Brain natriuretic peptide to predict successful liberation from mechanical ventilation in critically ill patients: a systematic review and meta-analysis

The primary endpoint was successful liberation from MV. Successful liberation was defined in accordance with existing literature as the absence of reintubation or application of new non-invasive ventilation in the 48 h following initial extubation [1]. We further analyzed additional data after 48 h as available. Secondary endpoints initially described in our protocol could not be analyzed due to insufficient available data.

Diagnostic accuracy of BNP was measured using sensitivity and specificity. In this instance, a true positive was represented by the combination of a low absolute or low variation of BNP in a patient that had successful liberation from MV. A true negative was represented by the combination of a high absolute or high variation of BNP in a patient that had unsuccessful liberation from MV. Where possible, results across studies were simultaneously pooled using a bi-variate model to create both a joint estimate of sensitivity and specificity (with 95% confidence regions) as well as hierarchical summary ROC curve (HSROC). For subgroup analyses, we were unable to use the bi-variate model due to insufficient studies. In these cases, the Moses-Littenberg model was used to estimate a summary ROC curve. When considered sufficiently clinically homogeneous, areas under the ROC curve (AUROC) were pooled when possible using a DerSimonian-Laird random effects model.

Quality of each study (JD and SA) was assessed using the QUADAS-2 questionnaire for systematic reviews [19]. Ten parameters across four domains (patient selection, index test, reference test, and flow/timing) were analyzed independently by two authors (JD and SA) and disagreements resolved through discussion. This was performed in duplicate by two independent reviewers (JD and SA). We applied the strict QUADAS-2 method of assigning low risk or at risk status for each domain. We did not perform a GRADE assessment as described in the protocol, as this was deemed not applicable to this type of research.

Of the 18 adult studies, there were 28 individual analyses of BNP relative to SBT (Table 1). The methods of BNP measurements included pre-SBT BNP measures (BNP-pre), post-SBT BNP measures (BNP-post), absolute BNP change during SBT (ΔBNP), and the relative BNP change during SBT (ΔBNP% = [post-SBT BNP − pre-SBT BNP]/pre-SBT BNP). Four studies described more than one measure of BNP. Some studies did not provide a ROC analysis and relied on odds ratio (OR). A subset of studies assessed BNP as an alternative tool to reclassify all patients by including SBT failure in the liberation failure group for analysis (group 1; Fig. 2). The other subset assessed BNP as an incremental tool in addition to SBT testing and excluded SBT failure from the analysis (group 2; Fig. 2). This was unclear in one study [20]. All studies relied on clinical indices of SBT to determine readiness of extubation, and no decision to extubate was done based on BNP measures. All studies excluded preventative non-invasive ventilation from the cohort of patients being analyzed.

Patient baseline characteristics, primary diagnoses, acuity scores, and ventilatory parameters were reported by status of liberation from MV (Additional file 4). These baseline data points were not classified according to BNP assessment results combined. All studies except one [19] evaluated the first liberation attempt (n = 17). All studies observed patients for extubation failure at 48 h (n = 18), with one study extending observation to 7 days [19]. The quality of included studies focused on adults suggested a risk of bias for all but one low-risk study [21] (Table 3). The main reasons for risk of bias were patient selection and lack of blinding of the index test and reference test.

Of the two studies focused on children, one study addressed preterm infants with respiratory distress syndrome, and the other addressed congenital heart surgery patients (Table 2). Pooled analysis could not be performed due to the lack of available data. As such, these studies were excluded from the pooled analysis (Table 3). Both studies were at risk of bias (Table 4).

Only 13 of the 20 studies included had sufficient data for pooled meta-analysis. We performed a single bi-variate estimate of sensitivity and specifity combining studies that reported absolute and relative changes of BNP (ΔBNP or ΔBNP%) in studies that either included or specifically excluded patients with SBT failure from analysis (n = 5). The sensitivity and specificity were 0.889 (0.831–0.929) and 0.828 (0.730–0.896), respectively (Fig. 3). This was further stratified in Moses-Littenberg summary ROC curves for either measure of BNP (ΔBNP or ΔBNP%) and for ΔBNP% only in studies that excluded patients with SBT failure (Additional file 5).

AUROC were pooled separately for ΔBNP (n = 3, 0.89 [0.83–0.95], I² = 67%) and ΔBNP% (n = 5, 0.92 [0.87–0.96], I² = 28%) regardless of inclusion or exclusion of patients failing a SBT (Fig. 4); for ΔBNP% for studies that excluded patients failing a SBT (n = 3, 0.92 [0.88–0.97], I² = 0%) (Fig. 5); and for BNP-pre (n = 4, 0.77 [0.63–0.91], I² = 87%) and BNP-post (n = 4, 0.85 [0.80–0.90], I² = 16%) regardless of inclusion or exclusion of patients failing a SBT (Fig. 6).

