Sample and Design Characteristics of Studies

Fourteen articles met all of the inclusion criteria and were selected for data abstraction.^{7, 16– 18, 48– 57} Of the 14 selected studies, four were from the Breathing Not Properly Multinational Study.¹⁸ The data from the sub-studies^{49, 51, 54} were excluded from the meta- analyses. This study and seven others examined only BNP.^{17, 48, 50, 52, 55– 57} Two other studies examined only NT-proBNP, ^{16, 53} and one study examined variations of both BNP and NT-proBNP.⁷ The included studies were published over a period of four years (2001 - 2004) with the majority published in 2002 and 2004 (Table 4). The patients enrolled in all studies presented to emergency departments with shortness of breath and were over 18 years of age. One study⁵⁷ limited enrolment to patients over 65 years of age while another¹⁶ limited enrolment to patients between 44 and 88 years of age.

Table 4

Diagnostic properties of studies that evaluated BNP and NT-proBNP in patients with symptoms suggestive of HF in emergency or urgent care settings.

Diagnosis of HF in studies. All studies except two^{52, 53} selected for data abstraction employed a cohort design and a reference standard agreed upon by consensus of at least two physicians (mostly cardiologists). Two studies based the diagnosis on the opinion of a single cardiologist^{48, 52} and the third only stated that the definitive diagnosis was based on the Framingham criteria and echocardiography results.⁵³ The adjudicating physicians each arrived at a diagnosis of HF based on their interpretation of all available clinical data, often including echocardiography results. The Boston Criteria were employed in the diagnosis in one study⁵² and the Framingham criteria in three studies,^{18, 53, 55} one of which also applied the National Health and Nutrition Examination Survey (NHANES).¹⁸

Diagnostic properties. Table 4 presents the results to answer the question “What are the clinical performance characteristics of both BNP and NT-proBNP measurement in patients with symptoms suggestive of HF or with known HF presenting to an emergency department?”

The 12 studies evaluating BNP utilized several cut point values ranging from 50 to 400 pg/mL and reported sensitivities from 60 percent to 100 percent, specificities from 27 to 99 percent, and areas under the curve (AUC) of 0.67 to 0.99.^{7, 17, 18, 48– 52, 54– 57} In addition, the reported positive likelihood ratio (LR+) ranged from 0.69 to 70 and the negative likelihood ratio (LR-) ranged from 0.00 to 0.44 (Table 4). The three studies evaluating NT-proBNP utilized several values ranging from 254 to 4567 pg/mL and reported sensitivities from 74 percent to 98.6 percent, specificities from 47 to 93 percent, and AUC values of 0.89 to 0.96. It was possible to do meta-analysis on eight studies for BNP^{7, 17, 18, 48, 52, 55– 57} and three studies for NT-proBNP.^{7, 16, 53} To maximize sensitivity the lowest cut point was used if multiple cut point data were given. The data are summarized in Table 7 and Figures 3 and 4. The sensitivities of the BNP studies were similar with a summary sensitivity of 97 percent and a CI of 96 to 98 percent. In contrast, the specificity data was very heterogeneous with a summary estimate of 70 percent and a CI ranging from 56 to 85 percent (see Appendix C, Table 13–15 for results of tests for heterogeneity with regards to setting). The corresponding likelihood ratios (LRs) showed that the LR- (0.06, 95 percent CI: 0.03 to 0.10) was better than the LR+ (3.63, 95 percent CI: 2.49 to 5.31) in terms of diagnostic value. The diagnostic odds ratio (DOR) for BNP ranged from 13 to 1635 with a summary estimate of 81 (95 percent CI: 29 to 219). With the exception of the pooled sensitivity estimates (Figure 3a) for BNP in the ED, all other combined estimates (for specificity, LR+, LR-, and DOR) had positive tests for heterogeneity; as such our confidence in these pooled estimates is decreased.

Table 7

Diagnostic odds ratios for studies that evaluated BNP and NT-proBNP in patients with symptoms suggestive of or with HF across all settings.

Figure

Figure 3. Forest plots for BNP in all settings using the lowest cut point provided in each study: a) sensitivity, b) specificity, c) LR+, d) LR-, e) DOR.

Figure

Figure 4. Forest plots for NT-proBNP in all settings using the lowest cut point provided in each study: a) sensitivity, b) specificity, c) LR+, d) LR-, e) DOR.

For NT-proBNP the summary estimates were similar to BNP in that the sensitivity (95 percent, 95 percent CI: 90 to 101) was much higher than the specificity (72 percent, 95 percent CI: 53 to 90). The LR- (0.07, 95 percent CI: 0.02 to 0.27) was also better than the LR+ (3.35, 95 percent CI: 1.75 to 6.41). The DOR from these three studies assessing NT-proBNP ranged from 17 to 291 with a summary estimate of 60 (94 percent CI: 9 to 407). All pooled estimates of NT-proBNP diagnostic accuracy measures were significant for heterogeneity for ED studies.

There were six studies that provided diagnostic information at a BNP cut point of 100 (±5) pg/mL.^{7, 17, 18, 55– 57} The meta-analysis on these studies shows a similar pattern to that described for varying cut points (Figure 5); with the exception of LR-, all other diagnostic pooled estimates were positive for heterogeneity. The sensitivity summary estimation is 95 percent (95 percent CI: 91 to 96) with a lower and broader specificity summary estimation (67 percent, 95 percent CI: 53 to 80). The LR+ was 3.4 (95 percent CI: 2.14 to 5.42) and the LR- was 0.11 (95 percent CI: 0.08 to 0.15), which is higher than the lowest cut point summary estimate. The overall DOR for this group of studies was reduced to 38 but the 95 percent CI was tighter (17 to 85) compared to the lowest cut point summary estimate.

Figure

Figure 5. Forest plots for BNP in the ED using a cut point of 100 (±5) pg/mL: a) sensitivity, b) specificity, c) LR+, d) LR-, e) DOR.

Quality assessment of studies. Results from the application of the QUADAS Question 14, quality assessment tool are as follows (see also Appendix C Evidence Figures, Figure 1): one (9.1 percent) of the studies clearly addressed the issue of disease progression bias (QUADAS Question 4); the reference standard was independent of the index test result (QUADAS Question 7) in 10 (90.9 percent) of the studies; the reference standard was described in sufficient detail (QUADAS Question 9) in seven (63.6 percent) of the studies; interpretation of the peptide marker (BNP or NT-proBNP) measurement was clearly without knowledge of the reference test results (QUADAS Question 10) in one (9.1 percent) of the studies; interpretation of the reference test results was clearly without knowledge of the B-type natriuretic peptide marker results (QUADAS Question 11) in seven (63.6 percent) of the studies; none (100 percent) of the studies stated whether or not the clinical data was available when the B-type natriuretic peptide test results were interpreted as would be the case when the test is used in practice (QUADAS Question 12); one (9.1 percent) of the studies reported uninterpretable or intermediate test results (QUADAS Question 13) and; withdrawals were not explained in two (18.2 percent) of the studies (QUADAS Question 14). Overall, the quality of these studies was good.

Sample and Design Characteristics of Studies

There were a total of six papers eligible for review published between 1997 and 2004.^{11, 58– 62} All studies evaluated BNP with the exception of two^{58, 60} which compared both BNP and NT-proBNP. The studies were conducted in Austria, Japan, Portugal and USA. Two studies were based on patients referred to a HF clinic^{11, 59} and the remaining, to outpatient settings.^{58, 60– 62} All studies provided evaluation on BNP and two compared BNP and NT-proBNP.^{58, 60} Three of these papers provided data on sensitivity, specificity, and ROC curves for BNP or NT-proBNP ^{11, 58, 60}. Three of these papers did not provide ROC characteristics or sensitivity or specificity data; instead, only correlation data for BNP or NT-proBNP with different variables of cardiac structure, function and symptoms were provided.

Diagnosis of HF in studies. Three studies^{58, 60, 62} used echocardiography as the reference standard, one study used echocardiography plus clinical criteria¹¹ and two studies^{59, 61} used the NYHA classification.

