NT-proBNP for Predicting All-Cause Death and Heart Transplant in Children and Adults with Heart Failure

Data for this study came from two sources. Electronic medical record data for children with HF were extracted from the IBM Explorys™ database, which had also provided the data used in another study to evaluate adults with HF [19]. The Pediatric Cardiomyopathy Registry (PCMR) [NCT00005391] [21] provided registry data from the National Institutes of Health-supported PCMR sub-study called Pediatric Cardiomyopathy Biomarkers [NCT01873976] [22].

The period covered by Explorys is about 20 years; that is, records go back to about 2000. Data were abstracted on the 10th and 18th of March 2020 from a subsection of the database specifically comprising all records of patients with HF, as defined by the Systematized Nomenclature of Medicine, Clinical Terms (SNOMED-CT) concept ID 84114007.

Data for children meeting the following criteria were abstracted:

The study endpoint was the composite of “all-cause death” or “heart transplant” (SNOMED codes 32413006, 47058000, or 32477003) over 2 years as in the adult cohort [19]. We wanted to study children with chronic HF at entry into the study. Accordingly, the baseline was set at 1 month after the first recorded HF diagnosis because the first diagnosis for HF is often associated with an acute event, such as decompensated HF. Mortality in such a crisis is higher in this comparatively short period, in which children often either recover or die [5, 18]. This definition of baseline means that the date of the first documented HF diagnosis is one month before the date of “enrollment” into our virtual real world data-based study. For simplicity, we will therefore refer to “enrollment date” for the remainder of this paper for PCMR and for Explorys. We used the median NP-proBNP value in the follow-up period for modeling.

The Pediatric Cardiomyopathy Biomarker sub-study was a prospective cohort study of children with dilated cardiomyopathy (DCM) [22]. Children were enrolled between 2013 and 2016 and followed for up to 5 years. The date of diagnosis was the date of the earliest diagnostic echocardiogram, cardiac magnetic resonance imaging study, or report that confirmed the diagnosis of DCM. Only children with a diagnosis of DCM at baseline were included in these analyses. As in the case of the Explorys cohort, the primary endpoint for children in the PCMR was the composite of all-cause death or heart transplantation, but the rates of events were determined within 5 years after enrollment in the registry to include a sufficient number of events because the incidence was lower than that in the Explorys population. The NT-proBNP values used for modeling were measured at the time of enrollment. Each participating center obtained Institutional Review Board or Ethics Committee approval for the study.

IBM Explorys is a commercial database fully compliant with the Health Insurance Portability and Accountability Act and the Health Information Technology for Economic and Clinical Health Act. Therefore, no approval by the institutions’ human research committee was required, and informed consent by the patients was not obtained. In the PCMR database, written informed consent was obtained from parents or guardians when children were enrolled.

In adults with HF irrespective of left ventricular ejection fraction (LVEF), an E_max-type function describes the relationship between individual NT-proBNP concentrations and probabilities for clinical events [19]. The same equation was used to model the data for children because a comparable relationship was expected.

Where, P_i: probability for an event, P₀: baseline probability for an event when the NT-proBNP concentration is very low, P_max: maximum probability added onto the baseline value, EC₅₀: NT-proBNP concentrations corresponding to the half-maximum risk for events, γ: Shape parameter determining the steepness of the curve, NT-proBNP_i: individual plasma NT-proBNP concentrations.

E_max functions are usually used to describe the concentration–response relationships of drugs and have a non-linear shape, flattening out at high concentrations and resulting in a maximal effect. In addition, the most general E_max function starts at a potential non-zero (placebo) effect at a concentration of zero and may have a more or less s-shape, as determined by the shape parameter γ. The location of the efficacy curve on the concentration axis is determined by the parameter EC₅₀.

