Study design and setting
This prospective observational study used emergency department (ED) and ICU data from 5 university hospitals in Japan. All hospitals received individual local institutional review board approval for conducting research with human subjects. The Ethics Committee at the Keio University School of Medicine approved this study (approval number 16-03-007). Informed consent was obtained from all patients for being included in the study.
Study Population
The study enrolled critically ill patients admitted to the participating centers between September 2016 and September 2018. Inclusion criteria were: age ≥ 20 years; ≥2 systemic inflammatory response syndrome (SIRS) criteria of the American College of Chest Physicians/Society of Critical Care Medicine on ED/ICU admission [22]; expected ICU stay ≥ 48 hours. In addition to meeting these criteria, burn patients were included with burn index ≥ 15 and trauma patients with injuries in ≥ 2 body regions on the abbreviated injury scale coding system and with Injury Severity Score ≥ 10. Exclusion criteria were: current medications that affect serum IL-6 concentration (e.g., corticosteroids, immunosuppressants) within 1 week before study inclusion; discharge or death within 48 hours after admission; deviation from study protocol for biomarker tests and SOFA score calculation; HIV infection; pregnancy; and any other condition precluding suitability for enrollment in the investigators’ opinion.
Data Collection And Definitions
Patient information included: demographic characteristics; admission source; comorbidities; medications administered within 1 week before study inclusion; etiology on admission; episode of cardiac arrest before study inclusion; presence of hemodynamic instability defined as vasopressor requirement or persistent hypotension despite fluid resuscitation; and any ED/ICU treatments.
Blood samples were obtained within 6 hours after ED/ICU admission (Day 0) and the next morning (Day 1). Blood tests were then performed daily from Days 2–3 and as need until 7 days after admission. These inflammatory biological markers were: C-reactive protein (CRP), IL-6, IL-8, IL-10, tumor necrosis factor (TNF)-α, and procalcitonin (PCT). Serum CRP was measured immediately with commercially available assays at each hospital; ILs, TNF-α, and PCT were measured blindly to treating physicians at an outside facility after serum samples were frozen and stored at − 20 °C (IL-6, PCT, Roche Diagnostics, Mannheim, Germany; IL-8, IL-10, BioSource Europe, Nivelles, Belgium; TNF-α, R&D Systems, Minneapolis, MN, USA).
Arterial blood gas analysis and other blood tests for calculating each SOFA score component were performed at each hospital at the same time of blood sampling for biological markers. The SOFA score was recorded daily until Day 3 and as needed until Day 7; APACHE II score was also calculated on ED/ICU admission.
Outcome Measures
Primary outcome was 28-day all-cause mortality. Secondary outcome was ICU-free days, defined as the number of days alive and out of the ICU between admission and Day 28.
Statistical analysis
To assess improved accuracy for mortality prediction by adding a biomarker test to SOFA score, a baseline model was developed using logistic regression analysis to predict 28-day mortality, in which SOFA score at Day 2 was chosen as a sole explanatory variable. The Day 2 score was chosen because the original SOFA score validation study showed that the SOFA score at 48 hours post-admission had a high discrimination power for predicting ICU mortality [8]. To identify the best time for each biomarker to predict mortality, receiver operating characteristic (ROC) curves were drawn for serum concentration of each biomarker at Days 0–3 based on 28-day mortality, and area under the ROC curve (AUROC) was calculated. The day with the highest AUROC was considered as best time point for each biomarker.
Logistic regression analyses to predict 28-day mortality were performed again to derive the linear combination of the baseline model (Day 2 SOFA score) and an additional biomarker as measured based on the best time point just described. Some biomarkers were analyzed with sex as suggested in other studies [23, 24]. The ROC curves were drawn and AUROC was compared between the baseline model and combination model developed with the additional biomarker. Improvement of AUROC from baseline was shown with 95% confidence interval (CI). Sensitivity and specificity of each model were also obtained at a best cutoff point defined as the Youden index(25). To assess optimism of the combination model using the additional biomarker, a corrected AUROC was calculated with bootstrap analysis (resampling the model 1000 times) [26]. The combination models with additional biomarkers were also examined in linear regression analyses to predict ICU-free days. The clinical applicability of biomarker was then assessed by calculating observed 28-day mortalities in subgroups classified as SOFA score and the biomarker dichotomized at the median value.
Descriptive statistics are presented as the mean (standard deviation), median (interquartile range), or number (percentage). No imputation was used to estimate missing data. The improvement of predictive ability for mortality by adding biomarkers was unclear before the study, and sample size estimation was not performed for the main analysis. Sample size estimation for ROC analysis in which 0.7 of AUROC was expected for an event with 15% of incident rate indicated that 150 cases were needed with power of 80% and α error of 0.05 [27]. Results were compared using Mann-Whitney U tests, chi-square tests, or Fisher's exact tests, as appropriate. For testing all hypotheses, a two-sided α threshold of 0.05 was considered statistically significant. All statistical analyses were conducted using SPSS, version 26.0 (IBM, Armonk, NY, USA) and R Version 4.0.2 (R Foundation for Statistical Computing, Vienna, Austria).