Diagnostic properties. Hammerer-Lercher et al.⁶⁰ directly compared the diagnostic values of NT-proBNP with BNP in 57 patients with stable chronic HF. In the analysis of normal (echocardiographic ejection fraction (EF) ≤ to 48 percent or radionuclide angiographic EF ≤ to 55 percent) versus impaired (echocardiographic EF < 48 percent or radionuclide angiographic EF < 55 percent) LVEF the AUC for BNP was 0.75 (SE ± 0.06), and for NT-proBNP was 0.67 (SE ± 0.07) (Table 5). In the analysis of LVEF less than 40 percent versus greater than or equal to 40 percent, the AUC for BNP was 0.83 (SE ± 0.06), and for NT-proBNP was 0.79 (SE ± 0.07). Positive and negative LRs were 3.17 and 0.35, respectively for BNP and 2.59 and 0.41, respectively for NT-proBNP. NT-proBNP did not differ significantly from BNP in either of the analyses. The optimal discriminator values were 142 pg/mL for BNP, and 4127 pg/mL for NT-proBNP for the detection of LVEF less than 40 percent compared to greater than or equal to 40 percent. For these discriminators the sensitivity was 73 percent for BNP, and 70 percent for NT-proBNP. The specificities were 77 percent for BNP, and 73 percent for NT-proBNP.

Table 5

Diagnostic properties of studies that evaluated BNP and NT-proBNP in patients with symptoms suggestive of or with HF in outpatient or specialty clinic settings.

Bettencourt et al.¹¹ studied 100 patients with symptoms suggestive of HF referred to a HF clinic. These patients had suspected or not previously investigated HF. Since healthy controls were included in this study, this suggested the potential for spectrum bias, although it was not clear if the control data was used in the estimates of diagnostic accuracy. For a cut point value of 39.7 pg/mL, the positive predictive value was 95.5 percent. For the diagnosis of HF regardless of LVEF, the AUC was 0.92 (95 percent CI: 0.86 to 0.99; p < 0.0001). The accuracy of BNP for the detection of systolic dysfunction was slightly less with an AUC of 0.78 (95 percent CI: 0.69 to 0.88; p < 0.0001). The BNP performance for detection of diastolic dysfunction expressed by the AUC was 0.89 (95 percent CI: 0.78 to1.00; p < 0.0001) for the patients without systolic dysfunction. Multiple regression analysis demonstrated that age, left ventricular mass index, and LVEF were independently associated with BNP.

Seino et al.⁵⁸ compared BNP and NT-proBNP relative to LVEF less than 40 percent and less than 50 percent in patients with HF. Their data indicate that detection of LVEF less than 50 percent was slightly greater for NT-proBNP than BNP (AUC 0.820 and 0.794, respectively). The reverse was true for LVEF less than 40 percent, with BNP having slightly greater AUC (0.770) compared to NT-proBNP (0.754). The optimum cut point values were determined to be 135 pg/mL for BNP and 695 pg/mL for NT-proBNP. There were four papers 11,59,61,62 that examined the relationship between BNP only and other HF variables but did not provide any data about the sensitivity, specificity or accuracy of these measurements. In general, it was demonstrated that BNP was related to cardiac function measured either as LVEF⁶² or left ventricular end diastolic pressure.⁶¹ One study⁵⁹ examined 41 HF patients and found BNP was related to the NYHA class.

There was only one study⁵⁸ that contained sufficient information to conduct meta-analysis by clinic setting alone. Therefore, no overall estimates for the individual clinic setting are possible. However, this one study was used to conduct meta-analysis for all sites combined (Figure 3 and 4). The results of this analysis are described elsewhere.

Quality assessment of studies. Results from the application of the QUADAS Question 14 quality assessment tool are as follows (see also Appendix C Evidence Figures - Figure 2): three (50.0 percent) of the studies clearly addressed the issue of spectrum bias(QUADAS Question 1); the selection criteria were only described in four (66.7 percent) of the studies (QUADAS Question 2) and in remaining two (33.3 percent) of the studies it was difficult to assess the selection criteria; four (66.7 percent) of the studies described if the reference standard was likely to correctly classify the HF (QUADAS Question 3); three (50 percent) of the studies clearly described the issue of disease progression (QUADAS Question 4); the reference standard was described in sufficient detail (QUADAS Question 9) in five (83.3 percent) of the studies; interpretation of the peptide marker (BNP or NT-proBNP) measurement was clearly made without knowledge of the reference test results (QUADAS Question 10) in five (83.3 percent) of the studies; interpretation of the reference test results was clearly made without knowledge of the B-type natriuretic peptide marker results (QUADAS Question 11) in all of the studies; five (83.3 percent) of the studies stated whether or not the clinical data was available when the B-type natriuretic peptide test results were interpreted as would be the case when the test is used in practice (QUADAS Question 12); one (16.7 percent) of the studies reported uninterpretable or intermediate test results (QUADAS Question 13); and withdrawals were explained in five (83.3 percent) of the studies (QUADAS Question 14). Overall, the quality of these studies was good.

Sample and Design Characteristics of Studies

There were seven papers eligible for review that selected patients from a primary care setting. Five of these studies were cross-sectional in design^{34, 65– 68} and one was a RCT⁶⁴. There was one study that selected patients randomly and identified a high risk cohort group.⁶³

Two of the studies restricted their recruitment by age; one to 40 years of age and above,⁶⁴ and one to more than 45 years of age.⁶³ One study presented data stratified by gender.⁶⁵ The RCT⁶⁴ examined the effect of BNP measurement on diagnostic accuracy in primary care. All patients had BNP measured, but the groups were randomized as to whether or not the primary care physician received the results. Nevertheless, this paper is useful because the BNP concentrations can be compared against the reference standard of HF (expert diagnosis) in both arms of the study.

Two studies^{34, 67} either did not provide estimates of the diagnostic performance of the BNP test, or presented the data in a manner such that these diagnostic characteristics could not be calculated.

Diagnosis of HF in studies. Four studies used evaluation of left ventricular systolic function by echocardiogram as the reference standard for HF.^{63, 65, 67, 68} Three used LVEF of less than or equal to 40 percent,^{63, 67, 68} one used LVEF less than or equal to 45 percent,⁶⁵ and one³⁴ did not state the reference cut point. Another study used x-ray or echocardiogram with evidence of pulmonary edema or cardiomegaly as the reference cut point.⁶⁶ The European Society of Cardiology criteria were used as the reference standard for the RCT study.⁶⁴

Diagnostic properties. Table 6 presents the results to answer the question, “What are the clinical performance characteristics of both BNP and NT-proBNP measurement in patients with symptoms suggestive of HF or with known HF presenting to a primary care physician?” Three papers evaluated only BNP,^{34, 66, 67} three evaluated only NT-proBNP,^{64, 65, 68} and one evaluated both.⁶³ Two of seven failed to indicate the cut point for BNP or NT-proBNP used. Two studies presented data for more than one cut point.^{65, 66} The range in cut points were 10 to 115 pg/mL for BNP and 67 to 338 pg/mL for NT-proBNP. Where possible, sensitivities, specificities, and LRs, either reported or calculated, are presented. Area under the ROC curve is presented when reported. Sensitivity ranged from 66 to 92 percent for BNP and 80 to 100 percent for NT-proBNP. Specificity ranged from 18 to 88 percent for BNP and 18 to 84 percent for NT-proBNP. For BNP, LR+ ranged from 1.12 to 6.71 and LR- ranged from 0.022 to 0.75. For NT-proBNP the LR+ ranged from 1.22 to 5.7 and the LR- ranged from 0 to 0.27.

Table 6

Diagnostic properties of studies that evaluated BNP and NT-proBNP in patients with symptoms suggestive of HF in primary care settings.

Meta-analysis was done on two studies^{63, 66} for BNP and three studies for NT-proBNP.^{63– 65} To maximize sensitivity the lowest cut point was used if data for multiple cut points were given. The data for the meta-analysis are summarized in Table 7 and the results presented in Figures 3 and 4. Since there were few studies available, the accompanying pooled summary statistics must be interpreted with caution. However, looking at the three studies for NT-proBNP the DOR summary estimate was 17 (95 percent CI: 9 to 32) whereas it was only 2 (95 percent CI: 1 to 6) for BNP. Furthermore, in the Hobbs study⁶³where both BNP and NT-proBNP were measured, the DOR was about eight times higher (2 and 17 for BNP and NT-proBNP, respectively). Tests for heterogeneity were not significant for either of the B-type natriuretic peptides for the pooled LR- or DOR.

Quality assessment of studies. Results from the application of the QUADAS Question 14 quality assessment tool are as follows (see also Appendix C Evidence Figures - Figure 3): six (85.7 percent) of the studies clearly addressed the issue of spectrum bias(QUADAS Question 1); the selection criteria were only described in six (85.7 percent) of the studies (QUADAS Question 2); all of the studies described if the reference standard was likely to classify the HF properly (QUADAS Question 3); five (71.4 percent) of the studies clearly described the issue of disease progression (QUADAS Question 4); the reference standard was described in sufficient detail (QUADAS Question 9) in six (85.7 percent) of the studies; interpretation of the peptide marker (BNP or NT-proBNP) measurement was clearly without knowledge of the reference test results (QUADAS Question 10) in all of the studies; interpretation of the reference test results was clearly without knowledge of the B-type natriuretic peptide marker results (QUADAS Question 11) in all of the studies; six (85.7 percent) of the studies stated whether or not the clinical data was available when the B-type natriuretic peptide test results were interpreted as would be the case when the test is used in practice (QUADAS Question 12); five (71.4 percent) studies reported uninterpretable or intermediate test results (QUADAS Question 13) and; withdrawals were explained in five (71.4 percent) studies (QUADAS Question 14). Overall, the quality of these studies was good.