It is important to note that the parameters P₀ and P_max cannot be expected to be independent of the scenario the observed data are derived from, while for EC₅₀, this might very well be the case. Thus, both, P₀ and P_max, will in any case depend on the patients’ ages, as well as on their observation periods. Also co-morbidities may have an influence. However, the rise of the risk from baseline to the maximal risk with increasing NT-proBNP concentration, which is determined by EC₅₀ and γ, may follow the same shape in different cases. Therefore, the model is not able to predict absolute risks, but within a patient cohort for which P₀ and P_max are determined, the relative risks of different subgroups or individuals can be predicted.

Model parameters were estimated with a maximum likelihood estimator. The number of adjustable parameters was reduced if necessary for a stable regression. Confidence intervals were estimated using a local approximation of the variance–covariance matrix on finite difference-based second derivatives of the likelihood function. Confidence intervals for estimates were derived through forward propagation of the parameter uncertainty using 10,000 random samples drawn from the approximated variance–covariance matrix. All evaluations and modeling were done with the statistical analysis software R V 4.0.2 and SAS Enterprise Guide V8.3.

In the Explorys database, 109 records met the inclusion criteria and were included in the analysis. Of the 150 eligible children included in the PCMR, NT-proBNP data were available for 146 (Table 1).

By definition, children in the PCMR cohort all had DCM, whereas 36% of the children in Explorys cohort had cardiomyopathy and a larger group had congenital cardiovascular malformations. For example, 40% had congenital cardiovascular malformations. Median plasma NT-proBNP concentration was considerably higher in the Explorys cohort than in the PCMR cohort (1250 pg/mL vs. 184 pg/mL). The use of diuretics and inotropic agents was much higher in the Explorys cohort than in the PCMR cohort (85% vs. 32% and 73% vs. 18%, respectively). The percentages of children treated with other cardiovascular medications were similar in both cohorts. The Explorys cohort seemed to include more severely ill children, given the proportion receiving inotropic agents (73%) and their higher rates of death or heart transplantation.

The age distributions were largely similar in the cohorts (Fig. 1). There was a bimodal age distribution in both cohorts. Interestingly, NT-proBNP concentrations were bimodal for both cohorts, in contrast to those of adults with HF, where the concentrations usually had a log-normal distribution [23]. The PCMR cohort includes more children with NT-proBNP concentrations below 400 pg/mL than does the Explorys cohort. The distribution of NT-proBNP concentrations above 400 pg/mL, however, is similar in both cohorts.

The dashed and dotted lines in Fig. 2 represent event rates and 90% confidence intervals in NT-proBNP bins defined by quintiles (PCMR; 146 patients) and quartiles (Explorys; 109 patients) of the respective NT-proBNP distributions. Both curves feature a steep increase in event rates around NT-proBNP above ~ 1000 pg/ml. Note that in both cases and in particular in the case of the PCMR cohort there are little data in the high NT-proBNP range as it would be required for identifying the P_max value (1) with high accuracy.

The Explorys data could be described very well using the model described in Eq. 1 keeping the shape parameter γ fixed at 1 to keep the number of free parameters as small as possible. The estimated values of the optimized model parameters (90% CI) are P₀ = 0.094 (0.065), EC₅₀ = 3730 (3140) pg/mL, and P_max = 0.65 (0.18). The substantial width of the confidence intervals reflects the small number of patients and consequently, the large uncertainty in the data.

Simulated values for the PCMR cohort (Eq. 1) were also similar to the observed values (Fig. 2). In this cohort, however, the shape parameter γ was included in the parameter estimation, whereas P₀ and P_max were fixed at 0 and 1, respectively. By doing this, the approach is equivalent to a log-logistic regression. The choice of P₀ = 0 was justified because there was no event in children in the lowest quintiles of the NT-proBNP distribution. E_max on the other hand could not be identified in the optimization process because of a lack of data at very high NT-proBNP concentrations. The estimated values of the two remaining parameters in the model were 4199 pg/mL for EC₅₀ and a mean (SE) of 1.21 (0.24) for γ.

The relationship between NT-proBNP concentration and mortality in adults with HF in the Explorys database [19] was determined with the same model equation used above, so the resulting models can easily be compared (Table 2).

