Results of the search

The search identified 3394 references. Once duplicates were removed, the titles and abstracts of the remaining 3379 references were screened using our inclusion /exclusion criteria and 3044 removed as not relevant to the review. Full texts were examined for the remaining 335 references and from these 18 studies were included in this review (see Figure 1). Full details of all the studies are given in the Characteristics of included studies, Table 1, Table 2, Characteristics of excluded studies, and Characteristics of ongoing studies. Each study is identified by the name of the first author and year of publication of the main results paper (Study ID). Additional references are listed together with this main publication under the study ID.

Figure 1

---

---

Study flow diagram.

Figure 4

---

---

Funnel plot of comparison: NP‐guided versus no NP‐guided treatment for all‐cause mortality.

Included studies

The Characteristics of included studies, Table 1 and Table 2 provide details of each of the 18 included studies.

The earliest study was published in 2000 (Troughton 2000) and the latest in 2015 (Skvortsov 2015). For two of the studies, data were only available through conference abstracts and direct contact with the authors (Krupicka 2010; Shochat 2012).

Ten of the studies were completed in Europe (two in Sweden/Norway (Karlstrom 2011; Persson 2010), two in Switzerland/Germany (Maeder 2013; Pfisterer 2009), one in Austria (Berger 2010), France (Jourdain 2007), the Netherlands (Eurlings 2010), Spain (Anguita 2010), Denmark (Schou 2013). and the Czech Republic (Krupicka 2010)); three studies were completed in North America (two in the USA (Januzzi 2011; Shah 2011) and one in Canada (Beck‐da‐Silva 2005)); two were completed in New Zealand (Lainchbury 2010; Troughton 2000), one in Israel (Shochat 2012), one in Russia (Skvortsov 2015), and one in China (Li 2015).

Two of the 18 studies (Berger 2010; Lainchbury 2010) had three comparison arms comparing NP‐guided treatment both to clinical assessment and to usual care. For usual care there were no scheduled visits and the participants were managed in primary care. Studies recruited 3660 participants ranging from 41 to 499 participants per study. The average age of participants in all the studies ranged from 62 to 80 years old. Studies followed up participants from baseline to between one and 54 months.

Seven studies (Anguita 2010; Beck‐da‐Silva 2005; Jourdain 2007; Karlstrom 2011; Krupicka 2010; Li 2015; Shah 2011) used BNP as the biomarker; the remainder used NT‐proBNP. Only seven studies (Eurlings 2010; Maeder 2013; Persson 2010; Pfisterer 2009; Schou 2013; Shochat 2012; Skvortsov 2015) stated an NP level as an inclusion criterion. All studies set a NP target except for Beck‐da‐Silva 2005; Schou 2013 and Shochat 2012 who stated a change in NP level (See Table 2).

Two studies (Beck‐da‐Silva 2005; Li 2015), compared the effect of NP‐guided treatment with clinical assessment exclusively for the up‐titration of beta‐blockers. Beck‐da‐Silva 2005 changed the dose of bisoprolol, but all other drugs remained unchanged, during a three‐month follow‐up period. Li 2015 started and increased the dose of metoprolol succinate over one month; for these patients intravenous cardiotonic, vasodilator or diuretic was applied if signs or symptoms of heart failure were observed.

Beck‐da‐Silva 2005 was the only study to report an algorithm where medication (beta blocker) was decreased for patients whom the BNP measurement was increasing, but the clinical assessment was worse.

All, bar three studies (Eurlings 2010, Lainchbury 2010; Schou 2013), reported inclusion criteria for classifying participants according to the New York Heart Association (NYHA) functional classification. This classifies patients with heart disease into four stages based on limitations on physical activity, symptoms with ordinary physical activity and status at rest. Stage four indicating the highest severity of symptoms. At baseline, most studies grouped participants by NYHA stage and overall, the participants ranged between stages II and IV. Three studies reported baseline NYHA as percentages in each stage: for Eurlings 2010 and Lainchbury 2010, over 60% of participants were in class II and for Schou 2013 over 85% were in stages I to II.