Meta-Analysis

We chose studies for meta-analysis from Questions 2ai, 2aii and 2aiii where sufficient information was presented to allow calculation of sensitivity, specificity, LRs and DOR for as many diagnostic cut points as were presented (Table 7). Using this information, we developed summary estimates of these parameters (Figures 3, 4 and 5) as well as summary receiver operator characteristic (SROC) curves (Figure 6). In the pooling these data, we observed significant heterogeneity. As a result, we tried to explore the sources of the heterogeneity using meta-regressions and stratifications. We evaluated potential sources of heterogeneity for B-type natriuretics by stratifying groups according to the following factors: a) study setting (clinic, emergency department, and primary care), b) study design (cross-sectional, prospective cohort, randomized trials, and diagnostic types), c) sample size (greater than or equal to 500 and less than 500), d) comparison to reference standard (LVEF, compared to other signs and symptoms, and HF defined by clinical criteria), and e) cut points (exactly 100 pg/mL, greater than 100 pg/mL, and less than 100 pg/mL). Across the 5 different metrics of diagnostic accuracy (sensitivity, specificity, LR+, LR-, DOR), many of these were observed to be positive for heterogeneity, suggesting that no single factor helped to explain the variation between studies (Appendix C, Tables 13–26 detail the results of the heterogeneity tests by factors). The small number of studies within each of the various categories was also a limiting factor in exploring the relative contribution of these covariates to the observed heterogeneity.

Figure

Figure 6. Summary ROC curves for a) BNP and b) NT-proBNP from all settings using the lowest cut point provided in each study.

We also used the Moses-Littenberg regression model to develop a summary ROC curve and test for the presence of a threshold effect. Using both weighted and unweighted regressions, the slope parameter was small and not statistically significant (BNP p = 0.4183, NT-proBNP p = 0.3430), thus indicating the lack of a threshold effect. These data show that despite the various cut points and patient cohorts studied there was fairly high concordance among the studies.

Figure 6 presents the summary ROC curves for BNP and NT-proBNP. In both cases the curve tends strongly towards the upper left hand corner. The cut points ranged from 10 to 200 pg/mL (mean = 95 pg/mL) for BNP and 125 to 1691 pg/mL (mean = 642 pg/mL) for NT-proBNP. Sensitivities for BNP and NT-proBNP ranged from 50 to 99 percent and 83 to 99 percent, respectively. Specificities for BNP and NT-proBNP ranged from 19 to 97 percent and 46 to 89 percent, respectively. The areas under the curves, however, are 0.86 for both BNP and NT-proBNP, suggesting that regardless of the clinical setting, the cut point chosen, or the test used, measurement of B-type natriuretic peptides are useful in the diagnosis of HF. The standard error (SE) for the BNP AUC was slightly higher than for the NT-proBNP AUC (0.068 compared to 0.034, respectively). There are two noted outliers in the BNP SROC with respect to sensitivity that can account for this. One is from the clinic setting⁵⁸ and the other is from the primary care setting.⁶³ Although there were two studies from the primary care setting in the SROC, the study that appeared as an outlier selected patients who were at high risk for HF (prevalence = 7.5 percent) compared to the other study which selected patients suspected of having HF (prevalence = 32 percent).⁶⁶

Design and Sample Characteristics of Studies

The prognostic value of BNP or NT-proBNP was examined in 12 studies of people with risk factors for CAD, ten were prospective cohorts,^{9, 10, 15, 24, 69– 74} one was a RCT,⁴ and one was cross-sectional.⁶ Sample sizes ranged from 111⁹ to 3346.⁷⁴ Most study participants were between 50 and 75 years of age, although eight studies included patients over 50 years.^{6, 10, 15, 24, 69, 71, 73, 74} Four studies included participants under 75 years of age.^{9, 10, 24, 71} The widest age range, (19 to 105 years) was from an Irish study⁷¹ in an emergency room. In one study,⁷⁰ only the mean age of participants was reported so it was not possible to assess the age range of the sample. The percentage of males in the studies ranged from 21 percent⁹ to 96 percent.⁶⁹ Follow up averaged 8 to 9 years in two studies^{15, 70} and 8 to 12 days in another.⁷¹ Excluding the cross-sectional study,⁶ and another study for which the follow up time was reported as “until discharge,”⁷² follow up time ranged from approximately 1 to 5 years.^{4, 9, 10, 24, 69, 73, 74}

CAD Risk Factors

Study participants were recruited in an emergency room or when admitted to a cardiac care unit. CAD risk factors included diabetes,^{15, 69, 70, 74} suspicion of cardiac dysfunction,⁶⁹ ACS or chest pain,²⁴ suspected heart disease,^{9, 71, 74} cardiac arrest with cardiac cause,⁷² hypertension,^{4, 10, 74} prior MI,⁷⁴ aortic stenosis or aortic valve replacement,⁶ left atrial enlargement or left ventricular hypertrophy,⁷⁴ significant heart disease upon admission, or a cardiac event within 90 days of admission.⁷³

CAD risk factors were assessed using a mixture of electrocardiography, chest x-ray, clinical examination, LVEF, or NYHA classification criteria. Diabetes was assessed by clinical criteria such as persistent macroalbuminuria (> 300 mg/24 hours) in at least two out of three consecutive 24-hour urine collections, with the presence of diabetic retinopathy and the absence of other kidney or urinary tract disease. Hypertension was defined as systolic blood pressure greater than or equal to 140 mmHg or diastolic blood pressure greater than or equal to 90 mmHg.

BNP and NT-proBNP Tests and Threshold Values

BNP was measured using the Triage BNP method in two studies,^{69, 71} and the Shionoria-IRMA method in four studies.^{9, 10, 72, 74} NT-proBNP was measured using the Elecsys method in six studies.^{4, 6, 15, 24, 70, 73}

BNP or NT-proBNP cut point values were not uniform; six studies^{6, 9, 24, 69, 71, 73} reported multiple and six^{6, 24, 70, 71, 73, 74} reported single cut points. Cut points were chosen based on median or percentile levels of fasting plasma NT-proBNP^{4, 70} or plasma BNP,^{9, 74} mean BNP,¹⁰ ROC curves,^{6, 72, 73} information from the test package insert,⁷¹ or miscellaneous external sources.¹⁵ In two studies, the selection of cut points was arbitrary or unexplained.^{24, 69}

Definition of Outcomes

Death was a primary outcome in 10 studies,^{4, 9, 10, 15, 24, 69– 71, 73, 74} although in four ^{4, 10, 70, 73} it was part of a composite outcome including other cardiovascular events (e.g., non-fatal MI). Death was limited to cardiovascular events in four studies^{4, 10, 70, 73} and included all-cause mortality in six studies.^{9, 15, 24, 69, 71, 74} Deaths were ascertained using public records (e.g., death certificates) in four studies,^{15, 24, 69, 73} while in three there was mention of “clinical assessment” of cardiovascular outcomes (including death).^{4, 10, 74} Assessment of death was not described in three studies.^{9, 24, 71}

The two studies with non-death outcomes were the only studies that enrolled patients with prior surgeries. In the first study,⁷² 27 to 40 percent of patients had an intra-aortic balloon counterpulsation or a coronary revascularization. The outcomes ranged from survival to hospital discharge. In the second study,⁶ 12 percent of patients had coronary artery bypass graft (CABG) and 9 percent had percutaneous coronary intervention. The outcome was the severity of aortic stenosis. Outcomes other than death were measured by clinical tests such as the mean transvalvular pressure gradient⁶ or LVEF.⁷³

Four studies had secondary outcomes, including non-cardiac causes of death⁶⁹ and cardiovascular mortality plus hospitalization for HF.⁷⁰ One study⁷⁴ had three secondary outcomes: MI, heart disease, and atrial fibrillation. The fourth study⁷² had four secondary outcomes: return of spontaneous circulation, hospital admission, 24-hour survival, and favorable neurologic outcome after discharge.

