Procalcitonin for the diagnosis of postoperative bacterial infection after adult cardiac surgery: a systematic review and meta-analysis

The PRISMA flow diagram for identification, screening, and inclusion of studies is shown in Fig. 1.

The original search performed on April 12th 2023 retrieved a total of 1855 records, which were uploaded into the Rayyan platform. After deduplication, 1544 records underwent manual title and abstract screening, of which 57 were identified as potentially eligible and underwent full text review. We excluded 46 reports [33‐78] (see Additional file 2: Table S2), leaving 11 eligible studies which were included in our systematic review [17, 79‐88]. Search rerun on September 15th, 2023 retrieved additional 130 deduplicated records, none of which was selected for full text review. Also, no additional eligible study was identified from reference lists of relevant papers.

Table 1 displays the characteristics of the 11 included studies. Overall 3803 patients (range from 40 to 819 per study) were involved.

All studies were single-center with observational design, five of which with retrospective data collection [17, 83, 84, 86, 87]. The vast majority was conducted in Asia (eight in China [17, 80, 82‐84, 86‐88], two in India [79, 85]), and only one in Europe [81].

The target condition was generically indicated as bacterial infection in six studies [17, 79, 85‐88], whereas five studies focused exclusively on pulmonary infection [80‐84]. The reference standards used to define infection varied. Three studies applied Centers for Disease Control (CDC) criteria [81, 83, 86], and the others all used positive cultures, either alone [17, 79, 85], or in combination with different parameters including cultures, imaging, laboratory findings, and clinical signs [80, 82, 84, 87, 88] (Table 1). Only one study did not report the technique adopted for measuring plasmatic PCT [88], while all other studies used the chemiluminescence immunoassay. However, only five studies provided information on the specific assay and its sensitivity range [79‐81, 84, 87].

Timing of PCT measurement also varied, with four studies performing only one measurement, three studies on the first PostOperative Day (POD) [83‐85], and the other at ICU admission [88]. The longest reported monitoring period was POD 5 in four studies [17, 80, 82, 86].

The methodological quality assessments with the QUADAS-2 tool results are summarized in Fig. 2 and further illustrated for individual studies in Fig. 3.

No study had a low risk of bias in all 4 domains. For the domain of risk of bias in patient selection, only five studies provided clear definitions of exclusion criteria and were judged as ‘low’ risk. Regarding the risk of bias for index tests, none of the studies prespecified a threshold and therefore they were all rated as ‘high risk’. Only one of the studies was judged to be at high risk of bias for the reference standard domain and for the patient flow and timing domain [79]. Seven studies were rated as ‘low’. Only three studies [79, 86, 88] were considered to have concerns about applicability, all in terms of patient selection. Further details on how judgments were made for each individual study are provided in Additional file 3: Table S3.

In the light of the issues that emerged from the risk of bias assessment, ten of the eleven studies were included in the meta-analysis. The study by Chakravarthy et al. [79] was excluded, because it exhibited high risk of bias in three domains and because it did not specify the execution time of the index test, making it impossible to attribute the outcome to a specific postoperative day.

Figure 4 shows the diagnostic accuracy of PCT in detecting bacterial infection after cardiac surgery, as reported in each of the 10 studies (2984 patients) included in the meta-analysis. The forest plots highlight the heterogeneity in test timing and in thresholds reported by each study, and in the corresponding values of Se and Sp and their 95%CI. The two diamonds represent, respectively, the pooled estimation of Se (0.70, 95%CI 0.67–0.73) and Sp (0.76, 95%CI 0.71–0.81). Concerning heterogeneity, through univariate analysis independent by thresholds, we determined values of I² = 15.5 and Q = 28.4, which do not highlight significant heterogeneity (p = 0.243).

Concerning other diagnostic accuracy values, pooled median PLR, NLR and DOR of PCT were 2.96 (95%CI 2.33–3.74), 0.40 (95%CI 0.35–0.46), and 7.53 (95%CI 5.18–10.60), respectively. Based on the meaning attributed to the PLR value, a diseased patient is nearly three times more likely to have a positive test compared to a non-diseased patient; conversely, considering NLR, a non-diseased patient is 2.5 times more likely to have a negative test compared to a diseased patient. Furthermore, the value of DOR indicated that for PCT the odds for positivity among subjects with bacterial infection were nearly eight times higher than the odds for positivity among subjects without bacterial infections.

Results of bivariate model analysis using multiple thresholds within a study are depicted in Fig. 5.

The first two scatterplots from the top (panel A and B) show the optimal threshold as 3 ng/mL (with corresponding Se 0.67 (95%CI 0.47–0.82) and Sp 0.73 (95%CI 0.65–0.79)), which allows to best identify the diseased and non-diseased groups (solid and dashed lines) in terms of probability positive test and in terms of the corresponding maximum value of the Youden index.

The two lower scatterplots (panel C and D) display the individual ROC curves for each study and the SROC curve corresponding to the optimal threshold. The AUC of the SROC is of 0.75 (IC95% 0.29–0.95), which is considered to be “good” diagnostic accuracy [89] even though wide variability was observed.