Further classification was determined by percentage left ventricular ejection fraction (LVEF); 12 of the studies stated as an inclusion criterion a maximum level for percentage LVEF which ranged between < 35% to < 50%; five studies did not stipulate any inclusion level (Anguita 2010; Eurlings 2010; Lainchbury 2010; Li 2015; Shochat 2012); and Maeder 2013 was the only study to have participants solely with percentage > 45% LVEF or preserved LVEF. Although six of the studies did not stipulate an inclusion level percentage LVEF, Lainchbury 2010 was the only other study to state participants with preserved LVEF were not excluded. At baseline, Berger 2010 did not report LVEF percentage, Maeder 2013 reported all participants averaged 56% LVEF, Karlstrom 2011 reported 57% of participants were < 30% LVEF, whilst the remaining studies reported overall averages ranging from 20% to 46% LVEF.

Six studies (Felker 2014; Jourdain 2014; Metra 2012; Moe 2007; Saraya 2015; Steinen 2014) are classified as ongoing. Of these, four studies (Felker 2014; Jourdain 2014; Moe 2007; Steinen 2014) are currently recruiting or have just finished recruiting. Metra 2012 finished recruiting in August 2009 and is due to publish shortly. Saraya 2015 has been completed, but currently only published as a conference abstract. All six are listed in the Characteristics of ongoing studies.

Excluded studies

Thirty‐five references are included in the Characteristics of excluded studies tables where the title or abstract or both appeared to suggest a relevant study to this review. Of these 68% were excluded as the study was not a randomised control trial. Other reasons included not NP‐guided treatment (20%), trial terminated, not treatment for heart failure, or not a baseline heart failure population.

Allocation

All studies clearly stated the study was randomised, but not all studies reported on how randomisation was completed or if allocation concealment was achieved. Five studies confirmed sequence generation and allocation concealment and methods were judged to be at low risk of bias (Berger 2010; Karlstrom 2011; Maeder 2013; Pfisterer 2009; Shah 2011). Januzzi 2011; Lainchbury 2010; Shochat 2012; Skvortsov 2015 and Troughton 2000 were low risk for sequence generation only and Beck‐da‐Silva 2005; Eurlings 2010 and Krupicka 2010 only for allocation concealment. The remaining studies were classified as unclear.

Blinding

Blinding of participants and study personnel was only judged to be low risk if both were blinded to the treatment allocation; only one study met this standard (Lainchbury 2010). Five studies did not report or it was unclear whether participants or personnel were blinded to treatment allocation (Anguita 2010; Li 2015; Persson 2010; Shochat 2012; Skvortsov 2015). In all the remaining studies one or more of these groups were not blinded. Blinding of outcome assessments was not achieved or not reported in the majority of studies; only five studies blinded outcome assessment (Berger 2010; Eurlings 2010; Karlstrom 2011; Lainchbury 2010; Schou 2013).

Incomplete outcome data

For the primary outcome, all‐cause mortality, eight studies (Anguita 2010; Berger 2010; Jourdain 2007; Li 2015; Schou 2013; Shah 2011; Skvortsov 2015; Troughton 2000) were judged to be low risk with regard to incomplete outcome data, in fact they all had no attrition except for Skvortsov 2015 where the numbers and reasons were fully reported. The remaining studies either did not report attrition, or the studies did confirm attrition with break down by intervention arm, but did not explain how missing data were handled. For those studies reporting dropouts, the overall attrition rates were no more than 23%.

All of the studies, bar four, completed intention‐to‐treat (ITT) analyses; Beck‐da‐Silva 2005 did not complete an ITT analysis, whilst Anguita 2010; Jourdain 2007 and Li 2015 did not report whether this method was used.

Selective reporting

Nine out of 18 studies reported on all stated outcomes and were considered low risk for reporting bias. Six studies have not yet reported on some secondary outcomes (Berger 2010 on heart failure mortality and all‐cause admission, Eurlings 2010 on all‐cause admission, Persson 2010 and Maeder 2013 on quality of life, Schou 2013 and Shah 2011 on treatment costs). Lainchbury 2010 partially reported quality of life data. Skvortsov 2015 is currently awaiting further publications. It was not possible to assess reporting bias for Shochat 2012 as data were provided from conference abstracts and direct contact with the author and any pre‐specified outcomes were not stated.