Adjusted Results — Multiple Regression Analysis

Eleven of the 12 studies (Neilson et al.⁷³ excepted) featured regression analysis to examine the association between levels of BNP or NT-proBNP and the outcome of interest. In two of the 11 studies, one BNP⁶⁹ and one NT-proBNP,⁶ the reported regression results consisted only of p-values or chi-square test statistics. See Tables 11 and 12 for summary results for all 12 studies.

Table 11

Summary of studies in patients with risk of CAD: BNP.

Table 12

Summary of studies in patients with risk of CAD: NT-proBNP.

BNP was treated as a categorical variable in four studies,^{9, 71, 72, 74} with the categories based on the cut points discussed above. Higher levels of BNP were consistently found to be positively associated with all-cause mortality or the occurrence of cardiac events (e.g., HF).^{9, 71, 74} The adjusted measures of association (OR or hazard ratio (HR)) ranged from 1.10 to 4.26 and did not appear to differ by outcome in two studies (point estimates < 2.00 in both studies).^{9, 74} In one mortality study,⁷¹ though, the estimated OR for BNP (≥ 700 pg/mL versus < 700 pg/mL) was found to be much larger at 4.26.

In the final study where BNP was treated as categorical,⁷² the outcome was survival to hospital discharge. For the study participants, some of whom had prior surgery, all of the adjusted, estimated ORs were less than 1.00. Since the ORs decreased as the level of BNP increased, higher levels of BNP were negatively associated with survival.

The results of the study¹⁰ where BNP was treated as a continuous variable showed that higher values of BNP were positively associated with the occurrence of cardiovascular events (adjusted risk ratio (RR) = 1.02; 95 percent CI: 1.01 to 1.02).

NT-proBNP was also treated as a categorical variable in four studies.^{4, 15, 24, 70} Again, the categories were based on the cut points discussed above. Higher levels of NT-proBNP were consistently found to be positively associated with composite endpoints that included both cardiovascular mortality and other cardiac events such as non-fatal MI.^{4, 70} Positive associations were also found between levels of NT-proBNP and all-cause mortality.^{15, 24} The adjusted measures of association (RR or HR) ranged from 1.85 to 5.40, with a concentration between 2.8 and 3.6.

In 10 of the studies, all of the adjusted measures of association (i.e., OR, HR, RR) were statistically significant at the 5 percent level. In the other two studies,^{9, 24} all of the measures except for three were also statistically significant at the 5 percent level.

Several variables were included in many of the regression models as covariates, including age, gender, blood pressure, cholesterol level, left ventricular mass or function, and smoking.

Quality Assessment of Studies

Selection and information bias are two common threats to the internal validity of observational studies. One method of minimizing selection bias is to evaluate all patients who present at the research site between a certain set of fixed dates. Patients who meet the inclusion criteria are then entered into the study. This method of enrollment — often called ‘consecutive enrollment’ — helps to prevent a conscious or unconscious bias from affecting the selection of patients into the study. The bias would occur if, for example, patients with multiple risk factors for CAD, or with higher risk factors for CAD, were the only persons selected for the study. In the 12 studies of persons with risk factors for CAD, the authors of only four studies^{15, 24, 71, 73} explicitly wrote that patients were enrolled consecutively. The other eight studies contained no mention of consecutive enrollment. The authors of future studies should be explicit about the order of patient enrollment so as to allow readers to assess study quality.

Information bias occurs when study subjects are misclassified on their exposure or disease status. Misclassification can occur due to random chance (e.g., inaccurate measures of BNP or NT-proBNP), or because the persons who take study measurements have knowledge of a subject's exposure and disease status. For example, knowing that subjects have very high levels of BNP or NT-proBNP could trigger additional clinical investigations that lead to what would have otherwise been unmade diagnoses or treatments. This could inflate the association between BNP or NT-proBNP and the outcome of interest. Blinding is one way to avoid the problem. In the 12 studies involving persons at risk for CAD, the authors of five publications^{4, 6, 70, 73, 74} reported that blinding had occurred and the authors of seven studies did not mention anything about blinding. Again, authors should be explicit about what they do (blinding, no blinding) so as to facilitate the assessment of study quality.

The absence of information about consecutive enrollment or blinding makes the presence of bias impossible to rule out for a majority of the studies. The same absence of information prevents the extent of any bias from being ascertained.

Design and Sample Characteristics of Studies

Thirty-eight studies examined the association between BNP or NT-proBNP and outcomes such as mortality or re-infarction in persons with CAD. Twenty-eight of the studies were prospective cohorts,^{3, 13, 19– 22, 29, 33, 75– 94} nine were RCTs,^{8, 14, 27, 28, 95– 99} and one was cross-sectional.¹⁰⁰ Sample sizes ranged from a low of 14⁹⁴ to a high of 7800.^{8, 97} The mean age of study participants was clustered around 55 and 65 years, with a range of approximately 40 to 70 years. The widest age range spanned 54 years and was observed in two studies (21 to 75 years,²⁷ and 26 to 80 years⁸⁰). The percentage of males in 31 of the studies ranged from a low of 45 percent³ to a high of 100 percent.⁹¹ The percentage was not reported in five studies.^{20, 22, 78, 83, 94} Lengths of follow up varied greatly between the studies. The mean, median, or maximum follow up was up to and including 6 months in 12 studies,^{3, 21, 27, 28, 76, 86, 87, 90, 92, 93, 96, 99} 7 to 12 months in eight studies,^{8, 33, 77, 78, 83, 91, 95, 97} 13 to 24 months in eight studies,^{14, 20, 29, 75, 84, 85, 89, 94} and more than 24 months in eight studies.^{13, 22, 79– 82, 88, 98} Follow up time was not reported in two studies.^{19, 100} The shortest follow up time was 72 hours⁹⁹ and the longest was 4.9 years.⁸⁸

CAD Diagnosis

Study participants were recruited after admission to hospital for a CAD related event. CAD related events included MI (16 studies),^{3, 13, 27, 29, 33, 77, 78, 80, 82, 84, 85, 88, 91, 92, 95, 100} angina (10 studies),^{8, 20, 76, 77, 82, 85, 92, 95– 97} ischemia (five studies),^{22, 81, 90, 96, 100} chest pain (three studies),^{21, 77, 86} stenosis (three studies),^{19, 79, 85} LVD (one study),⁹⁸ ACSs (one study),³ cardiac arrest (one study),⁹⁹ and hypertension (one study).⁸⁹ More than one CAD event was involved in several studies.

CAD was diagnosed with a plethora of clinical tests such as electrocardiography, ST elevation, development of left bundle blockade, rises in creatinine kinase levels, T-wave inversion, and blood pressure. In some studies,^{14, 28, 83, 87} persons were enrolled on the basis of whether test results exceeded a certain threshold value (e.g., ST elevation ≥ 1 mm). In other studies, persons were enrolled if they were undergoing percutaneous transluminal coronary angioplasty (PTCA)⁷⁵ or CABG.^{91, 93, 94, 99}

BNP and NT-proBNP Tests and Threshold Values

BNP was measured using the Triage method in nine studies,^{3, 19, 28, 29, 77, 78, 83, 95, 96} the Shionoria-IRMA method in seven studies,^{13, 33, 79, 88, 90, 93, 94} and the ADVIA Centaur method in one study.²⁷ NT-proBNP was measured using either the Elecsys system in 13 studies^{8, 14, 21, 22, 75, 76, 85– 87, 92, 97, 99, 100} or by a variety of other methods in seven studies.^{20, 80– 82, 84, 89, 98}

BNP or NT-proBNP cut points were not uniform. Twenty-four of the studies^{3, 13, 19– 21, 27– 29, 75, 77– 79, 81– 83, 87– 90, 93, 95, 96, 98, 100} reported a single cut point and six^{8, 76, 86, 91, 92, 97} reported multiple cut points. Another six studies reported a single cut point, but the analyses were stratified by disease,^{33, 84, 85} gender,²² BNP versus NT-proBNP,⁸⁰ or time period during follow up.¹⁴ The cut points were based on the medians or quartiles of measured BNP or NT-proBNP in the study participants,^{8, 13, 14, 19, 22, 75, 76, 78, 80– 82, 85– 87, 89, 95, 97, 98, 100} ROC curves,^{3, 21, 27, 29, 33, 79, 84, 88, 91, 92} previously published literature,^{20, 28, 77, 83, 90, 96} or regression analysis⁹³Cut points were not reported in two studies.^{94, 99}

