Study design, patients and definitions
This study was conducted as part of the “Molecular Diagnosis and Risk Stratification of Sepsis” (MARS) project, a prospective observational study in the mixed ICUs of two tertiary teaching hospitals (Academic Medical Center in Amsterdam and University Medical Center Utrecht) in the Netherlands [9,10,11]. Trained physicians prospectively collected the following data: demographics, comorbidities, chronic medication use, ICU admission characteristics, daily physiological measurements, severity scores, antibiotic use, and culture results. The plausibility of infection was post hoc scored based on all available evidence and classified on a 4-point scale (none, possible, probable or definite) according to Center for Disease Control and Prevention [15] and International Sepsis Forum consensus definitions [16], as described in detail previously [9]. For the current analysis, we selected all patients included in the MARS-study between January 2011 and July 2013 with sepsis, diagnosed within 24 h after admission, defined by the presence of a definite or probable infection [9] combined with at least one of general, inflammatory, hemodynamic, organ dysfunction or tissue perfusion parameters derived from the 2001 International Sepsis Definitions Conference [17]. Readmissions and patients transferred from another ICU were excluded, except for patients referred to one of the study centers on the day of admission. Organ failure was defined as a score of 3 or greater on the SOFA score, except for cardiovascular failure for which a score of 1 or more was used [12]. Shock was defined as use of vasopressors (noradrenaline) for hypotension in a dose of 0.1 mcg/kg/min during at least 50% of the ICU day. Patients were assessed daily for the presence of acute kidney injury and acute lung injury using strict preset criteria [13, 14]. Left-over plasma (obtained from blood drawn for patient care) was obtained within 24 h of admission to the ICU and stored within 4 h at − 80 °C. The Medical Ethical Committees of both study centers gave approval for an opt-out consent method (IRB no. 10-056C) [9, 10]. The Municipal Personal Records Database was queried to determine survival up to 1 year after ICU admission.
Biomarker assays
All measurements were performed in EDTA anticoagulated plasma obtained on admission. Tumor necrosis factor alpha (TNF-α), interleukin-1beta (IL-1β), IL-6, IL-8, IL-10, IL-13, interferon-γ, granulocyte-macrophage colony-stimulating factor (GM-CSF), soluble intercellular adhesion molecule-1 (ICAM-1), soluble E-selectin and fractalkine were measured using FlexSet cytometric bead arrays (BD Bioscience, San Jose, CA) using a FACS Calibur (Becton Dickenson, Franklin Lakes, NJ, USA). Angiopoietin-1, angiopoietin-2, protein C, antithrombin, matrix metalloproteinase (MMP)-8, tissue inhibitor of metalloproteinase (TIMP)-1 (R&D systems, Abingdon, UK), and D-dimer (Procartaplex, eBioscience, San Diego, CA) were measured by Luminex multiplex assay using a BioPlex 200 (BioRad, Hercules, CA). C-reactive protein (CRP) was determined by an immunoturbidimetric assay (Roche diagnostics). Platelet counts were determined by hemocytometry, prothrombin time (PT) and activated partial thromboplastin time (aPTT) by using a photometric method with Dade Innovin Reagent or by Dade Actin FS Activated PTT Reagent, respectively (both Siemens Healthcare Diagnostics). Normal biomarker values were acquired from 27 age- and gender-matched healthy volunteers, from whom written informed consent was obtained, except for CRP, platelet counts, PT and aPTT (routine laboratory reference values).
Blood gene expression microarrays
Whole blood was collected in PAXgene™ tubes (Becton–Dickinson, Breda, the Netherlands) within 24 h after ICU admission. PAXgene blood samples were also obtained from 42 healthy controls [median age 35 (interquartile range 30–63) years; 57% male] after providing written informed consent. Total RNA was isolated using the PAXgene blood mRNA kit (Qiagen, Venlo, the Netherlands) in combination with QIAcube automated system (Qiagen, Venlo, the Netherlands), according to the manufacturer’s instructions. RNA (RNA integrity number > 6.0) was processed and hybridized to the Affymetrix Human Genome U219 96-array and scanned by using the GeneTitan instrument at the Cologne Center for Genomics (CCG), Cologne, Germany, as described by the manufacturer (Affymetrix).
Raw data scans (.CEL files) were read into the R language and environment for statistical computing (version 2.15.1; R Foundation for Statistical Computing, Vienna, Austria; http://www.R-project.org/). Pre-processing and quality control was performed by using the Affy package version 1.36.1. Array data were background corrected by robust multi-array average, quantiles-normalized and summarized by median polish using the expresso function (Affy package). The resultant 49,386 log-transformed probe intensities were filtered by means of a 0.5 variance cutoff using the genefilter method [18] to recover 24,646 expressed probes in at least one sample. The occurrence of non-experimental chip effects was evaluated by means of the Surrogate Variable Analysis (R package version 3.4.0) and corrected by the empirical Bayes method ComBat [19, 20]. The non-normalized and normalized MARS gene expression data sets are available at the Gene Expression Omnibus public repository of NCBI under accession number GSE65682. The 24,646 probes were assessed for differential abundance across healthy subject and patient samples by means of the limma method (version 3.14.4) [21]. Supervised analysis (comparison between pre-defined groups) was performed by moderated t statistics. Throughout Benjamini–Hochberg (BH) multiple comparison adjusted probabilities, correcting for the 24,646 probes (false discovery rate < 5%), defined significance. Ingenuity Pathway Analysis (Ingenuity Systems IPA, http://www.ingenuity.com) was used to identify the associating canonical signaling pathways stratifying genes by over- and under-expressed patterns. The ingenuity gene knowledgebase was selected as reference and human species specified. All other parameters were default. Multiple comparison adjusted Fisher test probabilities < 0.05 defined significance.
Statistical analysis
Data analyses were performed in R (v3.1.1) [22]. Baseline characteristics of study groups were compared with Chi-square test for categorical variables and t-test for continuous variables. Non-normally distributed continuous variables, including biomarker levels, were analyzed with Wilcoxon rank sum test. To account for differential likelihood of receiving statins, we constructed a propensity score [23], using logistic regression, including variables associated with use of statins and variables that we considered of relevance to our outcome. This score included age, gender, weight, race (white), cerebrovascular disease, chronic cardiovascular insufficiency, chronic renal insufficiency, congestive heart failure, chronic obstructive pulmonary disease (COPD), diabetes mellitus, hematologic malignancy, hypertension, metastatic malignancy, history of myocardial infarction, ACE-inhibitors/ARBs, antiplatelet drugs, beta-blockers, oral antidiabetic drugs, and site of infection (pulmonary, abdominal, urinary). Subjects were 1:1 matched by the estimated propensity score using nearest neighbor matching with a caliper of 0.2SD of the logit of the propensity score, using R package “MatchIt”. Patients whose plasma samples were not collected for biomarker analyses within 24 h of ICU admission and were excluded from the matching procedure. In addition, matching for analyses of gene expression profiles was done using only patients from whom gene expression data were available. Standardized differences were calculated to determine balance between the propensity-matched groups [24]. In order to retain enough power to detect differences in biomarker levels, we accepted standardized differences between propensity-matched groups for comorbidities and chronic medication up to 20%. To investigate the independent association between statin use and 30-day mortality in our propensity-matched plasma biomarker cohort, we performed logistic regression including statin use, variables associated with mortality and comorbidities not optimally matched between users and non-users. P values below 0.05 were considered statistically significant. In host response biomarker comparisons, a Bonferroni-corrected P value of 0.002 was taken as cutoff to define statistical significance.