Diagnostic value of serum soluble triggering expressed receptor on myeloid cells 1 (sTREM-1) in suspected sepsis: a meta-analysis

Background We aim to synthesize the up-to-date studies to investigate the diagnostic value of serum soluble triggering expressed receptor on myeloid cells 1 (sTREM-1) in suspected sepsis. Results A total of 19 studies with 2418 patients were finally enrolled in the meta-analysis. The pooled sensitivity was 0.82 (95% CI 0.73 to 0.89), specificity 0.81 (95% CI 0.75 to 0.86), positive likelihood ratio 4.3 (95% CI 3.02 to 6.12), negative likelihood ratio 0.22 (95% CI 0.24 to 0.35), diagnostic odds ratio 20 (95% CI 9 to 41) and AuROC 0.88 (95% CI 0.85 to 0.91). The meta-regression analysis revealed that the sample size, reference standard description, prevalence of sepsis in the trials and consecution of patient recruitment might be the source of heterogeneity. Conclusions The serum sTREM-1 had a moderate ability in diagnosis in suspected sepsis based on the current studies. However, more large-scale studies were needed to further evaluate the diagnostic accuracy of sTREM-1.


Background
Sepsis is defined as life-threatening organ dysfunction caused by a dysregulated host response to infections, which causes high mortality in the intensive care unit (ICU) and is a grave burden to the public health [1] . Early recognition and diagnosis of sepsis in the high-risk patients with suspected infection is essential for the prompt management and empirical antibiotics therapy, which could potentially improve the mortality in septic patients [2].
The utility of biomarkers in the early recognition, risk stratification, antibiotic stewardship and outcome prediction in septic patients has long been applied in the clinical practice [3]. A myriad of molecules has been under investigation in the early discrimination of sepsis, including Creactive protein, procalcitonin, cytokines and surface markers of circulating leukocytes [4], which could be promising biomarkers in the diagnosis of sepsis.
Triggering receptor expressed on myeloid cells 1 (TREM-1), a member of immunoglobulin family predominantly expressed on the neutrophils and monocytes, was first identified by Bouchon A. et al in 2000 [5], which is upregulated in response of bacterial and fungal infections but poorly expressed in non-infectious inflammation [6,7] . The soluble form of TREM-1 (sTREM-1) is shed from cell surface and released into body fluids including plasma, pleural effusion, sputum and urine during the process of infections through proteolytic cleavage by metalloproteinases triggered by lipopolysaccharide [8,9].
The elevated sTREM-1 in the body fluids during infection could be measured directly by immunosorbent assays and used as a tool in discriminating infection from non-infectious inflammation, which makes it a promising candidate in the diagnosis of sepsis. A plethora of studies have been conducted to investigate the value of sTREM-1 as an early biomarker in patients with suspected infections since its discovery [10], however, its diagnostic accuracy remains undetermined.
Previous meta-analyses has evaluate the diagnostic ability of sTREM-1 in suspected infections [11][12][13], however with the burgeoning clinical studies of sTREM-1 as a diagnostic toolkit in suspected sepsis in recent years, we thought it quite necessary to conduct an upto-date meta-analysis to assess the diagnostic value of serum sTREM-1 in patients with suspected sepsis.

Syntheses of results
The synthesis of the 19 studies by the bivariate model yielded a pooled sensitivity of 0.82 (95% CI 0.73 to 0.89), specificity of 0.81 (95% CI 0.74 to 0.86), PLR of 4.3 (95% CI 3.0 to 6.1), NLR of 0.22 (95% CI 0.14 to 0.35) and DOR of 20 (95% CI 9 to 41) (Fig. 3), with AuROC of 0.88 (95% CI 0.85 to 0.91) (Fig. 4). The proportion of heterogeneity likely due to threshold effect was 0.20. We assumed the pre-test probability of 60% as the overall average sepsis prevalence concluded from the trials enrolled and yielded the post-test positive probability of 87% and negative of 26%, as illustrated in the Fagan's nomogram (Fig. 5). The scattergram indicated that the  sTREM-1 of limited clinical diagnostic value with PLR < 10 and NLR > 0.1 (no exclusion or confirmation) (Fig. 6).