Definition of Outcomes

Primary outcomes were death in 32 studies,^{3, 8, 13, 14, 19– 22, 27– 29, 75– 77, 80– 90, 92, 94– 98, 100} non-fatal MI in 15 studies,^{8, 14, 21, 22, 75, 77, 80, 83, 85, 86, 92, 95– 98} HF or cardiogenic shock in 10 studies,^{28, 33, 78, 80, 83, 87, 89, 90, 92, 100} re-infarction in four studies,^{19, 28, 29, 100} repeat hospitalizations for ACS in five studies,^{28, 80, 91, 95, 96} angina in three studies,^{29, 79, 94} ischemia in two studies,^{33, 87} miscellaneous cardiovascular complications (e.g., arrhythmia, cardiogenic shock) in two studies,^{91, 92} BNP or NT-proBNP levels in two studies,^{93, 99} and LVD in one study.³³ Twenty-five studies^{8, 14, 19, 21, 22, 28, 29, 33, 75, 77, 80, 83, 85– 87, 89– 92, 94– 98, 100} included more than one outcome or had a composite outcome that was formed by aggregating two or more single outcomes. The outcomes were ascertained using clinical definitions (e.g., LVEF < 35 percent) in 20 studies.^{3, 8, 21, 29, 33, 75, 77– 79, 83, 84, 87, 89– 92, 97– 100} In the other 18 studies,^{13, 14, 19, 20, 22, 27, 28, 76, 80– 82, 85, 86, 88, 93– 96} the authors simply named the outcomes without specifying how they were assessed. This lack of specification was often the case with mortality: authors did not describe their method (e.g., review of death certificates) of determining whether and why a participant died.

Nine studies had secondary outcomes, including death,^{33, 88, 97} coronary HF or cardiogenic shock,⁷⁸ recurrent MI,⁹⁰ poor myocardial perfusion or failed ST segment resolution,²⁷ recurrent ischemic events and severe HF,⁷⁶ stroke,⁹⁸ and thrombosis in MI.³

Adjusted Results — Multiple Regression Analysis

Thirty-three of the studies used regression analysis to examine the association between levels of BNP or NT-proBNP and the outcome of interest. Multiple regression was used in 31 of the studies^{3, 8, 13, 14, 19– 22, 27, 28, 33, 75– 78, 80– 90, 93, 96– 98, 100} and simple regression was used in two of the studies.^{79, 94} Logistic regression was the sole approach in 17 studies.^{3, 8, 19– 21, 27, 28, 33, 76– 78, 84, 86, 87, 90, 96, 100} Logistic regression was also employed with Cox regression in two studies^{75, 97} and with linear regression in two studies.^{14, 89} Cox regression was utilized alone in ten studies.^{13, 22, 79– 83, 85, 88, 98} Regression analysis was not used in five studies.^{29, 91, 92, 95, 99} The studies were stratified according to type of B-type natriuretic peptide (BNP or NT-proBNP). Further stratification was done according to whether patients received prior cardiac related surgery (yes/no) and whether the outcome was mortality or non-fatal event (e.g., MI). For BNP, measures of association (OR, HR, etc.) ranged from 1.60 to 16.30 in studies of prior surgery patients and mortality^{3, 28, 77, 96} (Table 13). The measures of association were concentrated in the range of 1.60 to 2.96 and they were statistically significant at the 5 percent level in three of the four studies.^{3, 28, 77} In two studies of prior surgery patients and non-fatal outcomes, point estimates of the measures of association were 3.9²⁸ and 41.12⁷⁹ (Table 14). Both estimates were statistically significant at the 5 percent level.

Table 13

Summary of studies in patients with CAD with surgery: BNP.

Table 14

Summary of studies in patients with CAD, no surgery: BNP.

The predictive ability of BNP was found to be 2.53 (HR)¹³ or 7.20 (OR)²⁷ in two mortality studies of patients with no prior cardiac related surgery. Three studies^{33, 78, 83} were conducted to examine non-fatal outcomes in patients with no prior surgery and the measures of association ranged from 1.01 to 3.03 (Table 14). The measures of association in all five studies were statistically significant at the 5 percent level.

Turning to NT-proBNP, the measures of association spanned from 1.33 to 6.63 in six mortality studies of patients with prior cardiac related surgeries, with most measures concentrated in the range of 1.33 to 3.42.^{8, 21, 75, 76, 80, 82} All of the measures were statistically significant at the 5 percent level in five studies. In the sixth study,⁷⁶ only one of three ORs was significant at 5 percent. Studies^{8, 21, 80} of non-fatal outcomes in patients with prior surgery yielded measures of association that ranged from 1.01 to 3.51. However, the measures were statistically significant at the 5 percent level in only one study (Table 15).⁸⁰

Table 15

Summary of studies in patients with CAD not surgery: NT-proBNP.

In studies of NT-proBNP as a predictor of mortality in persons who did not have a prior cardiac related surgery, the measures of association ranged from 1.01 to 19.70, with a concentration in the range of 1.50 to 4.80. The measures were all statistically significant at the 5 percent level in four studies,^{20, 22, 81, 84} two of the three measures were significant in one study,⁹⁷ and none were significant in two studies.^{87, 98} In the case of non-fatal outcomes in persons who did not have a prior surgery, measures of association ranged from 0.64 to 5.90 and the concentration was between 1.30 and 5.50. The measures were all significant in two studies,^{84, 86} three measures of 13 were significant in three studies,^{85, 87, 97} and none were significant in two studies (Table 16).^{22, 98}

Table 16

Summary of studies in patients with CAD no surgery: NT-proBNP.

In one⁸⁹ of the two studies where linear regression was used as an analytic tool, NT-proBNP was found to be a predictor (p = 0.03) of left ventricular systolic volume after MI. However, the authors did not quantify the relationship by providing an estimated regression coefficient. In the other study, several variables were found to predict baseline levels of, and rates of change in, NT-proBNP.¹⁴ These variables were age, gender, diabetes, previous MI, cTnT level greater than or equal to 0.01 μg/L, calculated creatinine clearance less than 73 ml/min, C-reactive protein greater than 10 mg/l, ST-segment depression, and use of diuretics or nitrates on admission to hospital (Table 17).

Table 17

Summary of studies in patients with CAD no regression analyses.

Table 17 shows the studies for which no regression results were reported (even though the authors claimed to have conducted regression analyses) as well as the studies for which no regression analyses were performed.

Several variables were included in many of the regression models as covariates, including age, gender, biological markers (e.g., heart rate, blood pressure, baseline creatine kinase, LVEF), and disease history (e.g., history of MI, hypertension, diabetes).

Quality Assessment of Studies

The authors of 11 studies^{3, 20, 21, 33, 75, 77, 79, 82, 84, 91, 92} indicated that participants were consecutively enrolled in their research. The authors of the remaining 28 studies did not report whether or not enrolment was consecutive. For blinding, reporting was only slightly better as the authors of 14 studies^{3, 8, 13, 19, 21, 28, 29, 78, 81, 87, 92, 95, 97, 99} reported that outcomes were assessed in a blinded fashion, while the authors of the remaining 25 studies did not mention blinding in their published manuscripts. The lack of reporting in both areas meant that it was impossible to rule out the presence of selection or information bias in a majority of the studies.

Prognosis Studies Using BNP Levels

Design and Sample Characteristics of Studies. All the studies evaluating BNP levels were prospective cohort designs with the exception of three publications based on the ValHeFT^{106, 110} including a sub-study⁴¹which were randomized trials. The evaluation of the VaLHeFT cohort included both arms of the trial in the evaluations. In addition to the ValHeFT publications, an additional two publications reported on the same study cohort^{125, 128} with different follow up periods. Twelve studies^{12, 23, 104, 105, 107, 111, 113, 118– 120, 124, 129} recruited patients that were admitted as hospital inpatients for acute episodes; three studies contained patients recruited from both emergency and inpatients,¹⁰² and outpatient clinics and inpatients combined.^{25, 117} Twenty studies^{30, 36, 41, 101, 103, 106, 108– 110, 112, 115, 116, 121– 123, 125– 128, 130} indicated that patients were recruited from primary care, general population, specialty clinics and outpatient settings (see Tables 18 and 19). Three studies recruited patients with HF from emergency departments.^{32, 48, 114}

Table 18

Summary of studies in patients with HF and mortality outcomes: BNP.

Table 19

Summary of studies in patients with HF and mixed outcomes: BNP.

Sample sizes ranged from a low of 33³² subjects to a high of 4300 subjects.^{106, 110} The mean sample size for all 36 studies was 327 (± 829) and the median was 102. The mean age of study participants was clustered around 65 years, with the widest range of 68 years¹¹³ ranging from 17 to 85 years. The percentage of males in 32 of the studies ranged from a low of 44 percent¹¹¹ to a high of 100 percent.^{48, 119} The percentage was not reported in four studies.^{101, 106, 110, 122} One cohort¹⁰² based on the REDHOT study had a high proportion of African Americans (63.4 percent).

