Study objectives
The primary objective of this meta-analysis was to assess the prognostic value of elevated BNP or NT-proBNP levels to predict short-term mortality (in-hospital or up to 40-day all-cause mortality) in patients with acute PE. The secondary objectives were to evaluate whether BNP increases are associated with short-term mortality resulting from PE (cause-specific mortality) or with SAEs.
Study endpoints
Total death and death resulting from PE were adjudicated by the authors of the individual studies. Death resulting from PE was related to irreversible RV failure or recurrent PE. SAEs were the composite of death and any of the following adverse outcome events: shock, need for thrombolysis, nonfatal PE recurrence, cardiopulmonary resuscitation, mechanical ventilation, catecholamnine administration, and surgical embolectomy.
Search strategy
The authors reviewed PubMed, BioMed Central, and the Cochrane database and conducted a manual review of article bibliographies. Unrestricted database searches until March 2008 were performed using the combined medical subject headings for 'BNP', 'pulmonary embolism', 'outcome', 'prognostic', and 'NT-proBNP' with the exploded term 'acute pulmonary embolism'. The retrieved studies were carefully examined to exclude potentially duplicate or overlapping data. Meetings abstracts were excluded as they could not provide adequately detailed data and their results might not be final. Only papers evaluating the role of BNP or NT-proBNP on patient outcomes (death or SAE) were included. Studies were eligible regardless of whether they referred to subjects with small or severe PE.
Study eligibility
We included a study if (a) it used BNP or NT-proBNP biomarkers as a diagnostic test in patients with documented PE (using a conventional threshold for positivity of the test), (b) it reported death as the primary endpoint of the study and/or SAEs in relation to BNP testing, or (c) it reported deaths and SAEs in absolute numbers for calculation of true-positive (death with BNP increased), false-positive (survival with BNP increased), true-negative (survival with normal BNP level), and false-negative (death with normal BNP level) results or presented sufficiently detailed data for deriving these figures or were provided by the authors when their studies did not report the full data. Studies were excluded if they were performed (a) in patients without certitude of PE, (b) in a subset of patients with cardiogenic shock, or (c) with fewer than 20 enrolled patients as there is a higher risk of invalid results due to selection bias.
Data extraction
The following information was extracted from each study: first author, year of publication, and journal; study population characteristics, including sample size (number of subjects evaluated with BNP tests and number of patients excluded); number of patients with documented PE; gender; mean age (and standard deviation); relative timing of BNP assessment; technical characteristics of the BNP test and threshold, including type and brand of test used; and rate of short-term death and rate of SAEs as previously defined according to BNP or NT-proBNP tests. Two investigators (GC and MH) performed the data extraction independently. Disagreements were resolved by discussion and consensus. The study was conducted according to MOOSE (Meta-analysis Of Observational Studies in Epidemiology) guidelines [13]. Unlike randomized controlled trials, no generally accepted lists of appropriate quality criteria for observational studies are available. Rather than producing a simple arbitrary quality score, specific quality aspects were used to assess the studies such as control of confounding factors, minimization of selection bias with clear description of inclusion and exclusion criteria, description of the baseline characteristics of the cohort, completeness of follow-up, clear definition of study outcomes, relative timing of the biomarker assessment after patient admission, and whether or not the investigator responsible for BNP measurements was unaware of the patients' baseline characteristics or clinical course.
Data synthesis and statistical analysis
Categorical variables from individual studies are presented as n/N (number of cases/total number of patients, percentage), and continuous variables are presented as mean values. Measures of odds ratio (OR) and of diagnostic performance are reported as point estimates (with 95% confidence intervals [CIs]). The main analysis was performed on the prognostic value of BNP testing to predict death. Secondary analyses combined the available SAE data to calculate prognostic performance.
By means of true-positive, true-negative, false-positive, and false-negative rates, we computed sensitivity, specificity, positive and negative likelihood ratios, and ORs. While predictive values are well known as measures of diagnostic accuracy, their results may be influenced by the prevalence of disease in tested subjects. The positive likelihood ratio (the ratio between sensitivity and 1 – specificity) provides an estimate of the probability of a positive test in a patient with disease, and the negative likelihood ratio (the ratio between 1 – sensitivity and specificity) gives an estimate of the probability of a negative test among diseased subjects. Both likelihood ratios are roughly independent from prevalence rates, and there is consensus that a positive likelihood ratio of greater than 10 and a negative likelihood ratio of less than 0.1 provide reliable evidence of satisfactory diagnostic performance. While likelihood ratios are the recommended summary statistics for systematic reviews of diagnostic studies, predictive values may also be of interest for clinicians, even if these values vary widely in their dependence on disease prevalence. Such limitations of predictive values notwithstanding, these figures were also computed and reported as exploratory data in this review. Weighted symmetric summary receiver operating characteristic plots, with pertinent areas under the curve, were computed using the Moses-Shapiro-Littenberg method.
We computed all statistics for individual studies, then combined them using a random-effects model, weighting each point estimate by the inverse of the sum of its variance and the between-study variance. Between-study statistical heterogeneity was also assessed using the Cochran Q chi-square test and the I2 test. Separate analyses were performed on studies with BNP and proBNP assessments. Publication bias was assessed visually by examination of funnel plots. Statistical computations were performed with SPSS 11.0 (SPSS Inc., Chicago, IL, USA), Meta-DiSc [14], and Review Manager 4.2 [15], and significance testing was at the two-tailed 0.05 level.