A prediction model for estimating NT‐proBNP in a general Japanese population: the Toon Health Study

Abstract

Aims

As part of the Toon Health Study, which is an ongoing population‐based cohort study, we aimed to develop a prediction model for N‐terminal pro‐brain natriuretic peptide (NT‐proBNP) in a general Japanese population. We sought to explore the influence of various demographic and clinical factors on NT‐proBNP levels and assess the model's performance. In addition, our objectives included internal validation and investigation of the diagnostic potential of the observed‐to‐predicted NT‐proBNP ratio (OPR) at baseline for predicting the risk of heart failure with preserved ejection fraction (HFpEF).

Methods and results

In this prospective cohort study, participants were recruited from Toon City, Japan, as part of the larger Toon Health Study, focusing on cardiovascular risk factors. We measured the NT‐proBNP levels and used linear regression with penalization (ridge regression) to develop the model. The model incorporated 10 prespecified predictors (age, gender, body mass index, diastolic blood pressure, heart rate, haemoglobin, albumin, total cholesterol, haemoglobin A1c, and estimated glomerular filtration rate) and underwent assessment using R ² and root mean squared error (RMSE). Internal validation was conducted through bootstrapping. In a post hoc analysis, we explored the OPR's diagnostic potential using 5 year follow‐up data (n = 636) to predict the elevation of NT‐proBNP > 125 pg/mL at the 5 year follow‐up as the risk of HFpEF. A total of 2505 participants (age: 60.4 ± 12.9 years, men: 35%) were enrolled in this study. There was a linear relationship between the observed and predicted values of NT‐proBNP in which the logarithm of observed NT‐proBNP was <6, which corresponds to 403 pg/mL in NT‐proBNP. The prediction model demonstrated satisfactory performance (R ²: 0.291, RMSE: 0.688), with age identified as a dominant predictor. The stability of the model was underscored by the internal validation. The OPR at baseline predicted NT‐proBNP > 125 pg/mL at the 5 year follow‐up with an area under the curve of 0.793.

Conclusions

This study introduces the first prediction model for NT‐proBNP in a general Japanese population. Although the model has acceptable performance, ongoing refinement is essential. Our transparent approach to model development, alongside a web‐based interactive tool, lays the groundwork for further improvements and external validation. The OPR holds potential for predicting the future risk of HFpEF. This research contributes to understanding the nuanced influence of patient backgrounds on levels of NT‐proBNP in asymptomatic individuals within the context of a broader population‐based cohort study.

Introduction

As the aged population increases, the number of patients with heart failure is increasing considerably worldwide. ¹ Despite improvements made in therapeutic interventions, the reduction in heart failure morbidity and mortality rates remains unsatisfactory. ²

The 2022 American Heart Association/American College of Cardiology/Heart Failure Society of America guideline grades four stages of heart failure in this continuum as follows: patients at risk (stage A), asymptomatic (stage B), symptomatic (stage C), and advanced stage (stage D). ² The guideline emphasizes intervening with patients in an earlier stage (A and B) of heart failure by ameliorating the underlying heart failure risk factors. If patients are hospitalized due to heart failure (stage C), their quality of life and outcomes gradually worsen, even if medical and interventional treatments temporarily relieve symptoms and signs of heart failure. Thus, it is important to screen patients with stage A and B heart failure and prevent the transition from asymptomatic stage (stages A and B) to symptomatic stage (stage C).

Suga et al. demonstrated that brain natriuretic peptide (BNP), as a family of natriuretic peptides (NPs), was produced in cardiac myocytes. ³ Human BNP, a biologically active 32 amino acid, is separated from the N‐terminal pro‐brain natriuretic peptide (NT‐proBNP), which is a constitutive peptide with unknown biological effects. ³ BNP and NT‐proBNP are clinically useful for detecting heart failure; however, these NPs are well known as being affected by various factors such as age, gender, body mass index (BMI), blood pressure, renal function, haemoglobin, lipid profiles, and diabetes. ⁴ , ⁵ , ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰ Hence, it is conceivable that the significance of the same NT‐proBNP value in diagnosing heart failure may vary based on different patient backgrounds. Because these NP confounding factors have considerable overlap with heart failure risk factors, creating a prediction model of NT‐proBNP from clinical and laboratory data is beneficial.

The Toon Health Study is a prospective community‐based cohort study designed to investigate the risk factors of cardiovascular disease. Accordingly, we aimed to develop a population‐based prediction model of NT‐proBNP using the serum level of NT‐proBNP of voluntary residents.