Other potential sources of bias

Eight of the studies were part or fully funded by pharmaceutical companies (Berger 2010; Januzzi 2011; Jourdain 2007; Krupicka 2010; Maeder 2013; Persson 2010; Pfisterer 2009; Shochat 2012). Five studies (Eurlings 2010; Karlstrom 2011; Schou 2013; Shah 2011; Troughton 2000) were partially funded by either national research grants, lotteries, hospital funds and/or pharmaceutical companies. Four studies did report funding sources (Anguita 2010, Beck‐da‐Silva 2005; Li 2015; Skvortsov 2015). These studies were judged to be of unclear risk of bias.

One study (Lainchbury 2010) was solely funded from a national research body and therefore considered at low risk of bias from the funding source.

All‐cause mortality

(See Analysis 1.1)

Sixteen studies (Anguita 2010; Beck‐da‐Silva 2005; Berger 2010; Eurlings 2010; Jourdain 2007; Karlstrom 2011; Krupicka 2010; Lainchbury 2010; Maeder 2013; Persson 2010; Pfisterer 2009; Schou 2013; Shah 2011; Shochat 2012; Skvortsov 2015; Troughton 2000) with 3292 participants recruited, reported results for all‐cause mortality. Follow‐up ranged from one month to four and a half years. However, data for Maeder 2013 was presented as survival curves and it was not possible to extract or obtain data for this study. Therefore meta‐analysis was only possible for the remaining 15 studies: During the follow‐up period, 265 (18%) participants died in the NP‐guided treatment groups compared to 368 (22%) in the control groups. When the data were pooled for all studies using a fixed‐effect model, the evidence favoured the guided treatment groups, but overall the evidence showed uncertainty (risk ratio (RR) 0.87, 95% confidence interval (CI) 0.76 to 1.01; patients = 3169; studies = 15; low quality of evidence). Heterogeneity was low (I² = 16%).

The two studies that did not report results for all‐cause mortality were Januzzi 2011 and Li 2015.

Heart failure mortality

(See Analysis 1.2)

Only six studies (Jourdain 2007; Karlstrom 2011; Krupicka 2010; Li 2015; Skvortsov 2015; Troughton 2000) with 853 participants recruited reported results for heart failure mortality. In the NP‐guided treatment groups, 34 participants died and in the control groups 38 participants died due to heart failure, representing 8% and 9% respectively. Similar to all‐cause mortality, the pooled result, using a fixed‐effect model, favoured the intervention, but overall, the evidence showed uncertainty (RR 0.84, 95% CI 0.54 to 1.30; participants = 853; studies = 6; low quality of evidence). The heterogeneity was low (I² = 21%).

Heart failure admission

(See Analysis 1.3)

Ten studies (Anguita 2010; Berger 2010; Januzzi 2011; Jourdain 2007; Karlstrom 2011; Krupicka 2010; Lainchbury 2010; Schou 2013; Skvortsov 2015; Troughton 2000) with 1928 participants reported on heart failure admission. Out of 858 participants, 219 (26%) experienced a heart failure event causing an admission in the NP‐guided treatment groups; this compared to 403 out of 1070 (38%) participants in the control groups. Overall, the pooled evidence for all 10 studies, with a fixed‐effect model, showed an effect favouring NP‐guided treatment (RR 0.70, 95% CI 0.61 to 0.80; participants = 1928; studies = 10; low quality of evidence). Heterogeneity was substantial (I² = 60%). The robustness of this finding was tested by converting to a random‐effects model; the effect remained consistent (RR 0.67, 95% CI 0.53 to 0.84; participants = 1928; studies = 10; low quality of evidence).

All‐cause admission

(See Analysis 1.4)

Six studies (Beck‐da‐Silva 2005; Jourdain 2007; Karlstrom 2011; Schou 2013; Shah 2011; Troughton 2000) with 1142 participants recruited reported data for all‐cause admission. During the follow‐up, 304 (53%) participants experienced an event requiring admission in the NP‐guided treatment groups. This compared to 327 (57%) participants in the control groups. The pooled results for all studies, with a fixed‐effect model, favoured the intervention, but overall, the evidence showed uncertainty (RR 0.93, 95% CI 0.84 to 1.03; participants = 1142; studies = 6; low quality of evidence). No heterogeneity was identified (I² = 0%). Lainchbury 2010 commented that no difference was seen between intervention and control groups for all‐cause admission, but the data were not provided.

