The association of endothelial cell signaling, severity of illness, and organ dysfunction in sepsis

This was a prospective, cohort study of a convenience sample of adult patients (age 18 years or older) presenting to the ED with suspected infection. Suspected infection was defined as a clinical suspicion of an infectious etiology as assessed by the treating clinician, and determined by interviewing the treating physician to determine if infection was suspected based on the ED work-up including the results from history, physical exam, laboratory and diagnostic testing. The population was selectively enrolled to achieve a relatively even distribution of different sepsis severities. A sample of non-infected ED control patients was also assembled by identifying adult ED patients without evidence of infection during presentation. The study period was between February 2005 and November 2008. There were 221 patients enrolled in the study with 189 patients enrolled de novo, and 32 patients co-enrolled with another protocol [4]. The setting was Beth Israel Deaconess Medical Center (BIDMC), Boston, an urban teaching hospital. The study was approved by the hospital ethics board, and written informed consent was obtained.

In order to characterize the population, relevant components of demographics, history, co-morbid diseases, suspected source of infection, vital sign information, physical exam findings, and the results of laboratory and radiologic testing were collected. The Charlson comorbidity index, a well established methodology to quantify co-morbid disease burden, was calculated for each patient [5].

All subjects received a blood draw while in the emergency department. Samples were drawn in EDTA tubes, centrifuged at 2,500 × g at 4°C, and frozen at -80°C within one hour of collection. Plasma was assayed for sE-selectin, sICAM-1, sVCAM-1, and PAI-1 as a multiplex panel using the human cardiovascular-1 panel (Millipore, Billerica, MA, USA) and Interleukin-6 (IL-6) using the human cardiovascular-3 panel (Millipore) on the Luminex 200 instrument (Millipore). The sFlt-1 assays were performed using Quantikine ELISA kits (R&D systems, Minneapolis, MN, USA). All assays were performed in duplicate and the average levels were used for analysis.

Between January 2007 and January 2009, patients in our study with septic shock received additional blood draws every 24 hours for the first three days - a total of 52 patients were enrolled in this subset. This sub-study was performed to assess the changes in the circulating biomarkers of endothelial cell activation over time.

Means with standard deviations, medians with interquartile ranges, and proportions were used for descriptive statistics, as appropriate. To analyze the association between the biomarkers of endothelial cell activation and sepsis severity, we used generalized linear modeling. Next, we calculated Spearman rank correlation coefficients to assess the bivariable association among the biomarkers. We display the graphs with a regression line and reported the calculated Spearman correlation coefficient (r-value) along with the associated P-value. We performed a similar analysis between the target biomarkers and organ dysfunction (SOFA score), the inflammatory response marker IL-6, and APACHE-II score. Due to non-normal distribution, SOFA score was log transformed throughout the analysis. As a comparator, we also examined the correlation of IL-6 and serum lactate with SOFA score. Next, to compare the strength of association between each of the biomarkers and organ dysfunction, we standardized each of the biomarkers values through the following formula: ((biomarker - biomarker mean)/biomarker SD). We then used a linear regression model and adjusted for age, gender, and co-morbid illness burden (Charlson score). We report the beta coefficient with standard error as well as the adjusted r-squared value for each biomarker model. We also tested multi-marker models to determine the value of combinations of biomarkers. To assess the clinical predictive ability of the biomarkers, we calculated the area under the receiver operating characteristic curve (AUC) with 95% confidence interval for each biomarker to predict the outcomes of severe sepsis (including septic shock) within 72 hours and in-hospital mortality. The AUCs were compared nonparametric approach [16].

Finally, for the subset analysis of biomarkers from patients with septic shock collected daily over the first 72 hours of hospitalization, we used a linear mixed effects model to estimate the differences in biomarkers between survivors and non-survivors over time. The linear mixed-effects model took into account the multiple measurements (at 0, 24, 48, 72 hours) of biomarkers and outcomes and used compound symmetry variance-covariance structure to account for the within-subject correlation.

There were a total of 221 patients enrolled with a mean age of 58 (SD +/- 19) years; 52% were male, 76% Caucasian, and there was a high co-morbid burden: diabetes (26%), cancer (20%) and chronic heart failure in 13% (Table 1). On admission, sepsis without organ dysfunction was present in 32%, severe sepsis without shock in 30%, and septic shock in 32%. Six percent were non-infected ED patients who were used as controls. The overall in-hospital mortality in the population was 7.7% (13/221), and 42% (84/221) of patients were admitted to the intensive care unit (ICU).

We found an association between biomarker levels and sepsis severity (worst sepsis syndrome within 72 hours) for sFlt-1 (P < 0.001 for trend across groups), PAI-1 (P < 0.001), sE-selectin (P < 0.001), sICAM-1 (P < 0.05), and sVCAM-1 (P < 0.04) (Figure 2). The most significant increases were found in median sFlt-1 levels, which ranged from 41 ng/ml (IQR 31 to 51) in non-infected controls to 243 ng/ml (IQR 137 to 449) in septic shock; and, in PAI-1 which ranged from 25.3 ng/ml (IQR 17.6 to 36.8) to 76.7 ng/ml (IQR 49.4 to 136).

To assess whether there was evidence of endothelial cell activation in the response to infection, we correlated the selected biomarkers which individually represent various components of the endothelial cell signaling pathway. Using a Spearman rank correlation coefficient, we found a significant correlation between all biomarkers (sFlt-1, PAI-1, sE-selectin, sICAM-1, and sVCAM-1) (Figure 3). The strongest correlations were between sFlt-1 and sE-selectin (r = 0.55, P < 0.001) and sFlt-1 and PAI-1 (0.56, P < 0.001).