Methods

We followed the TRIPOD statement (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis) to develop and validate the prediction model. ¹¹ , ¹²

Study design and source of data

We conducted this study as a part of the Toon Health Study, which was a prospective population‐based cohort study with the aim of assessing cardiovascular disease risk factors in Toon City, Ehime Prefecture, Japan. ¹³ , ¹⁴ , ¹⁵ Toon City is a Japanese rural city in Ehime Prefecture with 33 000 inhabitants. In this cohort study, we recruited the participants and registered their baseline characteristics from 2009 to 2012 as the first term, followed by two additional recruitments from 2014 to 2017 as the second term and from 2018 to 2021 as the third term (Figure  1 ). Follow‐up was conducted every 5 years after registration, and the follow‐up assessment was performed from 2014 to 2021. All participants provided written informed consent. This study protocol was approved by the Institutional Review Board (IRB) of Ehime University Graduate School of Medicine (IRB: 2007004).

Figure 1.

Study flow chart in the Toon Health Study.

Study population

For this cohort study, we recruited participants from among the residents of Toon City. The inclusion criteria were as follows: (i) voluntarily participating residents in Toon City and (ii) no serious symptoms and diseases. Using these criteria, we enrolled 2505 participants (male: 888, female: 1617) in this study. Figure 1 shows the study flow chart. One subject was excluded from the analyses due to withdrawal of consent.

Study outcome

The study outcome was the serum level of NT‐proBNP measured from the serum samples stored at −80°C at baseline. NT‐proBNP was measured by immunoassay with Elecsys proBNP II (Roche Diagnostics). We included cases in which no stored blood remained after several blood tests as a sample with missing NT‐proBNP data.

Candidate predictors

It is recommended that in order to develop an adequate prediction model, it is reasonable to create a list of 5–20 potential predictors based on a review of the literature on the specific topic, combined with consulting experts in the field. ¹⁶ Accordingly, based on a review of previous literature and expert opinions, which were collected at the baseline assessment, we included the following 10 explanatory variables for NT‐proBNP: demographic variables (age, gender, and BMI), physiological measurements [diastolic blood pressure and heart rate (HR)], and blood test results [haemoglobin, albumin, total cholesterol, haemoglobin A1c (HbA1c), and estimated glomerular filtration rate (eGFR)]. ⁴ , ⁵ , ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰ All variables were treated as continuous values, except for gender, which was treated as a categorical value (male or female). We calculated BMI from height and weight, which were also collected at the baseline assessment. eGFR was calculated based on serum creatinine level using the following formula:

$$\begin{gathered} \text{Male}:194\times {\text{serum creatinine}}^{-1.094}\times 1.094\times {\mathrm{age}}^{-0.287}, \\ \text{female}:\text{eGFR}\left(\text{male}\right)\times 0.739. \end{gathered}$$

For all continuous variables, we checked the distribution. If we found the skewness, we transformed these variables using the appropriate method (log, square, or inverse). Continuous valuables were standardized, and a categorical variable (gender) was transformed into dummy variables.

Sample size calculation

Based on the criteria proposed by Riley et al., ¹⁷ we calculated the sample size needed for the linear regression. Because there is no previous proposed prediction model in the literature, we set a predicted R ² value conservatively to 0.2. We also set the mean and standard deviation (SD) to 64 and 53, respectively, calculated from previous reports based on the method proposed by Wan et al. ¹⁸ With 10 parameters and a possible shrinkage value of 0.9, we determined that a total of 313 samples were required, and our sample size of 2505 was thought to be sufficient for this study.

Missing data

For missing data, we imputed the data set using multiple imputation by chained equations ¹⁹ and created 10 imputed data sets. Each completed data set was used for the analysis separately, and the results from each data set were summarized according to Rubin's rules. ²⁰ , ²¹

Model development and model performance

For model development, we used linear regression with penalization to reduce overfitting and optimism. We used ridge regression to avoid selecting some of the candidate parameters (ridge). To choose the optimal hyperparameter of penalization (lambda), we performed a sequential search from 0 to 1 by 0.01 with 10‐fold cross‐validation. In addition, using standard linear regression without penalization and random forest method, we developed benchmark models for comparison. All models were developed after the parameters were standardized (centring and scaling). Regression coefficients were reported as a simple mean of the results from 10 imputed data sets according to Rubin's rules. ²⁰ , ²¹ To evaluate model performance, we assessed the predictive performance of each model by R ² statistic and root mean squared error (RMSE). The results were reported as the median with interquartile range (IQR) according to Rubin's rules. ²⁰ , ²¹