Adverse events

(See Table 3)

Six studies (Januzzi 2011; Krupicka 2010; Maeder 2013; Persson 2010; Pfisterer 2009; Troughton 2000) with 1144 participants reported number of adverse events during follow‐up. Maeder 2013 did not report the number of adverse events broken down by intervention group, only as a total for the study. For the remaining five studies, the NP‐guided treatment groups (511 participants) experienced 215 compared to 184 adverse events in the control groups (510 participants). Meta‐analysis was not viable for this outcome since it was possible to have multiple events per individual. Therefore, the results have been tabulated. Quality of evidence was low.

Nevertheless, three studies (Januzzi 2011; Pfisterer 2009; Troughton 2000) commented there was no difference between the NP‐guided treatment and control groups: Januzzi 2011 reported that there was no significant differences between the groups, whilst Pfisterer 2009 and Troughton 2000 reported P values greater than 0.05. Maeder 2013 reported the number of patients experiencing a serious adverse event did not differ between the groups. Two studies (Januzzi 2011; Krupicka 2010) reported a complete breakdown of the nature of the adverse events, whilst Pfisterer 2009 and Maeder 2013 only highlighted two areas (renal impairment and hypotension). For Maeder 2013, adverse events for renal failure were more frequent in the NP‐guided group, where as events were less frequent for hypotension compared to the control group. However, both Januzzi 2011 and Pfisterer 2009 confirmed no difference between the groups based on specific adverse events. Incomplete data meant it was not possible to comment on the most frequent types of adverse events.

Cost

Four studies (Berger 2010; Januzzi 2011; Maeder 2013; Pfisterer 2009) presented data on costs, two only as conference abstracts. It was not possible to pool results for these four studies because the outcome measure differed for each study. Pfisterer 2009 reported on total overall costs per intervention arm: $20,949 for the NT‐proBNP‐guided treatment group versus $23,928 in the symptom‐guided group (control). Generally, costs were comparable, the main difference occurred in the residency costs (staying in a nursing home or home for the elderly): $4157 in the NT‐proBNP‐guided treatment group versus $7564 in the symptom‐guided group.

Januzzi 2011 examined the mean costs in the duration of the study. Overall costs for the NT‐proBNP group totaled $35,262 ($451 per day) versus overall costs for the standard of care management (control) group of $42, 629 ($580 per day). Similar to Pfisterer 2009, the lower costs in the NT‐proBNP group was predominantly due to inpatient costs. Januzzi et al concluded that costs were reduced by approximately 20% in the NT‐proBNP‐guided treatment group over the 10‐month follow‐up period.

In Berger 2010 an economic analysis was completed for a subgroup of participants (n = 190) who had complete follow‐up data. This analysis suggested NP‐guided treatment was cost‐effective and cheaper than in the usual care control group (for the multidisciplinary care control group this was cost neutral).

In contrast to the above three studies Maeder 2013 reported NP‐guided therapy as unlikely to be cost‐effective. Overall costs being $38,876 per patient for the NP‐guided group compared to $21,419 per patient in the control group over 18 months.

Quality of evidence was low.

Quality of Life

(See Analysis 1.5)

Quality of life data were reported in eight studies ((Beck‐da‐Silva 2005; Eurlings 2010; Karlstrom 2011; Lainchbury 2010; Pfisterer 2009; Schou 2013; Skvortsov 2015; Troughton 2000) with 1812 participants recruited using the Minnesota Living with Heart Failure questionnaire. Lainchbury 2010 is only represented by one data set as data were only reported for the usual care control group. The pooled evidence for all studies, using a fixed‐effect model, marginally favoured NP‐guided groups, but overall, the evidence showed uncertainty (mean difference (MD) ‐0.03, 95% CI ‐1.18 to 1.13; very low quality of evidence). Heterogeneity was judged to be substantial (I² = 75%).

Pfisterer 2009 also reported results for quality of life using the Short Form 12 and Duke Activity Status Index questionnaires; though not included due to incompatibility, both of these showed an improvement in both guided treatment and control groups with no differences in the degree of improvement.

