranged from 21 to 43 days but also because this window covered the follow-up time for almost the entire population and reduced any potential bias by outliers. Patients discharged alive were censored at the discharge date. Because the data are directly derived from inpatient medical records, in-hospital mortality and survival can be considered accurate.

Confounding variables

We measured covariates thought to be associated with intravenous corticosteroid treatment or in-hospital mortality based on clinical judgement and prior literature as potential confounders [18, 32–35]. All variables were evaluated before the index time. These variables included age, sex, race, primary reason for hospitalization (surgery vs non-surgery), total and six subcategories of SOFA score, mechanical ventilation use, comorbidities (i.e. diabetes mellitus, arrhythmia, asthma, chronic obstructive pulmonary disease (COPD), coronary artery diseases, cancer history, cerebrovascular accident, gastrointestinal bleeding, heart failure, renal failure, respiratory failure), lab data (white blood cell count, blood glucose, serum lactate, liver aminotransferases, serum potassium, serum sodium, hemoglobin, percentage of band neutrophils, serum ammonia, and Troponin-I), and facility-level factors such as geographic region and number of beds. Comorbidities were evaluated based on past medical history. Lab data were extracted within 1-day prior to index. If the same labs were measured multiple times within this 1-day window, we chose the value representing the worst clinical condition. Individuals aged >89 were masked due to HIPPA regulation [22], so we classified all patients aged >89 as age 90. All laboratory values were dichotomized into clinically normal vs abnormal; untested laboratory data (S11 Table) were assumed to be normal.

Statistical analysis

Our primary objective was to examine the association between corticosteroid treatment and in-hospital survival up to 50 days in patients with sepsis defined by different algorithms. To control for confounding, we developed logistic models to generate stabilized inverse probability treatment weights (IPTW) and further apply the weights to the final outcome model. First, for the numerator of the weight, we used a saturated logistic model including only an intercept to estimate the probability of getting observed treatment. Second, for the denominator of the weights, we developed a logistic model to estimate the probability of getting observed treatment given all the confounding variables. After obtaining the stabilized weights, we further created a person-day IPTW weighted hazards outcome model in which we included time-varying intercept and product terms (time*treatment and time²*treatment) to allow the hazard ratio to vary over time [36–38]. We used absolute standardized differences to examine the balance of baseline characteristics between the treated and non-treated before and after IPTW weighting. Balance was achieved if absolute standardized differences were ≤0.1 for all the covariates [39, 40].

We evaluated the crude and IPTW weighted in-hospital survival up to 50 days and survival differences between treatment groups using 95% confidence intervals (95%CI). In specific, we calculated 95%CI for survival differences using bootstrapping with 500 iterations using random sampling with replacement [36, 41].

Subgroup analyses were conducted by sepsis severity (sepsis versus septic shock). In the first cohort, patients were stratified by having septic shock ICD codes or not (ICD-9: 785.52; ICD-10: R65.21), whereas in the second cohort, septic shock was defined as patients who had lactate ≥2 mmol/L and also received vasopressors [16]. We conducted the following sensitivity analyses: (1) assessing the total SOFA score over shorter time windows ranging from 3 hours to 1 day prior to sepsis onset; (2) using different windows to define corticosteroid treatment (i.e., 1-day and 5-day, adjusting accordingly for immortal time bias); (3) comparing the treated to those receiving any dose of corticosteroids; (4) including daily hydrocortisone-equivalent corticosteroid doses within 200mg and 400mg as inclusion criterion to better mimic RCTs [4, 11]; (5) excluding control patients who received any intravenous corticosteroids after sepsis onset to reduce the effects of short-term or very long-term corticosteroids; (6) excluding patients with asthma or COPD; (7) excluding hospitals that never prescribed corticosteroids to patients with sepsis to reduce potential facility-level confounding; (8) applying the g-formula [36] instead of IPTW with same controlled covariates to overcome any model misspecification; (9) excluding patients had >1 hospitalizations due to sepsis in the same calendar year; (10) assessing the in-hospital survival up to 90 days; (11) excluding patients without ventilation support; (12) excluding patients admitted for surgery; (13) selecting patients who received hydrocortisone as treated; (14) assuming all the untested values as abnormal; and (15) excluding those laboratory values with high percentage of untested patients including serum lactate, ALT, AST, % of band neutrophils, serum ammonia, and troponin-I from the analysis.

To avoid model over-fitting for some subgroup or sensitivity analyses in which the sample size of treated patients was smaller than 10*number of adjusted covariates, we used a stepwise model selection with significance levels of 0.1 and 0.2 for model entry, and inclusion, respectively; however, we forced total SOFA score to be included because of its hypothesized importance for confounding control. The final model that yielded the minimum AIC was chosen.

We used SAS (Version 9.4, Cary, NC) for all analyses.