Model validation

Using a bootstrap procedure with 200‐time repetitions to calculate the optimism‐corrected R ², we assessed the internal validation of the model to evaluate its performance. The optimism‐corrected R ² was calculated separately for each of the 10 imputed data sets, and the results were reported as a median with IQR according to Rubin's rules. ²⁰ , ²¹

Post hoc analysis

We also conducted two post hoc analyses based on the results of the prespecified analysis. First, because the influence of age on the predicted value was much higher than that of other predictive factors, we added a term of square of age to the prespecified 10 parameters as a sensitivity analysis to assess performance improvement.

Second, to assess the diagnostic possibility of predicting the future risk of heart failure with preserved ejection fraction (HFpEF), we analysed participants for whom 5 year follow‐up data were available. We defined the observed‐to‐predicted NT‐proBNP ratio (OPR) as a new indicator, which was calculated by the observed value of NT‐proBNP divided by the predicted value at baseline, and the risk of HFpEF was defined as an NT‐proBNP level > 125 pg/mL at the 5 year follow‐up. ²² We assessed diagnostic performance using the receiver operating characteristic (ROC) curve and area under the curve (AUC).

Statistical software

We used R Version 4.2.1 (R Foundation for Statistical Computing, Vienna, Austria) for all analyses. For the analysis, we used ‘pmsampsize’, ‘mice’, and ‘caret’ packages for the sample size calculation, missing imputation, and model development/validation, respectively. The final prediction model was presented on a website using the R Shiny application.

Results

Baseline characteristics

Figure 1 shows the study population included in this study. The number of recruited participants of each recruitment term was 2033, 381, and 91, respectively. Of the 2506 participants who were recruited across the three terms, one participant withdrew consent and was thus excluded from the study. The remaining 2505 participants were included, including 833 patients (33.3%) with missing outcome (NT‐proBNP). Supporting Information, Figure S1 shows the details of the missing values. Table 1 shows the characteristics of the prognostic parameters and outcome. Study participants included 888 men and 1617 women. The mean age was 60.4 (SD 12.9) years. The summary of other variables was expressed as mean and SD or median and IQR based on the visual inspection of the distribution.

Table 1.

Clinical and laboratory data in the Toon Health Study

Predictive variable	NA	Statistics
Age, years	21	Mean (SD)	60.4 (12.9)
Gender	0	Men/women (%)	888/1617 (35)
BMI, kg/m2	20	Median (IQR)	22.8 (20.8–25.1)
DBP, mmHg	20	Mean (SD)	75.3 (11.5)
Heart rate, b.p.m.	20	Median (IQR)	66.5 (61.0–72.5)
Haemoglobin, g/dL	114	Mean (SD)	13.6 (1.4)
T. Chol, mg/dL	114	Mean (SD)	205.8 (33.2)
Albumin, g/dL	114	Mean (SD)	4.30 (0.24)
HbA1c, %	2	Median (IQR)	5.50 (0.54)
eGFR, mL/min/1.73 m2	96	Mean (SD)	70.8 (14.0)
Outcome
NT‐proBNP, pg/mL	833	Median (IQR)	49.9 (30.2–80.7)

BMI, body mass index; DBP, diastolic blood pressure; eGFR, estimated glomerular filtration rate; HbA1c, haemoglobin A1c; IQR, interquartile range; NA, not available; NT‐proBNP, N‐terminal pro‐brain natriuretic peptide; SD, standard deviation; T. Chol, total cholesterol.

Model development

From a visual inspection of the scatter plot matrix, we checked the multicollinearity between parameters and did not detect problematic collinearity (Supporting Information, Figure S2 ). On the basis of a visual inspection of the skewness of each parameter and outcome, we decided to transform NT‐proBNP, HR, BMI, and HbA1c using logarithmic transformation. Table 2 shows the mean estimates of the coefficients of covariates from 10 imputed data sets using a ridge regression model. As mentioned in the Methods section, all parameters were standardized (centring and scaling), and the coefficients must be interpreted based on it. The estimate of the coefficient for age was relatively higher (0.27) than that of the other parameters (−0.14 to 0.08), and hence, age might be a dominant factor influencing the value of NT‐proBNP.

Table 2.