Lengths of follow up varied greatly between the studies, with the shortest time being 30 days¹¹⁹ and the longest being 92 months.¹²⁶ For studies with short term follow up, the period varied from 30 days¹¹⁹ to 90 days^{32, 48, 102} and six months.^{107, 109, 111, 114, 124} For the remaining studies the mean or median follow up was 7 to 12 months,^{118, 130} 13 to 24 months,^{12, 23, 30, 101, 103, 105, 115, 116, 125, 129} 25 to 48 months,^{25, 36, 41, 104, 113, 117, 120, 121, 123} and greater than 48 months.^{126, 127} Several other studies did not report median or mean follow up times but rather end points of 12 months,¹¹² 18 months,¹⁰⁸ 36 months,^{110, 128} and 60 months¹²² or did not specify.¹⁰⁶

HF Diagnosis and Severity. The diagnosis of HF was established in a number of ways, but predominately confirmed using echocardiography, carried out as part of the study, or obtained from previous medical records. The exceptions were three studies that used ventriculography^{41, 121, 187} and eight that used clinical evaluation (signs and symptoms).^{102, 103, 105, 108, 114, 122, 126, 188} All but two studies^{48, 107} reported some rating of the NYHA classification for all the subjects or a subgroup.

The majority of studies included subjects across all levels of the NYHA classification I-IV. The exceptions were eight studies^{23, 32, 108, 109, 119, 120, 123, 124} that restricted subjects to levels III-IV and one that possibly restricted to level IV.¹⁰⁷ A single study¹²¹ enrolled subjects with level I and II; three other studies^{36, 127, 130} had predominately level I and II with less than 10 percent of subjects with level III, and none at level IV. Three studies^{48, 114, 122} did not specify the NYHA classification of their subjects.

LVEF was not reported in seven studies^{102, 104, 105, 111, 114, 122, 127}. Mean LVEF percentages were reported in 22 studies and varied between 50 percent,¹¹⁵ 40 to 49 percent,^{32, 129} 30 to 39 percent,^{12, 23, 36, 107, 113, 118, 119, 123, 123, 126, 130} and 20 to 29 percent.^{25, 108, 109, 112, 117, 120, 125, 128} Seven studies reported a threshold LVEF of less than 45 percent,^{48, 101, 106, 110, 121} less than 40 percent,¹²⁴ and less than 30 percent.³⁰

BNP and Cut Points. BNP was measured using the Triage method in 11 studies^{48, 107– 109, 111, 112, 114, 119, 125, 128, 189} and by Shionoria-IRMA method in the remaining 25 studies (see Tables 18 and 19). Four studies^{112, 125, 126, 128} also measured NT-proBNP using the Biomedica method. Results are described in the NT-proBNP section of this report.

The rationale for selecting the BNP cut points differed and as such, the values varied significantly (Table 18 and 19). Sixteen studies^{12, 30, 32, 36, 101, 106, 110, 117– 121, 123, 124, 126, 130} selected the mean or median values as the cut point for categorizing high and low BNP groups. Eight studies^{23, 108, 114, 120, 125, 128, 129, 190} selected values based on ROC. Five studies categorized the sample BNP values into two levels,¹²⁷ three levels,^{107, 109} four levels,¹¹⁹ and five levels.¹²² Six studies^{104, 105, 111– 113, 115} did not specify a threshold as they used the BNP values as a continuous variable in their statistical analyses. Three studies used other rationale for threshold selection including: previous literature reference,⁴⁸ 75th percentile value,¹¹⁶ and an unspecified internal analysis.¹⁰²

Definition of Outcomes. All studies with the exception of one,⁴⁸ had a primary outcome of mortality or a composite endpoint. These endpoints typically included death, other cardiac events, readmission or worsening HF. There were 21 studies^{12, 23, 30, 32, 36, 102, 104, 106, 109, 112, 114, 117, 119– 123, 125– 127, 130} that evaluated either all-cause mortality, cardiac related mortality or both. There were 25 studies^{23, 25, 101– 103, 105– 108, 110– 120, 123, 124, 126, 128, 129} that evaluated composite endpoints that included a mixture of fatal and non-fatal cardiac events. One study¹¹⁶ evaluated the performance of BNP relative to the Heart Failure Survival Score. Most authors did not specify how the outcomes of death were verified or subsequent events (such as other events or re-admission to hospital) were evaluated.

Adjusted Results — Multiple Regression Analysis. Twenty-nine studies undertook Cox proportional hazards regression analyses and five studies undertook logistic regression analyses to evaluate the relationship between BNP levels and various outcomes (Table 18 and 19). For the studies using Cox regression analyses, three studies^{32, 111, 115} presented only univariate analyses, 10 studies^{101, 105, 110, 112, 113, 122– 124, 126, 127} presented only the results from multivariate analyses, and the remaining 16 studies^{12, 23, 30, 36, 104, 106– 109, 117, 118, 120, 121, 125, 129, 130} undertook both univariate and multivariate computations. Four studies undertook multivariate logistic regression^{25, 48, 102, 116} and one study¹¹⁹ univariate logistic regression, to evaluate the strength of the association between levels of BNP and the outcomes of interest. A single study¹¹⁴ reported unadjusted RRs and another study¹²¹ undertook an unspecified type of linear regression.

BNP Studies with Mortality Outcomes. Table 18 details the 21 studies with the outcomes of all-cause mortality, cardiac mortality and sudden death. The results are expressed as both univariate and multivariate HR,^{12, 23, 30, 36, 104, 106, 109, 113, 117, 122, 123, 127} Chi square statistic and probabilities values,^{12, 32, 102, 120, 125, 126, 130} beta values,^{102, 130} and unadjusted RR.¹¹⁴ One study measured, but did not report, findings specific to mortality.¹¹⁹ Some studies^{102, 123, 125}reported the estimates of risk based on the log of the BNP levels, which makes interpretation of the magnitude somewhat limited. In general, there were 11 studies^{12, 23, 30, 36, 106, 109, 113, 123, 125, 127, 130} that found baseline BNP levels to be significant predictors for mortality related outcomes after adjustment in multivariate models. For those studies that presented adjusted HR, the risk estimates varied from 2.48 (95 percent CI: 2.13 to 2.88)¹⁰⁶ to 7.2 (95 percent CI: 1.6 to 32.1),³⁰ with the majority point estimates clustering around 2.5 to 3.0. It should be noted that despite the differing cut points, (> 97 pg/mL,¹⁰⁶ > 172 pg/mL,³⁰ > 230 pg/mL,¹¹⁴ > 260.4 pg/mL,³⁶ > 700 pg/mL,³² 1000 pg/mL,¹⁰⁹ > 346 pg/mL,¹²⁷), the multivariate models found baseline BNP to be a significant predictor of mortality outcomes. All these regression models used a variety of variables within their models, which makes comparisons across studies challenging.

Six of the studies evaluating mortality recruited decompensated HF patients,^{12, 23, 102, 112} emergency department patients,^{32, 114} or mixed patients.¹¹⁴ Five of these studies reported 46 to 100 percent of patients in NYHA class III and IV suggesting relatively severe HF patients. Although all studies reported that BNP was a significant predictor of mortality, only one study²³provided an estimate of the HR; the multivariate model included a variable of troponin T and baseline BNP. The log BNP had a HR = 3.11 (95 percent CI: 1.61 to 6.01, p = 0.0005).

One study reported unadjusted RR for HF death for baseline BNP greater than 230 pg/mL vs. less than 230 pg/mL: RR = 24.1 (95 percent CI: 6.3.5 to 491.1); for the outcome of cardiac death unadjusted baseline BNP greater than 230 pg/mL vs. less than 230 pg/mL: RR = 37.9 (95 percent CI: 5.7.5 to 755.8). The widely varying CI suggests instability with the estimate; therefore, results must be interpreted with caution.

There were six studies that found BNP levels to be not significant predictors of mortality based on univariate analyses alone,^{32, 126} or on multivariate analyses;^{30, 104, 117, 122} one study¹²³ found sudden unexpected death to be significant but not HF death. One of these studies³² had a small sample size (n = 33) and 73 percent of patients were NYHA level IV. No clear trend from these studies can account for the non-significance.

BNP Studies with Composite Outcomes. Table 19 details the 27 studies with composite outcomes which included death in all but four publications.^{111, 113, 118, 128} The results are expressed as both univariate and multivariate HR,^{23, 25, 106, 110, 115, 117} chi square and probabilities values,^{113, 119, 121, 124, 128, 129, 191} and unadjusted RR.¹¹⁴ Some studies^{23, 102, 118, 119, 123, 128, 129} reported the estimates of risk based on log BNP levels, which makes comparison with studies not using log values difficult.

