This systematic review and meta-analysis was performed and reported in accordance with the guidelines of Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA).8 PubMed, EMBASE and The Cochrane Library were searched for trials in English, and China National Knowledge Infrastructure (CNKI) database were searched for trials in Chinese to identify the relevant published studies for clinical data. Searches were conducted separately for each literature databases up to May 2017. The search terms for each database were (“adrenomedullin” OR “proadrenomedullin” OR “mid-regional proadrenomedullin” OR “MR-proadrenomedullin”) AND (“sepsis” OR “septicemia” OR “septicaemia” OR “septic” OR “septic shock”). In addition to our computerized search, the bibliographies of all articles were reviewed to look for other potential articles with eligible data for the meta-analysis. The corresponding authors or first authors of the publications were contacted if additional information was needed, results were unclear, or relevant data were not reported.

Studies were considered eligible for the systematic review ad meta-analysis if they fulfilled the following criteria: 1) data was from an original, peer-reviewed study rather than review articles or meeting abstracts; 2) patients with sepsis or septic shock; 3) studies that provided mortality or diagnostic characteristics, that is, the area under the receiver operating characteristic (ROC) curve (AUC), or sensitivity and specificity. We excluded studies that were animal studies or duplicate studies.

After deduplication, references were screened based on their titles and abstracts against a set of pre-defined eligibility criteria. Only one reviewer was involved in this round of review and the criteria was relative wide. If any uncertainty on inclusion existed, a second reviewer would join to assess and reach a consensus. Then, the full texts of retained references after first-round screening were obtained and reviewed by two reviewers against stringent inclusion criteria. Disagreements were resolved through discussion, and consensus was reached through a third party. All papers included after completion of the full text review were retained for data extraction and quality assessment.

The following information was extracted from the eligible studies: first author, publication year, country of origin, setting, sample size, gender distribution, mean ages, mortality rate, test methods, and diagnostic performances (AUC, cut-off value, sensitivity, and specificity).

Two investigators independently assessed the quality of eligible studies using the revised Quality Assessment for Studies of Diagnostic Accuracy tool (QUADAS-2).9 Agreement between the two reviewers for assessment of methodological quality was evaluated using the Cohen's Kappa statistic. Any disagreement in data extraction and quality assessment was resolved by discussion.

The analyses were performed with Review Manager (REVMAN) 5.3 Copenhagen (The Nordic Cochrane Centre, The Cochrane Collaboration, 2014) and STATA (version 14; Stata Corp., College Station, Texas, USA). Reported values are 2-tailed, and hypothesis-testing results were considered statistically significant at P<0.05.

Relative risk (RR) was used to assess the predictive value of ADM and proADM for mortality. We tested heterogeneity with Cochran's Q statistic, with P<0.10 indicating heterogeneity, and quantified the degree of heterogeneity using the I2 and tau2 statistic. In case of significant heterogeneity between studies, DerSimonian and Laird random effects model was used; otherwise, the fixed effects model was performed.10 Publication bias was tested using Deek's funnel plot.

The presence of a threshold effect on the prognostic accuracy of ADM and proADM in patients with sepsis was evaluated with the Spearman correlation coefficient between the logits of sensitivity and specificity. If no threshold effect existed, a bivariate random effects regression model11 was used to calculate the pooled sensitivity, specificity, diagnostic odds ratio (DOR), positive likelihood ratio (PLR), and negative likelihood ratio (NLR). We also constructed a summary receiver operator characteristic (SROC) curve by plotting the individual and summary points of sensitivity and specificity to assess the overall diagnostic accuracy.12

Eventually, 13 articles were eventually included after examining the initial searched 279 articles. The process of study selection and the final results of the search were illustrated in the Flow Diagram (Fig. 1).

Figure 1.

Flowchart of study selection.

(0.12MB).

Thirteen studies involving 2556 patients were published from 1999 to 2017.13–25 Five studies reported the prognostic values of ADM and eight reported the proADM (Table 1). All studies were published in English. The mean age of patients varied between 60.0 and 72.0 years and the proportion of men ranged from 58.6% to 74.2%. Eight studies were performed in ICUs, two in EDs and three in both departments. Follow-up periods differed across studies, including 28-days mortality (5 studies), 90-days mortality (1 study), and in-hospital mortality (7 studies). The mortality rate varied from 23.5% to 47.6%.

Fig. 2 summarizes the methodological quality assessment with the QUADAS-2 tool of the thirteen included studies. Three studies were labeled as unclear in the patient selection domain as there is no report on whether patients were consecutive (Fig. 2A). For the index text domain, there is no cutoff value in three studies which has been labeled as unclear. The reference standard domain of two studies were labeled as high risk as the ADM or proADM levels were calculated before the endpoint was revealed. Seven studies were labeled as unclear because whether ADM or proADM were measured in a blinded fashion is unknown. There is neither risk in flow timing nor application concerns in all these thirteen studies (Fig. 2B). The Cohen's Kappa statistic for inter-rater agreement of assessment of quality between two reviewers was 0.84.

Figure 2.

Quality assessment (QUADAS-2) summary of 13 studies (A) and quality assessment (QUADAS-2) for each included study (B).

(0.41MB).

All studies showed that an elevated ADM or proADM levels were associated with a higher risk of death, with RR ranging from 1.58 to 12.67. Due to the substantial heterogeneity between studies (I2=82%, Tau2=0.30, Cochran's Q statistic P<0.01), a random-effects model was used to pool RR estimates (Fig. 3). The pooled RR in all the 13 studies was 3.31 (95%CI, 2.31–4.75). Subgroup analyses indicated the pooled RRs were 3.12 (95%CI, 1.75–5.56) and 3.43 (95%CI, 2.21–5.31) for ADM and proADM, respectively. To explore the source of between-study heterogeneity, further analyses was performed (data not shown), showing that the pooled RRs comparing high with low level of proADM were 2.96 (95%CI, 2.12–4.13) for in-hospital mortality and 5.21 (95%CI, 3.10–8.77) for 28-day mortality. The definition of cutoff value may also bring about heterogeneity as the pooled RRs were 3.18 (95%CI, 2.21–4.56) for high cutoff value (≥4nmol/l) and 5.43 (95%CI, 3.13–9.40) for low cutoff value (<4nmol/l).

Figure 3.

Forest plot of ADM or proADM to predict mortality in sepsis.

(0.41MB).

There is no statistically significant difference when exploring for threshold effect (Spearman correlation coefficient=0.071; P=0.74). Totally ten studies were included for assessing the diagnostic accuracy of ADM and proADM for mortality in patients with sepsis. As shown in Fig. 4, the pooled sensitivity and specificity of included 10 studies were 0.72 (95%CI, 0.64–0.78) and 0.77 (95%CI, 0.69–0.83), respectively. The PLR and NLR were 3.1 (95%CI, 2.3–4.1) and 0.37 (95%CI, 0.29–0.47), respectively (data not shown). The DOR was 8.11 (95%CI, 5.25–12.54) (data not shown). The overall area under the SROC curve was 0.80 (95%CI, 0.77–0.84) (Fig. 5).

Figure 4.

Sensitivity and specificity of ADM or proADM for the predicting of mortality in patients with sepsis.

(0.3MB).

Figure 5.

Summary receiver operating characteristic of ADM or proADM for the predicting of mortality in patients with sepsis.

(0.11MB).

The Deek's funnel plot of the included studies showed that publication bias was not statistically significant (P=0.24) (Fig. 6).

Figure 6.

Deek's funnel plot asymmetry test for publication bias.

(0.06MB).