Estimates of the coefficients of 10 predicted parameters for N‐terminal pro‐brain natriuretic peptide using a ridge regression model

Coefficient	Estimate
(Intercept)	3.90
Age	0.27
Gender	0.06
Body mass index a	−0.09
Diastolic blood pressure	0.08
Heart rate a	−0.02
Haemoglobin	−0.14
Total cholesterol	0.07
Albumin	−0.11
HbA1c a	−0.04
eGFR	−0.10

eGFR, estimated glomerular filtration rate; HbA1c, haemoglobin A1c.

The mean estimates of the coefficients of covariates from 10 imputed data sets using a ridge regression model.

^a These parameters were logarithmized. All parameters were standardized (centring and scaling). The estimate of coefficient for the age is relatively higher (0.27) than that of the other parameters (−0.14 to 0.08), and age may be a dominant factor influencing the value of N‐terminal pro‐brain natriuretic peptide.

Model performance and validation

In terms of model performance, Table 3 shows the apparent performance of ridge regression, linear regression, and random forest models. The performances of ridge regression and linear regression were almost similar, with an R ² ranging from 0.289 to 0.291 and an RMSE from 0.687 to 0.688, and were slightly better than that of random forest. In terms of model validation for ridge regression, we evaluated an internal validation using bootstrap with 200 resamples. Optimism was small, with a median of −0.0032 (IQR: −0.0151 to 0.0080) for R ² and 0.0022 (IQR: −0.0007 to 0.0111) for RMSE. Optimism‐corrected performance was shown in Table 3 .

Table 3.

Apparent performance of ridge regression, linear regression, and random forest models in our prediction model

Model performance	Ridge regression		Linear regression		Random forest
	Median	IQR	Median	IQR	Median	IQR
Apparent
R 2	0.291	0.275–0.301	0.289	0.278–0.301	0.268	0.254–0.283
RMSE	0.688	0.686–0.690	0.687	0.686–0.690	0.696	0.694–0.701
Optimism corrected
R 2	0.294	0.283–0.306
RMSE	0.686	0.677–0.689
Post hoc analysis (with square term of age)
Apparent
R 2	0.291	0.275–0.301
RMSE	0.688	0.686–0.690

IQR, interquartile range; RMSE, root mean squared error.

The performances of ridge regression and linear regression were almost similar and slightly better than that of random forest. The table also shows optimism‐corrected performance. The square term of age was incorporated as an additional parameter. The model's performance did not deviate significantly from the predefined models.

Prediction model

Figure 2 shows the scatter plot between the observed and predicted values of NT‐proBNP. The smooth line showed an almost linear relationship in which the logarithm of observed NT‐proBNP was <6, which corresponds to 403 pg/mL in NT‐proBNP; however, the slope of the line was much less steep after this point.

Figure 2.

Scatter plot between the observed and predicted values of N‐terminal pro‐brain natriuretic peptide (NT‐proBNP). The predicted NT‐proBNP was correlated with the observed NT‐proBNP. However, the correlation was not found in cases with observed NT‐proBNP of >400 pg/mL in NT‐proBNP.

Post hoc analysis

Square term of age

We added the square term of age as an additional parameter. With the predefined set of 10 parameters plus the square term of age, we conducted a parallel analysis to the original specifications using ridge regression with 10 imputed data sets. The model's performance did not deviate significantly from the predefined models, yielding an R ² of 0.290 (IQR: 0.282–0.304) and RMSE of 0.687 (IQR: 0.683–0.689) (Table  3 ).

Prediction of progression to heart failure with preserved ejection fraction

Five‐year follow‐up data were available for 636 patients, among whom 102 exhibited NT‐proBNP levels exceeding 125 pg/mL. Figure 3 illustrates the ROC curve depicting the performance of the OPR indicator, with an AUC of 0.793.

Figure 3.

Prediction of heart failure with preserved ejection fraction progression. Five‐year follow‐up data were available for 636 patients, among whom 102 exhibited N‐terminal pro‐brain natriuretic peptide levels exceeding 125 pg/mL at the 5 year follow‐up. The figure illustrates the receiver operating characteristic curve depicting the performance of the observed‐to‐predicted N‐terminal pro‐brain natriuretic peptide ratio at baseline, with an area under the curve (AUC) of 0.793.

Web‐based software for predicting N‐terminal pro‐brain natriuretic peptide

We created web‐based, interactive software using R Shiny, in which a predicted NT‐proBNP was calculated from the input, based on the above‐mentioned prediction model (Figure 4 ; https://airwaystenosis.shinyapps.io/NT‐proBNP_calculator/). This value is generated as the mean of the predicted value from all 10 models from each 10 imputed data set.

Figure 4.