In general, there were 9 studies^{12, 30, 36, 106, 109, 123, 125, 127, 130} that found baseline BNP levels to be significant predictors for composite outcomes after adjustment in multivariate models. Two studies^{105, 111} with only univariate analyses showed baseline BNP levels to be not significant predictors and three studies (multivariate estimates^{106, 117} and univariate estimate¹¹⁵) showed only marginal significance.

For those studies that presented adjusted risk estimates for baseline BNP, the values varied from HR = 1.66 (95 percent CI: 1.36 to 2.04, p less than 0.0001)¹¹⁰ to RR = 3.23 (95 percent CI: 1.32 to 7.94, p = 0.01)²⁵, with the majority point estimates clustering around 2.0. Three studies^{23, 108, 123} included estimates of HR that combined levels of troponin I, troponin T and LVEF with baseline BNP. One study¹¹⁴ reported unadjusted RR for HF events that varied from baseline BNP greater than 230 pg/mL vs. less than 230 pg/mL: RR = 15.5 (95 percent CI: 6.2 to 43.7) to BNP greater than 480 pg/mL vs. less than 230 pg/mL: RR = 8.2 (95 percent CI: 4.7 to 14.3).

Comparison of NT-proBNP and BNP. Six studies^{41, 103, 112, 125, 126, 192} evaluated both BNP and NT-proBNP levels within the same group of subjects. Van Beneden et al.¹²⁶ compared a small number of subjects who had mild to moderate HF (NYHA I-II) with severe HF patients (NYHA level IIII-IV). Their univariate analysis showed that NT-proBNP was a significant predictor of mortality. However, no association between BNP and mortality was observed using either assay method. Two studies evaluated both B-type natriuretic peptides in predominately NYHA level II patients (approximately 78 percent). One study¹⁰³ found log BNP levels significant in predicting worsening HF in both univariate and multivariate analyses while NT-proBNP was significant only in the univariate model. In contrast, a second study,⁴¹ with mild to moderate HF subjects, found that BNP and NT-proBNP levels differed in their ability to predict 4-year mortality with respect to whether baseline levels, or measurement taken at last follow up, were considered. In this study, BNP was significant in the univariate analysis for baseline and significant for last follow up in the multivariate analysis; NT-proBNP was not significant at baseline in the univariate analysis but significant for the multivariate analysis that included last follow up NT-proBNP level. The findings of this study⁴¹ would suggest that measures of either BNP or NT-proBNP may be independent predictors of mortality, but the sample size for this study was relatively small (n = 100).

Two other studies (based on three publications)^{112, 125, 128} from the same research team evaluated BNP and NT-proBNP. With respect to predicting event free survival at 1 year in an ambulatory group of patients, one study¹¹² found both BNP and NT-proBNP were significant predictors at 1 year in univariate analysis. The subsequent multivariate analysis found four of the nine variables to be significant and included BNP, N-ANP, RAAS antagonists, and Living with Heart Failure questionnaire. From the same research group, but in a different study cohort, Berger et al.,¹²⁵ using univariate analysis, found the log transformation of both B-type natriuretic peptides to be significant predictors of pump failure death; however, only the log of NT-proBNP was significant in multivariate analyses. A sub-analysis, which excluded patients with atrial fibrillation, showed that BNP was the best independent predictor of sudden death. Their findings suggest that the magnitude of the prognostic prediction is dependent on the specific mode of death and the specific form of the natriuretic peptide. A re-analysis of this same cohort of subjects was undertaken in a third study.¹²⁸ The sample was classified according to severity levels based on LVEF and NYHA classification; a multivariate analysis was undertaken to see which factors predicted combined death and urgent heart transplant (for each of the 3 years of follow up). For Group A (NYHA I-II, LVEF ≤ 20 percent) log BNP was significant in both year 2 and 3. For Group B (NYHA I-II, LVEF < 20 percent, or, NYHA III-IV, LVEF ≤ 20 percent) only the log BNP was a significant predictor for year 1, but only log NT-proBNP was significant for years 2 and 3. For Group C (NYHA III-IV, LVEF < 20 percent) both log BNP and log NT-proBNP were significant for year 1; only log NT-proBNP was a significant predictor for years 2 and 3. These results would suggest that the strength of prediction for the B-type natriuretic peptides is also dependent on the year of follow up and less so on the severity of the HF.

Overall, of the six studies evaluating both BNP and NT-proBNP, only two studies found both BNP and NT-proBNP to be independent predictors of mortality. In one study¹²⁸ year of follow up and group stratification, and in a second study⁴¹ the timing of the B-type natriuretic peptide measurement, similarly altered the predictive ability. No clear pattern emerges to suggest superiority of one type of B-type natiruetic peptide relative to the other in these head to head studies.

Comparison of Studies That Evaluated Baseline and Predischarge Measures.

All studies with the exception of seven publications^{36, 41, 107, 111, 118, 119, 124} evaluated only baseline BNP levels. One of these studies³⁶ evaluated outpatients and measured BNP levels at two time points with an interval of 8 to 12 months. Another study evaluated patients BNP and NT-proBNP levels at successive follow up intervals over a 4-year period.⁴¹ Three^{107, 119, 124} of these studies had severe HF patients (NYHA III-IV) and approximately LVEF less than 40 percent; a fourth study¹¹¹ had 88 percent of subjects at NYHA level III and IV (LVEF not reported). Logeart et al.,¹⁰⁷ which had the largest sample size of all these seven studies, found the univariate HR for each 100 pg/mL increase of BNP to be slightly larger for predischarge BNP levels (HR = 1.22, 95 percent CI: 1.15 to 1.30, p = 0.0001) than baseline BNP levels (HR = 1.06, 95 percent CI: 1.03 to 1.10, p = 0.0001) for the composite outcome (death and readmission). In addition, they demonstrated a gradient of increasing risk from the first quartile (0 to 130) to the last quartile (660 to 1725) with the latter having the largest HR risk estimate (HR = 13.77). Similarly, the adjusted HR in this study showed increasing risk for poor outcome with increasing predischarge ranges, with the highest threshold (> 700 pg/mL) having a HR = 15.2, (95 percent CI: 8.5 to 27). Hamada et al.¹²⁴ found predischarge BNP levels to be the only significant (p = 0.0086) predictor of re-hospitalization within 1 year in a multivariate analysis that included baseline admission BNP levels. Cheng et al.¹¹⁹ undertook only univariate logistic regression and found admission BNP and discharge BNP to both be significant for the composite endpoint and 30 day readmission; however, the small sample in this study size did not permit a true multivariate analysis. Given the likely strong correlation between admission and discharge BNP levels, it would be important to use multivariate analyses to adjust for strong correlations between these two measures of BNP. Similarly, Bettencourt et al.¹¹¹ in a univariate analysis did not find median levels of either admission BNP (> 541 pg/mL) or discharge BNP (> 321 pg/mL) to be significant predictors of their composite endpoint (death or readmission).

Cheng et al.¹¹⁹ and a second study¹¹⁸ specifically recruited for new onset or first episode of HF. However, Tamura et al.¹¹⁸ measured BNP at discharge only and found the discharge log BNP to have the largest risk (HR = 2.656, p = 0.015) relative to the other variables significant in the model (NYHA class, LVEF, and left ventricular mass index) for predicting cardiac events.

A single study³⁶ evaluated patients in an outpatient clinic (sample was 93 percent NYHA level I-II) and measured BNP levels at initial visit and second time-point (8 to12 months later), as well as the change in BNP levels (per 100 pg/mL) and any BNP level increase (versus decrease) during follow up. However, they did not include discharge BNP as a unique variable in subsequent univariate or multivariate models. In this study, the change in BNP (HR = 1.34, 95 percent CI: 1.10 to 1.63) and change in NYHA class (HR = 6.68, 95 percent CI: 2.23 to 19.12) were the only significant variables in the multivariate model.

Kaplan Meier Survival Analyses. Twenty-seven studies^{12, 23, 30, 32, 36, 41, 101, 103– 105, 108– 118, 120– 123, 129, 130} reported results from Kaplan Meier survival analyses using various cut points that were based primarily on median/mean or best values from ROC curves. All studies that undertook Kaplan Meier analysis, regardless of the outcome or cut point, found significant differences between the two groups.

Quality Assessment of Studies. For this research question, which evaluated prognosis, the eligible studies were based predominately on prospective cohort designs. As such, selection biases attributed to non-consecutive enrolment, and misclassification bias attributed to blinding, were chosen as the main criteria for methodological quality evaluation. Thirteen studies^{30, 104, 105, 107, 108, 112– 114, 118, 120, 121, 129, 130} selected patients in a consecutive manner. The remaining studies did not specify, and likely did not employ, this strategy to minimize bias as convenience samples were generally selected.