Table 2 reports performance measures, calculated considering prespecified ranges of thresholds and prevalence. Predictive values are further illustrated by continuous lines in Fig. 6, in which the threshold range is amplified (up to 20). As evident in Panel A, PPV varies approximately between 0.50 and 0.70, when prevalence is high (30%). Regarding NPV, the value is always > 90% when prevalence is < 20% (regardless of the threshold), and is reduced to 83% when prevalence is high (30%).

The results of the analysis where the interaction term threshold*day was included are displayed in Additional file 4: Table S4. The corresponding coefficient value is equal to − 0.24 (95%CI − 0.48 to 0.00), implying that the threshold should be decreased by 0.24 points per day. Although this finding is close to statistical significance (p = 0.053), for explorative purposes we examined it for each of the 6 PODs (Fig. 7). Starting from POD 1 to POD 4, the FN rate is reduced as the threshold decreases. This is especially true on POD 2, for which the finding is statistically significant (p = 0.019) (see Additional file 5: Table S5), identifying it as the probable best time point to use PCT for the diagnosis of infection.

To our knowledge, this is the first systematic review and meta-analysis investigating the role of PCT for the diagnosis of postoperative infection only including adult patients after cardiac surgery. Our meta-analysis, including 10 studies and 2984 patients, assessed the diagnostic test accuracy of PCT, considering different thresholds and different time points reported in included studies. Bivariate analysis using multiple thresholds within a study enabled us to highlight important characteristics of the diagnostic test. Specifically, we identified the optimal threshold value at 3 ng/mL, which is considerably higher than the 0.5 to 1.0 ng/mL range generally recommended for the diagnosis of postoperative infection[8]. However, even when considering this optimal threshold, test performance was limited, with a sensitivity of 67% and specificity of 73%. These findings may be due to the presence of systemic inflammation immediately after surgery, a hypothesis also supported by our analysis of the interaction between threshold and POD, which suggested that the threshold should be reduced daily to improve PCT diagnostic accuracy, and especially to increase the positive predictive value. Our analysis also suggested that POD 2 may be the best timing to diagnose infection with PCT, an indication also reported by other studies [82, 91]. Another interesting aspect worth noting, particularly relevant for clinical practice, is the test’s considerable ability to identify non-diseased individuals (NPV between 83 and 98%, with a prevalence range between 30 and 5%), and its poor utility in identifying diseased patients (PPV never exceeding 60%, even considering a high prevalence of 30%). This suggests that the use of procalcitonin in this context is useful to exclude, and not to confirm, the presence of a bacterial infection.

Concerning risk of bias assessment, various problems were detected. One of the main issue concerned the fact that threshold determination occurred a posteriori by ROC curve analysis in all studies, which may have led to optimistic test performance. Moreover, none of the studies was multicenter and none formally defined sample size a priori considering study endpoints.

Comparison of our results with other meta-analyses was not possible, because the only one published recently on this topic [16] considered both adults and children, and the analysis model used did not take into account the different thresholds reported in individual studies.

This systematic review was conducted following rigorous methodology, for search strategy development, evidence analysis and quality appraisal, involving a multiprofessional research team. One of the main strengths of this work lies in the advanced meta-analysis methods used to summarize data according to multiple threshold values in each study. Furthermore, the use of strict eligibility criteria for our review (clear definition of target condition diagnosis, only adult populations and only publications from the last 10 years) helped reduce heterogeneity, thus improving generalizability of results. In particular, the decision to apply a date restriction was due to the fact that perioperative standards of care (e.g. surgical techniques, extracorporeal circulation, Intensive Care Unit (ICU) care, etc.) have improved considerably in the last decade, leading to a reduction of surgery-related stress, and thus of SIRS, which may be misclassified as infection [92‐94]. Although minimally invasive cardiac surgery, miniaturized and biocompatible CPB circuits, and fast-track protocols were all introduced over 20 years ago, their implementation has accelerated over the past decade [94‐97].We also decided to exclude patients with transcatheter interventions, as these procedures are associated with a significantly lower degree of systemic inflammation, are usually performed on older, sicker patients, and could therefore impact on the generalizability of the results to the cardiac surgical population [98‐100].

Some limitations of this work should also be acknowledged. Firstly, we only included studies that clearly indicated the diagnostic criteria applied to confirm infection, which may have lead to exclude relevant studies that did not report this aspect accurately. Unfortunately, we could not verify this potential bias with funnel plots, since this is not feasible in meta-analyses of diagnostic studies with multiple thresholds. Furthermore, the decision to apply a date restriction might have led to the exclusion of relevant studies. Secondly, included studies used different reference standards, which may have affected reliability of results. Furthermore, we acknowledge that although the analyzed literature aimed to exclude patients with preoperative infection, cases of undiagnosed preoperative infection cannot be ruled out, and this may have influenced results. Thirdly, in all studies, even when PCT measurements were taken on different days, the number of patients at risk considered for measuring test accuracy remained constant. This may have influenced the determination of the optimal threshold. Moreover, this prevented an unbiased estimation of the threshold for each POD. Finally, all included studies are observational, five of which with retrospective data collection, including one case–control study. This may have influenced reliability of results.