50-day in-hospital survival and survival difference between treatment groups

Median follow-up for both cohorts were 9 days; the 50-day follow-up covered >95% of both cohorts (S2 Fig). Before IPTW adjustment, the treatment group had a higher in-hospital mortality rate: (97 vs 794 deaths; 13.2 vs 10.6 deaths/1000 person-days). In the second study cohort, the treatment group also had a higher in-hospital mortality rate: (89 vs 611 deaths; 14.3 vs 11.3 deaths/1000 person-days).

Table 3 shows the crude and IPTW adjusted 50-day in-hospital survival and survival difference between treated and non-treated patients. Fig 2 provides the survival curves and the survival difference between two treatment groups. In the first cohort, adjusted in-hospital 50-day survival in the treated was 0.62 vs 0.57 in the non-treated, and the in-hospital survival difference at day 50 was 0.05 (95%CI: -0.11, 0.17). In-hospital survival did not significantly differ between the groups throughout the entire follow-up period. Treated patients had slightly lower survival compared to controls in the first 19 days, with a maximum in-hospital survival difference -0.01 (95%CI: -0.05, 0.01) at day 10. The in-hospital survival difference became positive beginning on day 20 and diverged afterward. Subgroup analyses in which individuals were stratified by their having septic shock diagnosis codes showed similar results (Table 3), but the survival curves did not cross (S3 Fig). All results from the sensitivity analyses agreed with the primary analysis (S14 Table and S4 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. In-hospital survival and survival difference between treated and non-treated after inverse probability treatment weights.

A. In-hospital Survival up to 50 days for the Explicit Cohort; B. In-hospital Survival up to 50 days for the Sepsis-3 Cohort; C. In-hospital Survival Difference with 95% CI for the Explicit Cohort; D. In-hospital Survival Difference with 95% CI for the Sepsis-3 Cohort.

https://doi.org/10.1371/journal.pone.0243149.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. In-hospital survival and survival difference in the two study cohorts.

https://doi.org/10.1371/journal.pone.0243149.t003

In the second cohort where we further applied total SOFA≥2, the adjusted in-hospital 50-day survival results were similar to the first cohort (i.e. in-hospital 50-day survival difference, 0.02; 95%CI: -0.19, 0.17) (Table 3). Fig 2 shows that survival for treated patients was lower than non-treated patients from day 0 to day 19 and the maximum negative in-hospital survival difference was -0.02 (95%CI: -0.05, 0.02) at day 9. Survival curves for the two groups intersected at day 20 and then gradually diverged but with no significant survival differences. Subgroup analyses exploring whether a new septic shock definition produced different survival benefits were still non-significant (S5 Fig). All sensitivity analyses revealed consistent results with the primary approach (S15 Table and S6 Fig).

Clinical implications

To our knowledge, this is the first observational study focused on the effects of corticosteroids and in-hospital survival using Sepsis-3 criteria or using these methods to evaluate in-hospital survival [36]. As a result, absolute risks were reported as main estimates which may be more clinically relevant than hazard ratios [55]. According to the updated Surviving Sepsis Campaign guidelines [56], there has been no established evidence regarding the use of corticosteroids in septic patients defined by the Sepsis-3 criteria. Therefore, our findings can be used as updated evidence to help clinicians determine whether corticosteroid administration was necessary for patients with sepsis or septic shock defined by the Sepsis-3 criteria.

Limitations

This study has some potential limitations. First, using explicit diagnosis codes may have potentially included patients who actually did not have sepsis [15]. However, prior evidence has shown that the ICD-9 codes had very high PPV, so we believe the magnitude of the misclassification is limited and moreover unlikely to be differential between groups. Second, sicker patients tend to receive corticosteroids, and our study may suffer from confounding by indication even though we adjusted for numerous baseline confounders. However, the sensitivity analysis to explore this issue comparing the treated to those who received any dose of corticosteroids still agreed with the primary result. Third, although excluding patients who died within the first 96 hours successfully eliminated immortal time bias [28], this adjustment also introduced selection bias to our findings, limiting generalizability [57]. To minimize this, we conducted two additional sensitivity analyses shifting the exposure windows and changing immortal time adjustments, and the results were consistent with the primary approach. Fourth, we utilized sepsis onset as an anchor to calculate the SOFA score; while this may not perfectly represent Sepsis-3 definitions, sensitivity analyses changing SOFA score assessment windows were generally consistent. Fifth in Table 3, the adjusted survival of “all patients” was similar to the subgroup of patients with septic shock only, suggesting that the algorithm may identify patients with more critical conditions, leading to less generalizability. Sixth, the eICU database only contained patients admitted to ICUs from 2014 to 2015, which may limit generalizability to other years; some lab values were untested, and we assumed untested data to be normal. Finally, while we adjusted for numerous patient-level and facility-level factors, unmeasured confounding may affect the findings.