Studies retrieved and their characteristics
The database search identified 2525 records that potentially qualified for inclusion. The titles and abstracts of these records were then filtered. Full texts of 300 records were screened, and 40 met the inclusion criteria. Additional file 1: Table S2 lists the main characteristics of the 40 studies included in the meta-analysis. Of the 40 studies included, two studies included a control group that might have sepsis, and ten studies used healthy neonates as controls. We did not include these studies in the meta-analysis based on the inclusion criteria provided. Eventually, 28 studies (2661 participants) were included in the meta-analysis (Fig. 1), of which 9 studies were not written in English (Additional file 1: Table S1).
Overall, 1281 participants were assigned to the sepsis group and 1380 to the non-sepsis group. In terms of region, 17 (60.7%) trials recruited patients from Asia, 8 (28.6%) from Europe, 1 (3.6%) from North America, and 2 (7.1%) from Africa.
In terms of sepsis onset, three studies included only patients with early-onset neonatal sepsis (diagnosed in the first 72 h of life), five studies included patients with late-onset neonatal sepsis (diagnosed after 72 h of life), and the remaining trials included early-onset and late-onset neonatal sepsis or did not provide relevant information.
In terms of trial design, 13 studies were prospective cohort studies, 12 studies were case-control studies, and 3 studies were cross-sectional studies.
Risk-of-bias assessments
Figures 2 and 3 show the results of assessment for risk of bias. In terms of the risk of bias, of the 28 studies included in our meta-analysis, 12 studies had unclear bias in patient selection. There were 18 studies that were judged as having low bias in the index tests, 27 studies were allocated as having low bias in terms of reference standards, and 26 studies were judged as having low bias in terms of flow and timing. In terms of applicability concerns, 8 studies had high bias in patient selection, 17 studies were judged as having low bias in relation to index tests, and 21 studies were classified as causing high concern about reference standards.
Threshold effect and heterogeneity
The Spearman correlation coefficient and P value for PCT, CRP, PCT + CRP, and presepsin were 0.144 and 0.523, 0.301 and 0.174, 0.433 and 0.244, and 0.371 and 0.468, respectively, which indicated that there was no significant threshold effect, and thus we combined the sensitivity, specificity, PLR, NLR, DOR, and AUC. We used I2 and a bivariate boxplot (Additional file 2: Figure S1) to measure the heterogeneity caused by non-threshold effects. For PCT, CRP, PCT + CRP and presepsin, the I2 values were 96%, 98%, 0%, and 80%, respectively.
Forest plot and area under the summary ROC (SROC) curve
Forest plots of sensitivity and specificity are shown in Fig. 4. Additional file 1: Table S3 shows the pooled results of PCT, CRP, PCT + CRP and presepsin. Figure 5 shows the SROC curve for the diagnosis of neonatal sepsis. The pooled sensitivity, specificity, PLR, NLR, DOR, AUC and corresponding 95% CI (95% CI) of PCT, CRP, PCT + CRP, and presepsin were 0.85 (0.79, 0.89), 0.71 (0.63, 0.78), 0.91 (0.84, 0.95), 0.94 (0.80, 0.99); 0.84 (0.78, 0.89), 0.88 (0.80, 0.93), 0.89 (0.81, 0.93), 0.98 (0.87, 1.00); 5.4 (3.7, 7.9), 6.1 (3.6, 10.5), 8.0 (4.6, 14.0), 50.8 (6.5, 394.7); 0.18 (0.13, 0.25), 0.33 (0.26, 0.42), 0.10(0.05, 0.19), 0.06 (0.02, 0.23); 31 (17, 54), 19 (10, 35), 79 (26, 246), 864 (65, 11473); and 0.91(0.89–0.94), 0.85 (0.82–0.88), 0.96 (0.93–0.97), 0.99 (0.98–1.00), respectively.
Pair-wise comparisons
Additional file 1: Table S3 shows the results of pair-wise comparisons between statistical indicators for sensitivity, specificity, PLR, NLR, and AUC. The pooled sensitivity of CRP (0.71 (0.63, 0.78)) was weaker than that of PCT (0.85 (0.79, 0.89)), PCT + CRP (0.91 (0.84, 0.95)) and presepsin (0.94 (0.80, 0.99)) and the pooled NLR of presepsin (0.06 (0.02, 0.23)) and PCT + CRP (0.10 (0.05, 0.19)) were less than for CRP (0.33 (0.26, 0.42)), and the AUC of presepsin (0.99 (0.98–1.00)) was greater than for PCT + CRP (0.96 (0.93–0.97)), CRP (0.85 (0.82–0.88)), and PCT (0.91 (0.89–0.94)).
