A Molecular Host Response Assay to Discriminate Between Sepsis and Infection-Negative Systemic Inflammation in Critically Ill Patients: Discovery and Validation in Independent Cohorts

Background Systemic inflammation is a whole body reaction having an infection-positive (i.e., sepsis) or infection-negative origin. It is important to distinguish between these two etiologies early and accurately because this has significant therapeutic implications for critically ill patients. We hypothesized that a molecular classifier based on peripheral blood RNAs could be discovered that would (1) determine which patients with systemic inflammation had sepsis, (2) be robust across independent patient cohorts, (3) be insensitive to disease severity, and (4) provide diagnostic utility. The goal of this study was to identify and validate such a molecular classifier. Methods and Findings We conducted an observational, non-interventional study of adult patients recruited from tertiary intensive care units (ICUs). Biomarker discovery utilized an Australian cohort (n = 105) consisting of 74 cases (sepsis patients) and 31 controls (post-surgical patients with infection-negative systemic inflammation) recruited at five tertiary care settings in Brisbane, Australia, from June 3, 2008, to December 22, 2011. A four-gene classifier combining CEACAM4, LAMP1, PLA2G7, and PLAC8 RNA biomarkers was identified. This classifier, designated SeptiCyte Lab, was validated using reverse transcription quantitative PCR and receiver operating characteristic (ROC) curve analysis in five cohorts (n = 345) from the Netherlands. Patients for validation were selected from the Molecular Diagnosis and Risk Stratification of Sepsis study (ClinicalTrials.gov, NCT01905033), which recruited ICU patients from the Academic Medical Center in Amsterdam and the University Medical Center Utrecht. Patients recruited from November 30, 2012, to August 5, 2013, were eligible for inclusion in the present study. Validation cohort 1 (n = 59) consisted entirely of unambiguous cases and controls; SeptiCyte Lab gave an area under curve (AUC) of 0.95 (95% CI 0.91–1.00) in this cohort. ROC curve analysis of an independent, more heterogeneous group of patients (validation cohorts 2–5; 249 patients after excluding 37 patients with an infection likelihood of “possible”) gave an AUC of 0.89 (95% CI 0.85–0.93). Disease severity, as measured by Sequential Organ Failure Assessment (SOFA) score or Acute Physiology and Chronic Health Evaluation (APACHE) IV score, was not a significant confounding variable. The diagnostic utility of SeptiCyte Lab was evaluated by comparison to various clinical and laboratory parameters available to a clinician within 24 h of ICU admission. SeptiCyte Lab was significantly better at differentiating cases from controls than all tested parameters, both singly and in various logistic combinations, and more than halved the diagnostic error rate compared to procalcitonin in all tested cohorts and cohort combinations. Limitations of this study relate to (1) cohort compositions that do not perfectly reflect the composition of the intended use population, (2) potential biases that could be introduced as a result of the current lack of a gold standard for diagnosing sepsis, and (3) lack of a complete, unbiased comparison to C-reactive protein. Conclusions SeptiCyte Lab is a rapid molecular assay that may be clinically useful in managing ICU patients with systemic inflammation. Further study in population-based cohorts is needed to validate this assay for clinical use.