In Karlstrom 2011, changes in quality of life for participants was measured using the Swedish and Norwegian Short Form Health Survey 36; 68% from the NP‐guided group and 74% from the control group completed the survey at both the start and end of the study. For these participants NP‐guided treatment did not improve quality of life compared to clinical assessment alone.

Participants in Persson 2010 completed the Kanas City Cardiomyopathy Questionnaire at baseline and follow‐up. This symptom score tool contains a quality of life element. In Persson 2010, the scores improved in both groups (+3.6 (SEM 1.65) in the NT‐proBNP group and +6.2 (SEM 1.66) in the control group). There was no differences between the groups (P = 0.28).

Subgroup analysis

Except for age, it was not possible to explore subgroups within the study populations. Data were reported for severity of heart failure, baseline NT‐proBNP, target NT‐proBNP, achieved NT‐proBNP/BNP drop and gender, but generally only as totals, in varying categories, or as averages, for intervention and control groups (Table 1, Table 2). Post hoc, consideration was given to subgrouping by left ventricular ejection fraction, (LVEF), but this too was not reported in an appropriate form (Table 1). All studies were completed under supervision of the hospital, except for Berger 2010 and Lainchbury 2010 where supervision was jointly in hospital and the community, and therefore subgroup analysis for this factor was not completed.

Subgroup analysis was only possible by age for three studies (Eurlings 2010; Lainchbury 2010; Shochat 2012) and only for the primary outcome of all‐cause mortality (see Analysis 3.1). From the three studies, including Lainchbury 2010 with two control groups, there were 830 participants. For this analysis, the age threshold was set as equal or greater than 75 years old versus under 75 years old, though the data from Eurlings 2010 are reported marginally different as greater than 74 versus equal to or less than 74 years old. When the data from these three studies were pooled, the evidence showed uncertainty for either age subgroup. However, whilst showing uncertainty for either age subgroup the results suggest that for participants equal to or greater than 75 years old, the effect favoured the control groups (RR 1.23, 95% CI 0.96 to 1.57; participants = 410; studies = 3) whilst for participants less than 75, the effect favoured the guided‐treatment groups ((RR 0.73, 95% CI 0.49 to 1.10; participants = 420; studies = 3) (Analysis 3.1).

Lainchbury 2010 further reported data by age for heart failure admission (=/< 75 years: RR 1.13, 95% CI 0.77 to 1.64; participants = 188; < 75 years: RR 0.73, 95% CI 0.45 to 1.17; participants = 177) (Analysis 3.2). The data followed a similar trend to the pooled data for age and all‐cause mortality.

Despite data not being available to pool, three further studies did comment on the age of participants in their results. Januzzi 2011 concluded for their study that 'no interaction between NT‐proBNP‐guided care and age was found (P = 0.11)'. Persson 2010 commented 'levels of NT‐proBNP tended to decrease more in patients younger than 75 years than in patients older than 75 years (change ‐2.4% ≥75 versus ‐20.3% <75 years, P = 0.06). Finally, Pfisterer 2009 reported that in the first six months the BNP levels decreased similarly for both guided treatment and control groups and were similar for participants under 75 and equal to or over 75 years of age. Though Pfisterer 2009 did state that "there was a significant interaction between treatment and age groups, i.e. patients aged ≥ 75 years in the NT‐proBNP group had a smaller relative benefit on NT‐proBNP levels (p = 0.04) and symptoms (p = 0.05) than younger patients". At eighteen months, the interaction between treatment and age was significant for mortality (P = 0.01, Cox regression adjusting for baseline characteristics) indicating that 'NT‐proBNP‐guided treatment differed significantly between younger and older patients'.

Post hoc subgroup analysis was carried out to explore whether data from two studies (Berger 2010; Lainchbury 2010) using usual care differed to all other studies using clinical assessment as the comparator to NP‐guided treatment (Analysis 2.1). This was only possible for two outcomes. For the primary outcome of all‐cause mortality, the evidence showed very little difference for either subgroup (usual care RR 0.79, 95% CI 0.56 to 1.13; participants = 319; studies =2; clinical assessment RR 0.89, 95% CI 0.76 to 1.04; participants = 2850; studies = 15) to each other or compared to the overall pooled result (RR 0.87, 95% CI 0.76 to 1.01; participants = 3169; studies = 15; low quality evidence) (Analysis 1.1). Similarly, for heart failure admission there was very little difference for either subgroup (usual care RR 0.72, 95% CI 0.53 to 0.99; participants = 319, studies = 2; clinical assessment RR 0.70, 95% CI 0.60 to 0.81; participants = 1609, studies = 10) to each other or the overall pooled result (RR 0.70, 95% CI 0.61 to 0.80; participants = 1928; studies = 10; low quality evidence) (Analysis 1.3).