Likelihood ratio scattergram
For PCT and CRP, the summary LRP and LRN for index testing were on the right lower quadrant (RLQ), indicating that PCT or CRP could not exclude or confirm neonatal sepsis (Additional file 3: Figure S2). For PCT + CRP, the summary LRP and LRN for index testing was between the left lower quadrant (LLQ) and right lower quadrant (RLQ), suggesting that PCT + CRP may exclude but not confirm neonatal sepsis (Additional file 3: Figure S2). For presepsin, the summary LRP and LRN for index testing was between the left upper quadrant (LUQ), indicating that presepsin could exclude and confirm neonatal sepsis (Additional file 3: Figure S2).
Fagan diagram and publication bias
Additional file 4: Figure S3 shows the assessment of publication bias. Based on the P values of PCT, CRP, PCT + CRP and presepsin (0.430, 0.735, 0.825, and 0.410, respectively) and the corresponding Deek’s funnel plot, no significant publication bias was observed. Additional file 5: Figure S4 shows the Fagan diagrams. Based on the same pre-test probability of 20%, the post-test probability for presepsin (93%) was higher than for PCT + CRP (67%), PCT (58%), and CRP (60%).
Sensitivity analysis
Sensitivity analysis was performed with a method of reducing one article at a time, and the effect of a single study on the meta-analysis was evaluated. Additional file 1: Table S10 shows the combined DOR and 95% CI calculated after deleting a single study. We observed that regardless of the excluded study, the combined DOR after removal did not significantly change, suggesting that the results of this analysis were not excessively dependent on a certain study, and our findings were robust.
Meta-regression analysis
Meta-regression analysis of sensitivity, specificity, and joint models was performed to find potential sources of heterogeneity (Additional file 1: Tables S4, S5, and S6). According to the results of meta-regression analysis, we specified subgroups based on design, region, method, test time, and cutoff value.
Subgroup analysis
The results of the subgroup analysis are shown in Additional file 1: Tables S7, S8, and S9.
In terms of region, PCT had similar sensitivity in Asia and Europe, while its sensitivity in North America (0.98 (0.92–1.00)) was significantly higher than in Asia (0.85 (0.80–0.91)). In Asia, the sensitivity of PCT obtained at a cutoff level of 1.53 ng/mL (0.91 (0.77–1.00)) was higher than its sensitivity at a cutoff level of 1 ng/mL (0.59 (0.26–0.91)).
For CRP, in terms of region, the sensitivity of CRP in Africa (0.92 (0.80–1.00)) was significantly higher than in Asia (0.72 (0.63–0.80)) and Europe (0.63 (0.47–0.79)). In Europe, immunonephelometric assay (0.71 (0.54–0.83)) was significantly more sensitive than chemiluminescent immunoassay (0.28 (0.01–0.54)), but its specificity (0.85 (0.81–0.89)) was significantly lower than that of chemiluminescent immunoassay (0.98 (0.92–1.00)).
For PCT + CRP, in terms of region, the sensitivity of PCT + CRP in Asia (0.93 (0.90–0.97)) was significantly higher than in Europe (0.69 (0.51–0.87)).
For presepsin, in terms of region, the specificity of presepsin in Europe (1.00 (0.94–1.00)) was significantly higher than in Asia (0.90 (0.93–0.95)). In terms of study design, the sensitivity of case-control studies (0.92 (0.80–1.00)) was significantly higher than for cohort studies (0.80 (0.66–0.94)). The sensitivity and specificity of presepsin obtained at a cutoff level of 722 μg/L was higher than its sensitivity and specificity at a cutoff level of 539 μg/L.
In addition, we performed a study of the appropriate cutoff interval (Additional file 1: Table S9). For PCT, 0.5–2 ng/mL may be the appropriate cutoff interval. Moreover, the 0.5–1 ng/mL PCT range had high sensitivity (0.88 (0.82–0.95)), whereas the 1.5–2 ng/mL range had high specificity (0.90 (0.77–1.00)). For CRP, ae cutoff value > 10 mg/L had high sensitivity (0.85 (0.72–0.98)) and specificity (0.93 (0.82–1.00)).