Procedure for Classifying Patients
A total of 345 patients satisfying the above criteria were selected to define the five validation cohorts of the present study. Patients in each validation cohort were characterized and classified by the following three-step process: Step 1: Patients in each cohort were subjected to daily assessments by the attending physician(s), to estimate: 1) the severity of the Sepsis Event (if such an event occurred) and 2) the infection likelihood. In addition, physiological data were recorded daily. Patient information was analyzed each day to produce an infection likelihood assessment (definite, probable, possible or none) [5,6] and the previous day's assessment was overwritten.
Step 2: Retrospective analysis of all patient information, at a median time of 3 months post-admission, led to a final determination of infection likelihood and an identification of primary and any secondary sites of infection. Infection likelihood was assessed after taking into account all available information in the patient's medical chart. Infection likelihood was scored as definite (culture proven), probable, possible or none.
To perform the evaluation, two trial physicians on site (at either the Amsterdam or Utrecht ICU) examined each patient's chart. (The two physicians were randomly selected from a pool of 8 physicians tasked with assessing patients.) A third trial physician was used to adjudicate in cases of discordance over the infection likelihood as per [5].
Step 3: Patients were classified as follows: • Controls (Infection-Negative Systemic Inflammation): This group consisted of: 1) patients who exhibited ≥2 clinical signs of systemic inflammation, but who were never assessed for infection likelihood, because they did not have a Sepsis Event, were not suspected of sepsis, and were never treated for sepsis; 2) patients who exhibited ≥2 clinical signs of systemic inflammation and experienced a Sepsis Event, and therefore as a precaution were given therapeutic systemic antibiotics, but who were then retrospectively adjudicated to have an infection likelihood of none.
• Cases (Sepsis): This group consisted of: 1) patients who experienced a Sepsis Event and were then retrospectively adjudicated to have an infection likelihood of probable; 2) patients who experienced a Sepsis Event and were then retrospectively adjudicated to have an infection likelihood of definite (culture proven).
• Infection Likelihood of Possible: Patients who experienced a Sepsis Event and were given therapeutic systemic antibiotics as a precaution, but who were then retrospectively adjudicated to have an infection likelihood of possible could not be classified with certainty as either cases or controls. These patients were excluded from performance analyses.
Note that, in proceeding from Step 2 to Step 3, there has been a translation from an Infection Likelihood classification (ordinal scale with the four values of definite, probable, possible, none) to a case/control classification (ordinal scale with the three values of case, control, and infection likelihood possible). Difficulties inherent in the process of translating between reference scales, when each reference scale is imperfect, have been discussed in [5,[8][9][10]. Table 1 defines the five validation cohorts of this study. Validation Cohort 1 (n=59) consisted only of patients for which the diagnosis as case or control was definite or probable. By design, this cohort did not include any patients with an infection likelihood of possible, whereas in the entire MARS database ~15% of patients are classified with an infection likelihood of possible. The patients in this cohort were admitted to the Utrecht ICU from December 2012 to March 2013, but were not enrolled sequentially. RT-qPCR data for this cohort were generated in July 2013.

Definition of Validation Cohorts
Validation Cohort 2 (n=36) consisted of 36 patients exhibiting ≥ 2 criteria of systemic inflammation, randomly picked from the Amsterdam ICU (n=19) or Utrecht ICU (n=17) over the entire available dates of the study (December 2012 to July 2013). This cohort included 6 patients with an infection likelihood of possible. This cohort was used to check that the SeptiCyte Lab signature was not skewed with respect to study enrolment date. RT-qPCR data for this cohort were generated in July and October 2013 (two batches).
Validation Cohort 3 (n=106) was drawn from an initial set of n=1004 patients consecutively admitted to the Amsterdam ICU (n=273) and Utrecht ICU (n=731) from December 2012 to July 2013. From this initial set, 150 patients with an infection likelihood of possible were deliberately excluded. An additional four patients were excluded because insufficient data were captured to meet the minimum reporting requirements for retrospective physician evaluation. From the remaining pool of 850 patients, random draws were made of 75 patients from the Amsterdam ICU and 74 patients from the Utrecht ICU. After then applying the study exclusion criteria, the final numbers of patients from the two study centers were 52 patients from the Amsterdam ICU, and 54 patients from the Utrecht ICU. RT-qPCR data for this cohort were generated in November 2013.
Validation Cohort 4 (n=87) consisted of patients exhibiting ≥ 2 criteria of systemic inflammation, drawn sequentially from the Amsterdam ICU from March 2013 to June 2013. Patients with an infection likelihood of possible were included. This cohort was used to assess performance in a real-world setting (sequential patients), and also the feasibility of porting the SeptiCyte Lab assay from one brand of RT-qPCR reagents to another brand of reagents. RT-qPCR data for this cohort were generated in August 2014.
Validation Cohort 5 (n=57) consisted of patients of Black or Asian ethnicity consecutively enrolled at the Amsterdam ICU (N=46) or Utrecht ICU (n=11) from November 2012 to August 2013. This cohort was used to test the performance of the SeptiCyte Lab classifier for bias due to ethnicity. This cohort included 11 patients with an infection likelihood of possible. RT-qPCR data for this cohort were generated in April 2014.