No meta-analysis of the specific thresholds of BNP measures could be performed due to insufficient data. A net reclassification index could not be calculated given the limited data assessing BNP measures in SBT failure groups. Only the pediatric study by Zhang et al. [22] calculated a NRI of 0.224 for the addition of NT-ProBNP to SBT, suggesting an improvement in reclassification.

This meta-analysis supports the validity of the relative change of BNP (ΔBNP%) during a SBT to add incremental value and inform the likelihood of successful liberation from MV in adults. This meta-analysis also demonstrated high accuracy using a pooled AUROC of ΔBNP% for studies that excluded patients who failed a SBT. Combining methods of absolute and relative change in BNP measures, irrespective of inclusion or exclusion of patients who failed their SBT showed high sensitivity and specificity for predicting successful liberation. The data from pediatric studies and studies describing other BNP measures, such as ΔBNP, BNP-pre SBT, and BNP-post SBT, were insufficient to suggest incremental value and for clinical decision support about likelihood of liberation success.

There findings are noteworthy given the limited predictive ability of SBT alone, which is generally regarded as the best available assessment. In studies, SBT misclassified patients in 10–20% of patients who subsequently failed successful liberation from MV [1, 12]. While reintubation was described as occurring without immediate difficulty, these patients had greater risk of morbidity and mortality following a failed attempt at liberation from MV [1]. Better prediction by use of alternative tests that add incremental value, such as BNP, may lead to greater confidence in clinical decision-making to extubate, reduced reintubation, and improved outcomes. This possibility has been recognized as early as 2008 in two methods of analysis: as an incremental “value-added” test during SBT [3] and as a “stand-alone” alternative test [5]. These two approaches were well represented in the studies included in this meta-analysis. A subset of studies (group 1, n = 8) included patients failing their SBT in the analysis of the group that failed MV liberation; this in effect assessed BNP as an alternative test to conventional SBT. This method may have decreased accuracy compared to evaluation of pooled analysis where patients who failed their SBT were excluded. A second subset of studies (group 2, n = 9) excluded patients with a failed SBT in the analysis of the MV liberation failure group. This in effect assesses BNP as an incremental test to a successful SBT. The major benefit in this case is the potential reclassification of patients for whom liberation may have been attempted but may have failed. This distinction is important to determine the optimal use of BNP in assessment for MV liberation. In our view, the two ways in which SBT failure is incorporated in the analysis should ideally be pooled and analyzed separately. However, we expected limited data and pooled them for further analysis as planned in the protocol. Similarly, the different methods of BNP measures (ΔBNP, DeltaBNP%, BNP-pre, and BNP-post) should not be pooled, as some address a change in BNP values, whereas others only address a single value at a specific time. The only exception in which BNP measures could be pooled would be ΔBNP and ΔBNP% given the possibility that baseline BNP level may not be relevant in the case of substantial change occurring during a SBT.

Most of the data that could be pooled related to absolute and relative changes in BNP during a SBT (ΔBNP and ΔBNP%). In order to increase the breadth of our analyses, we pooled studies of either method of measures (ΔBNP and ΔBNP%) for studies that excluded patients who failed their SBT. The Moses-Littenberg summary ROC analysis showed high accuracy (Additional file 5). However, this analysis requires the availability of true-positive, true-negative, false-positive, and false-negative data to perform, limiting its applicability to certain subgroups. This summary ROC analysis was mostly driven by ΔBNP% (3 out of 4 studies, 148 out of 178 patients). We were able to perform a pooled AUROC of ΔBNP% for studies that excluded SBT failure from the analysis of the liberation failure (Fig. 5). This AUROC analysis further supports the initial findings and provides evidence of high accuracy (0.92 [0.88–0.97], I² 0%). This represents the most robust combination of BNP measures in pooled analysis obtained from the data.

Unfortunately, the data was insufficient to perform sensitivity and specificity estimates for this specific combination of BNP measures. The closest approximation was obtained by a bi-variate analysis using pooled data of studies of either ΔBNP or ΔBNP% measures, regardless of inclusion or exclusion of the SBT failure group. The sensitivity and specificity obtained were high [0.889 (0.831–0.929) and 0.828 (0.730–0.896), respectively]. It is important to note that these results were mostly driven by studies, where patients with SBT failure were excluded (4 out of 5 studies; 248 out of 278 patient s), and ΔBNP% (4 out of 5 studies; 178 out of 278 patients). This closely approximates the prior analysis performed on ΔBNP% for studies that excluded patients with SBT failure from the analysis.