To assess the association of endothelial cell related biomarkers with organ dysfunction, we analyzed the correlation between the endothelial related biomarkers with SOFA score in the ED. All biomarkers were significantly correlated with the concurrent SOFA score (Figure 4). Of note, sFlt-1 was highly correlated (r = 0.66, P < 0.001) with SOFA score, and compared favorably in predicting SOFA score to other common biomarkers of inflammation such as IL-6 (r = 0.45) and lactate (r = 0.43). In addition, biomarker levels at the time of presentation correlated with SOFA score at 24 hours: sE-selectin (0.37), sFlt-1 (0.64), sVCAM-1 (0.22), and PAI-1 (0.51), P < 0.001 for all comparisons; except sICAM-1 (0.13), P = 0.08.

We used circulating IL-6 concentrations as a read-out of the pro-inflammatory response. There was a notable association between the biomarker levels of endothelial activation and IL-6 (Figure 4). Here, sFlt-1 had a particularly strong correlation with IL-6 (r = 0.62, P < 0.001).

Endothelial cell activation markers correlated with two independent markers of disease severity, lactate and APACHE-II scores. There was a significant correlation using Spearman rank between the target biomarkers and APACHE-II score: sFlt-1 (r = 0.58, P < 0.01), pai1 (0.46, P < 0.01), sE-selectin (0.33, P < 0.01), sICAM-1 (0.15, P < 0.03), and sVCAM-1 (0.25, P < 0.01). These results compare favorably to the r-value for the correlation between classic biomarkers such as lactate with APACHE-II (0.38) and IL-6 with APACHE-II (0.43). There was a significant association between the endothelial related biomarkers and lactate level: sFlt-1 (0.51, P < 0.01), PAI-1 (0.40, P < 0.01), sE-selectin (0.33, P < 0.01), sICAM-1 (0.23, P < 0.01), and sVCAM-1 (0.20, P < 0.01). As a comparator, IL-6 correlation coefficient with lactate was 0.44.

We analyzed the association of the biomarkers with organ dysfunction (log SOFA score) with linear regression models adjusted for age, gender, and co-morbid illness burden (Table 2) using beta coefficients standardized to a 0 to 100 scale to allow equal comparison. We report the models testing one marker at a time (Table 2). We then checked to see if model fit (measured by adjusted R²) would be improved by any combination of multiple markers in the model. Interestingly, once sFlt-1 was included in the models, no additional marker becomes significant if added. The R² value in the adjusted model for sFlt-1 alone was 0.46, and adding any second marker did not improve the model fit above this level. Additionally, and there was no other combination of two or more markers that exceeds the R² of the model with sFlt-1 alone, including adding IL-6 and lactate as eligible covariates. Thus, the marker sFlt-1 appears to have the strongest association with organ dysfunction.

To further assess the clinical accuracy of the different markers, we report the area under the receiver operating characteristic curve for the ability of the biomarker drawn on ED presentation to predict two clinical outcomes: 1) severe sepsis (including septic shock as cardiovascular dysfunction) within 72 hours; and, 2) in-hospital mortality (Table 3). Again, sFlt-1 performed with the highest accuracy, and has a higher AUC (0.82; 95% CI 0.76 to 0.88) for severe sepsis when compared to all other endothelial related biomarkers (P < 0.05). For the outcome of in-hospital mortality, sFlt-1 had an AUC of 0.91 (0.87 to 0.95), and was also higher (P < 0.05) than the AUC for all other markers (Table 3).

There were a total of 52 patients with septic shock who in addition to the 0 hour draw had serial samples at 24, 48 and 72 hours. We compared biomarker levels in survivors (n = 43) to non-survivors (n = 9) (Figure 5). Using a linear mixed-effects model, adjusting for age, gender, and co-morbid burden, we found the following estimated mean differences in biomarker levels over time comparing the non-survivors to survivors: sFlt-1 366 pg/mL (95% CI: 218 to 514, P < 0.01); PAI-1 63.2 ng/ml (38.5 to 87.8, P < 0.01); sE-selectin 24.1 ng/mL (5.5 to 42.7, P < 0.01); sICAM-1 135 ng/mL (67 to 202, P < 0.01); and, sVCAM-1 683 ng/mL (320 to 1,046, P < 0.01).

This study has a number of important limitations. First, it was a convenience sample that may have suffered from selection bias. However, the population was selected to obtain a spectrum of severities as opposed to a consecutive sample of patients. Second, we primarily only analyzed blood from the initial draw, and except in the septic shock subset, did not follow biomarkers over time. Third, in our modeling, we adjusted for age, gender, and co-morbidity, but other important confounders may have affected our results. Fourth, circulating levels of endothelial biomarkers are only indirect measures of endothelial cell activation, and thus may not accurately reflect the degree, nature and site of activation of the intact endothelium. While we have selected representative biomarkers, others may still be more accurate. Fifth, we did not include a population of non-infected critically ill patients (for example, trauma patients) so we are unable to answer whether the endothelial cell changes are specific to sepsis, or broader markers of illness severity that would extend across disease states. Finally, our sample size is reasonable, but a larger study may have afforded the opportunity for more complete subset analysis. Both our sample over time analysis and mortality analysis was limited by a small sample size.