Web‐based N‐terminal pro‐brain natriuretic peptide (NT‐proBNP) calculator. Web‐based, interactive software using R Shiny, in which a predicted NT‐proBNP is calculated from 10 predictive variables of heart failure risk factors based on previous literatures and expert opinions. BMI, body mass index; eGFR, estimated glomerular filtration rate.

Discussion

In this study, we developed a model to predict NT‐proBNP and assessed its performance and internal validation, based on a prospective population‐based cohort with a relatively large sample size that incorporated various predictive factors and expert opinions. To our knowledge, this is the first prediction model for NT‐proBNP in the general population. The OPR, which serves as a novel diagnostic metric, holds potential for predicting the future risk of HFpEF. In addition, we presented the results of our prediction model as a web‐based interactive application using R Shiny. This could provide opportunities to validate our results with external data and update or improve the prediction model for NT‐proBNP.

Advancements in expert knowledge and improvements in computer performance in recent years have spurred a rapid increase in research on predictive modelling across various fields. Although a growing number of papers have focused on predictive models in medical research, sample sizes are typically limited, and fully data‐driven analyses, such as machine learning and deep learning, often result in unstable models with overfitting. Consequently, it is considered more advantageous to develop predictive models in medical research by incorporating insights from the existing literature and expert input, rather than relying solely on complete data‐driven analyses and using shrinkage methods such as ridge regression. ¹⁶ For these reasons, due to the relatively small sample size with missing data, we applied ridge regression with 10 prespecified parameters from the previous literature to develop the model. In addition, we used the random forest method, one of the popular machine learning methods, for comparison; however, it did not yield superior performance. Although this study represents an initial attempt to construct a predictive model for NT‐proBNP, and despite its exploratory nature, further research is warranted to refine and enhance the accuracy of the prediction model.

Generalizability is also a critical consideration when developing predictive models. Numerous studies rely on a single data set for model development, with initial emphasis placed on internal validation to scrutinize reproducibility. ²³ Furthermore, following internal validation, it is imperative to assess prediction accuracy using diverse data sets to evaluate its applicability across different scenarios. In this study, we assessed overfitting using bootstrapping to evaluate internal validity and evaluated the optimism‐corrected performance. However, to ensure external validity, evaluating the model using an independent external data set will be crucial in future investigations. Notably, as part of this study, we have made a predictive tool available online through R Shiny, and we anticipate that a subsequent study assessing external validity will be conducted using this tool.

Patients with heart failure present a spectrum of cardiovascular risk factors, encompassing advanced age, female gender, anaemia, hypoalbuminaemia, renal insufficiency, dyslipidaemia, and diabetes mellitus, all of which are particularly evident in those with HFpEF. ⁴ , ⁵ , ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰ , ²² , ²⁴ Although the mortality rate in HFpEF parallels that of heart failure with reduced ejection fraction (HFrEF), HFpEF poses a distinct challenge due to its resistance to conventional medical and interventional treatments, unlike HFrEF. ²² The cumulative impact of these risk factors imposes a haemodynamic burden on the ventricular wall, characterized by fluid retention and increased central blood pressure. Consequently, NPs are secreted in response to ventricular end‐diastolic wall stress. It is noteworthy that, independent of specific heart failure risk factors, these risk factors might also accumulate in the general population during the ageing process. In this context, the elevation of NT‐proBNP levels may stem not only from cardiac factors such as structural and functional abnormalities but also from non‐cardiac factors such as ageing, anaemia, and chronic kidney disease. ⁴ , ⁵ , ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰

Our study addresses the following fundamental question: ‘In developing a prediction model + from a general population without apparent symptoms or diseases, if the NT‐proBNP value is influenced by the background characteristics of the patient, should the cut‐off value of NT‐proBNP be adjusted based on individual patient backgrounds, rather than applying a uniform value across the entire population?’ To explore this, we developed a prediction model using data from a general population without evident morbidities, with the aim of comparing its outcomes with the observed NT‐proBNP values. This approach allows for a nuanced understanding of how NT‐proBNP values may be influenced by diverse patient backgrounds, paving the way for a more personalized and context‐specific interpretation of NT‐proBNP levels.