Attempts were made to evaluate the potential for misclassification bias through lack of blinding of clinicians who evaluated subjects or those who ascertained the endpoints. Blinding of the clinicians to the BNP level was undertaken in only four studies^{23, 36, 102, 115} and this minimized the potential for clinicians to systematically provide differential treatments or request additional tests. Blinding to NT-proBNP levels was not undertaken in any study; however, three studies^{113, 120, 193} indicated that the outcome was judged by researchers external to the clinical setting and had some potential to minimize ascertainment bias.

Prognosis Studies Using NT-proBNP Levels

Study Design and Sample Characteristics of Studies. For those studies evaluating NT-proBNP, four were RCTs^{41, 136, 138, 140} and the remaining 14 were prospective cohort studies.^{26, 35, 103, 112, 125, 126, 128, 131– 135, 137, 139} Six of these studies evaluated both NT-proBNP and BNP^{41, 103, 112, 125, 126, 128}. Two publications were based on the same COPERNICUS study^{35, 140} and an additional two publications reported on the same study cohort^{125, 128} with different follow up periods.

Only four studies^{26, 135, 136, 139} recruited patients that were admitted to hospital for acute episodes. The remaining 14 studies^{35, 41, 103, 112, 125, 126, 128, 131– 134, 137, 138, 140} indicated that patients were recruited from primary care or specialty clinics and outpatient settings (see Tables 20 and 21).

Table 20

Summary of studies in patients with HF and mortality outcomes: NT-proBNP.

Table 21

Summary of studies in patients with HF and mixed outcomes: NT-proBNP.

Sample sizes ranged from a low of 48¹³¹ to a high of 2320 subjects.¹³³ The mean sample size for all 18 studies was 378 (± 596) and the median was 121. The mean age of study participants clustered around 65 years but varied from 51 to 78 years; the widest age range spanned 64 years (40 to 104 years).¹³³

Lengths of follow up varied greatly between the studies and the shortest time was 6 months¹³⁹ while the longest time was 92 months¹²⁶. The mean or median follow up time varied from 7 to 12 months,^{135, 136} 13 to 24 months,^{103, 134} 25 to 48 months,^{35, 128, 131, 194} and greater than 48 months.¹²⁶ Several studies specified only endpoints of 6 months,¹³⁹ 12 months,^{112, 133, 137} 17/18 months,^{26, 138} 24 months,¹⁴⁰ and 48 months⁴¹ (Tables 19 and 20).

HF Diagnosis and Severity. The diagnosis of HF was established in a number of ways, and these included echocardiography (carried out as part of the study or obtained from previous medical records) in six studies,^{35, 112, 125, 132, 133, 140} ventriculography in four studies,^{41, 134, 137, 138} clinical evaluation (signs and symptoms or NYHA classification) in six studies,^{103, 126, 131, 135, 136, 139} and other methods in one study.²⁶ All studies used the NYHA classification, with the exception of one study¹³⁵ which used the Killip method, and three studies^{35, 133, 140} which did not report any rating.

With respect to severity of HF, the majority of studies included subjects across all levels of the NYHA classification. The exception to this was one study¹³⁹ that restricted subjects to levels III-IV. A single study¹¹² enrolled subjects with predominately level II. Three studies^{35, 133, 140} did not specify the NYHA classification of their subjects. A single study¹³⁵ used the Killup method of classification, and these patients varied from levels II to IV.

LVEF was not reported in four studies on admission.^{26, 135, 136, 139} Four studies^{35, 132, 137, 138} reported a threshold LVEF of less than 45 percent, five studies^{112, 125, 126, 128, 134} less than 35 percent, and five studies^{41, 103, 131, 133, 140} less than 25 percent.

NT-proBNP and Cut Points. NT-proBNP was measured in nine studies using the Elecsys method, six studies used the Biomedica method,^{41, 103, 112, 125, 126, 128} two used the Roche ELISA method, ^{135, 137} and one study used the Christchurch method¹³⁸ (Tables 20 and 21). Six of these studies evaluated both NT-proBNP and BNP.^{41, 103, 112, 125, 126, 128}

From the two publications based on the COPERNICUS study, one of the reports³⁵ stated that a newly developed NT-proBNP Roche ELISA method was used and that samples would subsequently be retested using the Elecsys method. The re-analysis was published the same year but does not reference the previous report. Since this is the same cohort, we assume that the assay met our eligibility criteria but was not adequately reported.

Definition of Outcomes. Ten of these reports examined mortality as a discrete end point.^{35, 41, 125, 126, 131, 133, 134, 136, 138, 140} Nine studies^{103, 112, 128, 134– 137, 139, 140} reported composite end point of death or worsening HF. The remaining studies evaluated the ability to predict recommendation of heart transplantation,¹³² worsening HF alone,¹³⁸ and in one case no estimate of risk was provided.²⁶

Adjusted results — Multiple Regression Analysis. One²⁶ study examined stratification with troponin level and presented no relative measure of risk such as a HR. However, it did make the limited statement that there were fewer events when NT-proBNP levels were below rather than above the median admission level (1357 pg/mL, p < 0.01).

Eleven studies undertook Cox proportional hazard regressions analyses and six studies undertook logistic regression analyses to evaluate the relationship between BNP levels and the relationship to the various outcomes. For the studies using Cox proportional hazard regression analyses, two studies^{112, 138} presented only multivariate estimates and the remaining studies undertook both univariate and multivariate computations. Six studies undertook multivariate logistic regression^{126, 132, 133, 135– 137} to evaluate the strength of the association between levels of BNP and the outcomes of interest. A single study²⁶ did not report estimates of risk.

NT-proBNP Studies With Mortality Outcomes. Eleven studies^{35, 41, 112, 125, 126, 131, 133, 134, 136, 138, 140} evaluated mortality outcomes and these are detailed in Table 20. The results are expressed as both univariate and multivariate HR,^{35, 131, 140} chi square statistic,^{41, 112, 126, 134} OR,^{133, 134, 136}or not specified.¹³⁸ Some studies^{41, 126, 131, 133} reported the estimates of risk based on the log NT-proBNP levels. In general, all studies that undertook univariate or both univariate and multivariate analyses found NT-proBNP to be a significant predictor of mortality. One study¹³⁵ found multivariate estimate on a subsample with HF to be significant (OR = 5.30, 95 percent CI: 1.4 to 168.9, p = 0.026) but the CI was wide. For those studies that presented adjusted HR, the risk estimates varied from HR = 2.17 (NT-proBNP > 1767 pg/mL) (95 percent CI: 1.33 to 3.54, p < 0.02)¹⁴⁰ to HR = 9.35 (log NT-proBNP) (95 percent CI: 2.42 to 36.10, p = 0.001). These estimates encompass baseline, discharge and change of NT-proBNP levels.

NT-proBNP Studies With Composite Outcomes. Table 21 details the ten studies with the composite end points, which included death as a component of the outcome. The results are expressed as both univariate and multivariate HR,^{139, 140} chi square statistic,^{103, 112, 128, 134, 137} OR,^{134– 136} or not specified.¹⁹⁵ One study¹²⁸ reported the estimates of risk based on the log NT-proBNP levels.

In general, there were seven studies^{112, 128, 134– 137, 140} that found baseline, discharge or change levels of NT-proBNP levels to be significant predictors for composite outcomes after adjustment in multivariate models. Two studies^{103, 139} showed marginal or no statistical significance. All these regression models used a variety of variables within their models, which makes comparisons across studies challenging.

For those studies that presented adjusted risk estimates for NT-proBNP, the values varied from HR = 2.11 (95 percent CI: 1.54 to 2.90, p < 0.0001) for NT-proBNP greater than 1767 pg/mL,¹⁴⁰ to HR = 5.96 (95 percent CI: 3.23 to 11.01) for change in NT-proBNP vs. decrease greater than 30 percent or increase greater than 30 percent.

Quality Assessment of Studies. As with the BNP studies, potential for selection bias (attributed to non-consecutive enrolment) and misclassification bias (attributed to blinding) were chosen as the main criteria for methodological quality evaluation. Eight studies^{26, 112, 132– 135, 137, 139} selected patients in a consecutive manner. The remaining studies did not specify and likely did not employ this strategy to minimize bias, as convenience samples were generally selected. Blinding of the clinicians or the investigators to the NT-proBNP levels with respect to the outcomes was undertaken in only three studies^{132, 136, 139} suggesting some potential for ascertainment bias for of the outcome.