There were insufficient studies to analyze ΔBNP, BNP-pre, and BNP-post separately as an incremental test (i.e., excluding SBT failure from the liberation failure analysis) or an alternate “stand-alone” test (i.e., including SBT failure from the liberation failure analysis). Pooling studies of both methods of analysis for each BNP measure appears to support a high accuracy in these cases (Fig. 4). The main limitation is the inability to determine if it is of better use as an incremental or alternate test. Additionally, for BNP-pre and BNP-post, the studies were not described with as much detail as those for studies on ΔBNP and ΔBNP%. Furthermore, it is difficult to determine a superior measure or method, as only two studies directly compared various BNP measures. Both Cheng et al. [23] and Martini et al. [24] compared ΔBNP and ΔBNP%, and both studies suggested ΔBNP% was superior.

From a clinical standpoint, using these measures requires using a specific threshold for dichotomization between likelihood of liberation success versus failure (Table 1). This was determined through analysis of the AUC curve of best sensitivity, specificity, positive and negative predictive values, and diagnostic accuracy. No pre-specified threshold was studied prospectively across any of the studies. Studies of ΔBNP% suggested a threshold above 13.4–20% (n = 4) optimally predicted liberation failure, if both BNP types (BNP and NT-ProBNP) were pooled. The other BNP measures had three or fewer studies for each measure or combination (Table 1). As such, this systematic review cannot recommend a specific threshold for any of the BNP measures to optimally discriminate liberation success and failure that could be adopted in clinical practice.

Our study would suggest that BNP performs best if used as a relative change BNP during a SBT among those studies that only included adult patients who successfully passed an initial SBT by other clinical criteria. There is a potential role to the use of either ΔBNP% or ΔBNP irrespective of whether patients have successfully or unsuccessfully passed an SBT, but the data is less robust and requires further investigation.

As described above, the heterogeneity of BNP measures and varying analytic approaches limited our ability to perform pooled analysis; acknowledging this limits the inferences that can be made. Similarly, due to limitation in the reporting across studies, we were unable to perform stratified analysis by potentially important subgroups, such as case-mix and type of ICU admission. We opted to combine both general ICU populations and specific ICU subgroups to capture enough data to perform the AUROC analyses. The bi-variate analysis and Moses-Littenberg analyses were unaffected, as all studies included were of a mixed ICU population. We believe that this makes our results more generalizable to mixed medical/surgical ICU practice; however, we cannot provide strong inferences on the value of BNP for MV liberation in selected ICU subgroups, as each population was represented by a single study. There are several confounders to the accuracy of BNP testing. Heart disease (and specifically depressed left ventricular ejection fraction) and kidney failure can significantly alter BNP kinetics. Unfortunately, these patient characteristics were inconsistently included or excluded across studies. In the case of kidney failure, the distinction between acute and chronic renal failure was also poor. In balance, renal function was normal in most studies, and at most mildly impaired in the rest. As for heart disease, the definitions were variable. The etiology of respiratory failure has an impact on the accuracy of BNP: delirium, traumatic brain injury, inability to clear secretions, or stridor, among others, limit the accuracy of BNP as they may not lead to a change in BNP measurements. A low number of studies that were included in this review attempted to limit this impact by excluding stridor and TBI from the analysis. Unfortunately, capturing clearance of secretions or delirium as the cause of respiratory failure is understandably difficult and was not done in any study. Another limitation is the lack of studies that directly compared the accuracy of a successful SBT by clinical indices and by BNP measure. In this instance, a patient that has passed a SBT by clinical indices may fail by BNP measure, leading to a delay in extubating a patient that would have succeeded. Unfortunately, the relative accuracies were not directly assessed in any study.

Finally, the quality of studies (as defined by QUADAS-2) uniformally ranked as at risk of bias, except for one [21]. The main issue was lack of transparency regarding blinding of physicians to the BNP test. In our opinion, this is not a critical flaw, as the decision to extubate patients was most often based on clinical SBT criteria in all studies.

Research on mechanical ventilation liberation is complex and would benefit from greater standardization. Successful liberation from mechanical ventilation appears well-defined and this is reflected in the studies collected. Liberation failure, on the other hand, has a variable definition among studies, mostly relating to the inclusion or exclusion of SBT failure. Regardless of its importance for applicability of alternative or incremental testing, the terms used require standardization to facilitate research.

Additional data is needed to strengthen BNP as a liberation tool. We consider that this should take the form of a comparative study of BNP as an alternative or an incremental tool to the clinical indices after an SBT. This study would ideally take the form of an assessment of ΔBNP and ΔBNP% and compare inclusion versus exclusion of SBT in the analyzed subgroups. This would allow determination of whether BNP is superior to SBT on its own or simply incremental.

The potential benefits of improved tools to inform greater likelihood of success or failure of liberation from MV have far-reaching implications. On top of reclassifying individuals after initial assessment with clinical indices, this may allow stratification of the risk of failure. Such a stratification may help better determine targets for optimization (such as further volume de-escalation), better identification of the need for post-extubation therapies (such as high-flow oxygen therapy and BiPAP), and need for prolonged ICU observation. Clinical risk scores on this basis could be developed to aid in management of these patients after extubation.