Post‐hoc we explored the effect of duration of the intervention on outcomes. Analysis 6.1 shows that both at ≤ one year (RR 0.46, 95% CI 0.25 to 0.85; participants = 555; studies = 5; P =0.01; I² = 0%) and between one and two years (RR 0.83, 95% CI 0.69 to 0.99; participants = 1842; studies = 8; P =0.04; I² = 0%), there was a potential reduction for all‐cause mortality, but the evidence showed uncertainty at > two years (RR 1.11, 95% CI 0.87 to 1.41; participants = 772; studies = 2; P = 0.41; I² = 0%) and the subgroup test for difference was significant (P =0.02). The effect of duration on heart failure admission shows a similar trend for each subgroup (≤ one year: RR 0.37, 95% CI 0.23 to 0.58; participants = 278; studies = 3, one to two years: RR 0.65, 95% CI 0.54 to 0.79; participants = 878; studies = 5; > two years: RR 0.97, 95% CI 0.77 to 1.23; participants = 772; studies = 2), again the test for subgroup effect was significant (P = 0.0004) Analysis 6.3. For heart failure mortality (Analysis 6.2), all‐cause admission (Analysis 6.4) and quality of life (Analysis 6.5), the subgroups all showed uncertainty similar to the overall pooled result for each outcome.

Post hoc we also explored the assumption that the two biomarkers were sufficiently biologically and clinical similar to evaluate together. We investigated this by separating the pooled data by each biomarker. For all‐cause mortality (Analysis 7.1), heart failure mortality (Analysis 7.2), all‐cause admission (Analysis 7.4) and quality of life (Analysis 7.5), the pooled data for each biomarker showed uncertainty and were similar to the overall pooled result for each outcome. For heart failure admission, using a fixed‐effect model, the result grouping the trials by BNP (Anguita 2010; Jourdain 2007; Karlstrom 2011; Krupicka 2010), or NT‐ProBNP (Berger 2010; Januzzi 2011; Lainchbury 2010; Schou 2013; Skvortsov 2015; Troughton 2000) did not make a difference to the main findings (BNP: RR 0.70, 95% CI 0.56 to 0.87; participants = 600; studies = 4; NT‐proBNP: RR 0.70, 95% CI 0.59 to 0.84; participants = 1328; studies 6) Analysis 7.3. In view of the substantial heterogeneity we tested the robustness of this finding using a random‐effects model and found that the pooled result for studies using the BNP marker continued to favour NP‐guided treatment but now showed uncertainty (BNP: RR 0.68, 95% CI 0.43 to 1.05; participants = 600; studies = 4; NT‐proBNP: RR 0.65, 95% CI 0.48 to 0.89; participants = 1328; studies 6).

Sensitivity analysis

Risk of bias within the studies varied across the aspects of bias assessed. Blinding of participants and study personnel appeared to be poor (see Figure 2 and Figure 3), nevertheless, it was not always practical to blind participants and personnel in some studies. High risk in this category could still mean one party was blinded. Blinding of outcome assessment and attrition was judged to potentially impact on the pooled results.

Sensitivity analyses were completed restricting studies to those with low risk of bias for blinding of outcome assessment (Berger 2010; Eurlings 2010; Karlstrom 2011; Lainchbury 2010; Schou 2013) and for attrition (Anguita 2010; Berger 2010; Jourdain 2007; Li 2015; Schou 2013; Shah 2011; Skvortsov 2015; Troughton 2000). For all outcomes, the analyses produced a similar effect to the main findings (see Table 4). Though there was only one study (Karlstrom 2011) assessed as low risk for detection bias for heart failure mortality and therefore no comparison with the main findings could be made in this instance.