Risk of bias and sub-group analysis
The studies were divided according to the cut-off values of sTREM-1, and the results revealed the AuROC of 0.87 (95% CI 0.84 to 0.90), sensitivity of 0.81 (95% 0.73 to 0.87) and specificity of 0.80 (95% CI 0.73 to 0.86) at the cut-off range of 30 to 199.72 pg/mL; and AuROC of 0.89 (95% CI 0.86 to 0.91), sensitivity of 0.85 (95% CI 0.64 to 0.95) and specificity of 0.80 (95% CI 0.65 to 0.90) at the cut-off range of 230 pg/mL to 60 ng/mL (Additional file 4).
A categorical univariate meta-regression analysis was conducted including the factors which could potentially bring bias to the results. We found that the sample size, and the reference standard description was possibly related with the heterogeneity in both sensitivity and specificity, and the prevalence of sepsis and the consecution of patient recruitment was possibly related with the heterogeneity in specificity (Fig. 7). Sub-group analyses were conducted according to the factors possibly related to the heterogeneity. (Additional file 6: Figure S2). The Deek's funnel plot was constructed and potential bias could be inspected with the P value of 0.002, suggesting publication bias should be considered (Fig. 8).

Disscussion
In this study, we synthesized 19 trials and found that the sTREM-1 has a moderately accuracy in diagnosis of sepsis in high-risk patients with the pooled sensitivity of sensitivity of 0.82 (95% CI 0.73 to 0.89), specificity of 0.81 (95% CI 0.74 to 0.86) and the AuROC of 0.88 (95% CI 0.85 to 0.91).
In previous meta-analysis by Wu et al including 11 trials of 1795 patients, they found the pooled sensitivity and specificity of 0.79 (95% CI 0.65 to 0.89) and 0.80 (95% CI 0.69 to 0.88) respectively, and DOR of 4.0 (95% CI 2.4 to 6.9) [12], and in our study with enlarged sample size enrolling 19 trials of 2418 patients saw improved accuracy in diagnosis in patients with suspected sepsis.
However, the sample size of the trials we enrolled in this study was relatively small (in most of studies the participants not exceeding 100), which could lead to type II error with elevated false positive [33]. We thus conducted the sub-group analysis including studies enrolling more than 100 patients and found the pooled sensitivity of 0.71 (95% CI 0.58 to 0.82), specificity of 0.69 (95% CI 0.54 to 0.81) and DOR of 5.48 (2.10 to 14.27), which showed that the sensitivity, specificity and DOR all dropped compared with the overall results, suggesting more large-scale trials are needed for the evaluation of the diagnostic ability of sTREM-1.
The baseline characteristics of patients was complicated (post-operative, traumatic, mechanical ventilated),  and heterogeneity in the patient selection might bring bias to the results; the severity of sepsis was different between studies with mortality ranged from 11% [27] to 54.4% [22], suggesting different pathogen virulence, inflammatory response and organ dysfunction, could probably influence the serum sTREM-1 level and its ability in discrimination of sepsis from SIRS. In some of the studies we included, not only patients with bacterial or fungal, but also with viral or parasite infection were recruited [23], and the heterogeneous composition of infection may also potentially bring bias to the results. The diagnosis of sepsis was relied on the comprehensive combination of clinical manifestation, laboratory and radiographic results, and the microbiological isolation or even in some occasions, the response to the empirical antibiotics therapy. We noticed that in the studies we included, the authors used inconsistent methods of determination of infection and in some studies only patients with positive pathogen isolation were included in the sepsis group [22,27], all of which could potentially cause biases.
The level of serum sTREM-1 was also reported to be elevated in some non-infectious diseases in recent years, including chronic kidney disease on hemodialysis [34], chronic obstructive pulmonary disease [35], inflammatory bowel disease [36], atherosclerosis [37] and in malignant carcinomas like hepatocellular carcinoma [38] and non-small cell lung cancer [39], which could affect the specificity of sTREM-1 in sepsis diagnosis.
An ideal biomarker for sepsis should give information to the syndrome recognition, precise diagnosis and prognosis, and improve antibiotic stewardship [3], but unfortunately, no single biomarker could accomplish this task currently. In our study, sTREM-1 yielded a moderate diagnostic accuracy of sepsis. Currently, combination of biomarker for the diagnosis and prognostication of sepsis have been widely investigated as in the trial by Gibot et al [21], serum sTREM-1 were combined with PMN CD64 index and procalcitonin in diagnosis of sepsis.

Limitations
This study had several limitations. The sample size in most of the trials we included in this meta-analysis was relatively small (less than 100 participants). In some trials, the predictive ability of sTREM-1 in sepsis was also evaluated, however, was not included in our study. All the cut-off values in the trials included were not prespecified, but optimized by the sensitivity and specificity, so the cut-off value of sTREM-1 for sepsis diagnosis was not determined so far.