To comprehensively address our research question, we conducted a post hoc analysis using available 5 year follow‐up data that had not been integrated into the initial development of the model. To compare the observed and predicted NT‐proBNP values, we introduced a novel metric, the OPR, calculated as the observed value divided by the predicted value, serving as a robust indicator. Our methodology aligns with the guidelines set forth by the Heart Failure Association of the European Society of Cardiology, which highlight the negative predictive values associated with low levels (<125 pg/mL) of NT‐proBNP for excluding HFpEF. ²⁴ We conducted a thorough assessment of the discrimination capability of this ratio, specifically its efficacy in categorizing the NT‐proBNP values of 5 years later as either exceeding or not exceeding 125 pg/mL, presenting the results in a binary format. Although the resulting AUC of 0.79 was considered acceptable, it is essential to exercise caution due to the small sample size. Although suggestive, these results should be interpreted with care, as further research is necessary to validate and expand on this preliminary finding in the future.

There are several limitations in this study. First, the performance of the final model was acceptable but far from ideal (0.291 for R ² and 0.688 for RMSE), suggesting the possible presence of other important predictive factors. Because this is the first exploratory research to develop the prediction model of NT‐proBNP from patients' characteristics, further investigation on possible predictive factors is awaited to brush up the model. Second, although we established the prediction model of NT‐proBNP in a general Japanese population, we selected 10 explanatory variables for NT‐proBNP based on previous papers that were mainly reported in countries other than Japan because there were a few studies about BNP or NT‐proBNP in a general Japanese population. Third, the ratio of missing values was relatively high (833 missing in observed NT‐proBNP among 2505 patients). This study of the development of prediction models of NT‐proBNP was conducted as a part of the Toon Health Study, which had been initiated as a prospective population‐based cohort study for other purposes. Most of the missing values were from the early part of data in a first arm from 2009 to 2012 because of the lack of blood volume to measure it. For these reasons, we considered that the missing pattern was not missing at random, and we conducted multiple imputation. Fourth, the population is limited to Toon City, and a selection bias could occur in comparison with the ‘true’ general population. Toon City is a rural city situated in the Shikoku region with a different age distribution compared with large cities. However, this is the first prediction model of NT‐proBNP in the general population, and this could be the basis for further development of a more accurate prediction model. Finally, we could not confirm whether this study excluded participants having cardiac structural and functional abnormalities. In fact, the predicted value was not accurate after the observed value was higher than 400 pg/mL, which was a cut‐off point for the presence of heart failure, suggesting stage B heart failure with left ventricular hypertrophy, cardiomyopathy, and valvular heart diseases even in asymptomatic individuals. In this sense, the comparison of the observed and predicted values of NT‐proBNP might be helpful for the diagnosis of asymptomatic heart failure.

Conclusions

Our study introduces a robust prediction model for NT‐proBNP levels in the general population that addresses various demographic and clinical factors. Although the model shows acceptable performance, further refinement is required to enhance accuracy. The post hoc analysis, in which the OPR was introduced, offers a novel perspective on NT‐proBNP interpretation. Our transparent approach, coupled with the development of an interactive web‐based tool, sets the stage for future improvements. This study lays the foundation for ongoing efforts to enhance the predictive capacity of NT‐proBNP models.

Conflict of interest

Institute for Airway Disease received a software engineering fee from Department of Cardiology, Pulmonology, Hypertension and Nephrology, Ehime University Graduate School of Medicine.

Funding

Roche Diagnostics supported the research funding for this study. Roche Diagnostics also provided support for the test kits for immunoassay for the in vitro quantitative determination of NT‐proBNP (Elecsys proBNP II).

Supporting information

Figure S1. Missing data of NT‐proBNP and its predictive factors in this study. From the 2505 participants, 833 patients (33.3%) without enough stored blood to measure NT‐proBNP were judged as the missing outcome in this study. Alb, albumin; BMI, body mass index; DBP, diastolic blood pressure; eGFR, estimated glomerular filtration rate; HR, heart rate; HbA1c, haemoglobinA1c; T Chol, total cholesterol.

Figure S2. Visual inspection of the scatter plot matrix of NT‐proBNP and its predictive factors. There was no problematic collinearity among the variables. Alb, albumin; BMI, body mass index; Corr, correlation; DBP, diastolic blood pressure; eGFR, estimated glomerular filtration rate; HR, heart rate; HbA1c, haemoglobinA1c; T Chol, total cholesterol.

Inoue, K. , Yamamoto, K. , Higashi, H. , Takata, Y. , Inaba, S. , Miyazaki, S. , Higaki, A. , Saito, M. , Osawa, H. , and Yamaguchi, O. (2024) A prediction model for estimating NT‐proBNP in a general Japanese population: the Toon Health Study. ESC Heart Failure, 11: 1483–1492. 10.1002/ehf2.14691.