In the Explorys database, the number of heart transplants in adults was negligible compared to deaths, so this endpoint was ignored in the comparison. None of the estimated parameters differed markedly between children and adults in the Explorys cohorts, as shown by their overlapping confidence intervals. This implies that the relationship between NT-proBNP and mortality and heart transplant is comparable in both cohorts (Fig. 3). In the case of the PCMR cohort, however, the fact that baseline and maximum probability for an event was fixed to 0 and 1 does not allow a statistical assessment of the deviation of these parameters from those for the other two cohorts. The EC₅₀ values, however, are statistically indistinguishable in all three cases.

Respectively, all model curves (of both the pediatric and adult cohorts) coincide in the range of their EC₅₀ values and their confidence intervals (Fig. 3). Differences in P₀ between the adult and pediatric Explorys cohorts could be explained by the severity of the disease status of the affected children. The value of P_max for the PCMR cohort is not informed by the data, which was insufficient in the high NT-proBNP range and was thus set to 1 in the model. Therefore, the apparent deviation from the other two curves may be purely artificial and cannot be interpreted.

We combined data from two small cohorts of children with HF from different causes. Whereas the Pediatric Cardiomyopathy Biomarker sub-study of the PCMR was prospective and had pre-specified sampling timepoints, the Explorys analysis was retrospective and used NT-proBNP concentrations as collected. Furthermore, the PCMR cohort met pre-specified enrollment criteria, whereas the Explorys cohort was selected only by diagnostic codes resulting in a less well-defined group. Heart failure in children in both cohorts could be new or chronic. For children in the Explorys cohort, we used the median concentration of the available NT-proBNP values in the follow-up period, whereas we used the NT-proBNP concentration at enrollment for children in the PCMR cohort. Additionally, serial data were not available. Such data are needed to assess the predictive value of changes in NT-proBNP concentrations and whether natriuretic peptides or other biomarkers could be part of goal-directed therapy [32]. Furthermore, as described in the discussion both datasets were too small for testing equivalence of models in a statistically robust sense and hence, the assessment has to be semi-quantitative.

A possible limitation is that there is a substantial number of pediatric heart failure patients with congenital cardiovascular malformation diagnoses in the Explorys database. This may be viewed as somewhat concerning as the cause for serum NT-proBNP concentrations theoretically could be increased due to this (such as shunting) beyond what might be seen in DCM (pump failure). We have previously compared the association between serum NT-proBNP concentrations in children with heart failure from with DCM with children with heart failure with congenital cardiovascular malformations and published that when we compared the entire sample to the subgroup of children with DCM without congenital cardiovascular malformations assess confounding we found that NT-proBNP increased significantly with the severity of symptoms with a volume under the receiver operating characteristic surface estimate of 58% which was significantly more frequent than what would have been expected for an uninformed marker [31, 33]. However, we did find that when patients with congenital cardiovascular malformations were excluded NT-proBNP still increased significantly with the severity of symptoms but its discriminatory ability was lower [33]. We concluded that comparing the effect estimates in the entire sample to those within the subgroup of children with congenital cardiovascular malformations may have confounded the associations between NT-proBNP and symptom severity in the direction of minimizing a statistically significant association [33]. Therefore, theoretically comparing the Explorys cohort with thee PCMR cohort may have a bias toward not having a significant association. Since these two groups of pediatric HF patients were quite similar in this analysis, this suggests that the true association of serum NT-proBNP concentration in pediatric HF patients from either DCM or congenital cardiovascular malformations may be similar and in fact may be even more significant in pediatric DCM HF patients that demonstrated here since the direction of relation to symptoms in congenital cardiovascular malformation patients may be lower. In our prior work, we found that there were differences between children with HF from DCM when compared to children with HF from congenital cardiovascular malformations in terms of medications and measurements of LV structure and function [31, 33]. Nevertheless, these groups were comparable using serum NT-proBNP concentrations with respect to HF severity in our prior analyses [33] and in terms of mortality in this paper. So, although this is a limitation the serum NT-proBNP concentrations between these two pediatric HF populations and an adult HF population are important comparisons.