Conclusions
Serum sTREM-1 has a moderate accuracy in diagnosis in patients with suspected sepsis, however, the heterogeneity was high between studies. More large-scale studies are needed for validation the diagnostic value of sTREM-1 in suspected sepsis.

Methods
This manuscript was prepared following the guidelines of Preferred Reporting Items for Systemic Reviews and Meta-analyses (PRISMA) statement [40,41] and Metaanalysis of Observational Studies in Epidemiology (MOOSE) statement [42]. This study has been registered in PROSPERO (CRD42018083695).

Eligibility criteria
We aimed to include all the clinical trials investigating serum sTREM-1 as an early biomarker in the patients with suspected sepsis. The including criteria were as follows: (1) clinical trials of adult patients (> 18 year-old) with suspected sepsis; (2) serum or plasma sTREM-1 protein expression was measured, if multiple measures were taken in the studies, only the earliest one was used; (3) a 2 × 2 contingency table with true positive (TP), false positive (FP), false negative (FN) and true negative (TN) could be constructed. As most of studies were conducted in prior of the upgrade to sepsis-3, sepsis was  [43], and the determination of infection was remained to the researchers in individual studies.

Information sources
Two reviewers searched the electronic database including PubMed, EMBASE, Cochrane Central Register and Web of Science updated to June, 2018 separately, with no language restrictions. When relevant reviews or meta-analysis were reviewed, a backwards snowballing search was conducted for further studies.

Study selection
The titles and abstracts of the articles initially reviewed separately by two reviewers, and the full manuscripts were recruited if potentially relevant for further assessment. Disagreements were solved by consensus.

Assessment of risk of Bias
The internal validity and risk of bias of the included studies were evaluated by the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) tool [44], which consist of four domains including patient selection, index test, reference standard, and flow and timing, with the risk of bias assessed as "low", "unclear" and "high". Discrepancies were solved by consensus. (Additional file 2).

Statistical analysis
The number of patients classified in TP, FP, FN and TN were calculated from the prevalence of sepsis, sensitivity and specificity as provided in the studies. Data synthesis was performed within the bivariate mixed-effects regression framework to calculate average sensitivity and specificity, and also positive/negative likelihood ratio (PLR & NLR) and diagnostic odds ratio (DOR, defined as PLR divided by NLR, which reflected the effectiveness of diagnosis), presented with 95% confidence intervals (CI) [45].
The derived logit estimates of sensitivity, specificity and respective variance were used to construct a summary ROC curve, and the area under ROC (AuROC) was calculated for the global measure of the test performance, with 0.5 > = AuROC <= 0.7 as low, 0.7 > = AuROC <= 0.9 as moderate, and 0.9 > = AuROC <= 1 as high diagnostic accuracy [46]. The threshold effects were also visually assessed from the summary ROC, and the proportion of variance due to threshold effects was calculated as the squared correlation coefficient estimated from the between-study covariance parameter tested by rank correlation test.
Post-test probability was calculated using likelihood ratios based on Bayes' theorem and depicted visually with Fagan's nomograms [47]. The likelihood ratio scattergram was also plotted, with the definition of left upper quadrant of exclusion and confirmation, right upper quadrant of confirmation only, left lower quadrant of exclusion only and right lower quadrant of no exclusion or confirmation, respectively [48].
The factors that could potentially bring heterogeneity were extract and introduced in the univariate metaregression analysis, as we speculated from the designation of the studies, including study design (prospective or not, and consecutive or not), settings (in the ICU exclusively or not), assays (methods and kit brands), sample size (> 100 participants or not), prevalence of sepsis (according to the average prevalence of sepsis from the studies included), the description of the reference standard (described in detail or not) and literature source (published in English or not). Sub-analyses were subsequently conducted according to the results of the Fig. 8 Deeks' funnel plot. The P value for the slope coefficient indicating significant asymmetry was 0.02, which indicating a high likelihood of publication bias univariate meta-regression to investigate the diagnostic performances of sTREM-1 in sub-group population of patients.
The Deeks' funnel plot was used to evaluate the publication bias, with P value < 0.10 for the slope coefficient indicating significant asymmetry and a high likelihood of publication bias [49].
STATA (ver. 14, StataCorp LP, TX, USA) was used for the analyses, and midas command was used for the calculations. A two-tailed P value < 0.05 was considered statistical significance.