Risk for Incident Heart Failure: A Subject‐Level Meta‐Analysis From the Heart “OMics” in AGEing (HOMAGE) Study

Abstract

Background

To address the need for personalized prevention, we conducted a subject‐level meta‐analysis within the framework of the Heart “OMics” in AGEing (HOMAGE) study to develop a risk prediction model for heart failure (HF) based on routinely available clinical measurements.

Methods and Results

Three studies with elderly persons (Health Aging and Body Composition [Health ABC], Valutazione della PREvalenza di DIsfunzione Cardiaca asinTOmatica e di scompenso cardiaco [PREDICTOR], and Prospective Study of Pravastatin in the Elderly at Risk [PROSPER]) were included to develop the HF risk function, while a fourth study (Anglo‐Scandinavian Cardiac Outcomes Trial [ASCOT]) was used as a validation cohort. Time‐to‐event analysis was conducted using the Cox proportional hazard model. Incident HF was defined as HF hospitalization. The Cox regression model was evaluated for its discriminatory performance (area under the receiver operating characteristic curve) and calibration (Grønnesby‐Borgan χ² statistic). During a follow‐up of 3.5 years, 470 of 10 236 elderly persons (mean age, 74.5 years; 51.3% women) developed HF. Higher age, BMI, systolic blood pressure, heart rate, serum creatinine, smoking, diabetes mellitus, history of coronary artery disease, and use of antihypertensive medication were associated with increased HF risk. The area under the receiver operating characteristic curve of the model was 0.71, with a good calibration (χ² 7.9, P=0.54). A web‐based calculator was developed to allow easy calculations of the HF risk.

Conclusions

Simple measurements allow reliable estimation of the short‐term HF risk in populations and patients. The risk model may aid in risk stratification and future HF prevention strategies.

Introduction

Heart failure (HF) is a complex syndrome often characterized by an abnormal cardiac structure and function leading to diminished quality of life, recurrent and costly hospital admissions, and premature death.1 HF remains a major public health concern because of the ageing of the populations and the increasing prevalence of HF‐associated comorbidities such as diabetes mellitus and obesity.2 The estimated prevalence of HF amounts to 1% to 2% of the adult population in developed countries, with the prevalence rising to >10% in those 70 years and older. The incidence is ≈5 to 10 per 1000 person‐years.3 Remarkable therapeutic advances in the past decades entailed a ≈3‐fold decrease in fatal HF, thereby increasing its prevalence and incidence. However, these advances are limited to HF with reduced ejection fraction, while its counterpart, HF with preserved ejection fraction, representing 50% of all HF cases, still lacks specific therapeutic strategies other than treating risk factors.4

Given the overall magnitude and lack of specific treatment for >50% of cases, identifying individuals at high risk for HF remains a priority in clinical research. Early identification will potentially allow testing preventive strategies in at‐risk populations thereby delaying or preventing HF onset and subsequently the disease burden. In this regard, a well‐calibrated, easily applicable, and generalizable risk model will help clinicians and trialists identify “high‐risk” patients. The former should concentrate on risk factor control with potential to delay HF onset, while the latter must focus on developing and testing novel therapeutic approaches.5 In this study, we aimed to identify routinely measured risk factors that reproducibly predict incident HF in an elderly population. We developed an easy‐to‐use calculator, allowing estimation of an individuals’ risk for incident HF in daily, routine clinical practice. The Heart “OMics” in AGEing (HOMAGE) study6 was used to conduct this subject‐level meta‐analysis and to develop an adequate scoring model.

Methods

Heart “OMics” in AGEing Database

The HOMAGE study database is constructed and maintained at the Studies Coordinating Centre in Leuven, Belgium. By sending anonymized data, the contributing partners confirmed that their study complies with good clinical practice (Helsinki Declaration), that all participants provided written informed consent, and that at the time of its conduct the study conformed to national regulations on clinical research in humans and on the protection of privacy. A detailed description of the database is available elsewhere.6 The database used for this analysis was locked on January 16, 2016.

Cohorts Analyzed

The entire HOMAGE database consists of 21 studies (n=46 134). Studies were eligible for inclusion in this analysis if information was available on fatal and/or nonfatal HF, with at least 20 events. We excluded 4 HF cohorts (n=4312), as well as 9 studies (n=7565) without information on HF outcome, and 4 studies with <20 HF events (n=4120). Among the 4 included studies (n=30 137), persons with a history of HF at baseline (n=223), those with missing information on incident HF (n=144), or those lost to follow‐up (n=279) were excluded. Figure 1 gives a detailed overview of the selection of the studies. Table 1 describes the 4 cohorts included in this analysis.

Figure 1.

Flow chart of studies and participants included in the analysis. HF indicates heart failure.

Table 1.

Description of Included Studies

	Study	Type of Patients	Age Range, y	Total No. (HF Events)	Median Follow‐Up (5th–95th Percentile)	Definition of HF
Study population	Health ABC	Population‐based (elderly persons)	70.0–80.0	3075 (574)	12.1 (2.0–14.1)	HF hospitalization: confirmation by reviewing hospital records
	PREDICTOR	Population‐based (elderly persons)	65.0–86.4	2001 (44)	3.9 (2.5–5.0)	HF hospitalization: main cause of hospital admission=ICD‐9‐CM code 428
	PROSPER	Patients at high risk of developing vascular disease	70.0–83.4	5804 (234)	3.3 (1.4–3.8)	HF hospitalization: confirmation by reviewing hospital records
External validation	ASCOT	Hypertensive patients	40.0–80.0	19 257 (292)	5.6 (3.5–6.6)	HF hospitalization and physician‐based diagnosis/HF death

ASCOT indicates Anglo‐Scandinavian Cardiac Outcomes Trial; Health ABC, Health Aging and Body Composition; HF, heart failure; ICD‐9‐CM, International Classification of Diseases, Ninth Revision, Clinical Modification; PREDICTOR, Valutazione della PREvalenza di DIsfunzione Cardiaca asinTOmatica e di scompenso cardiac; PROSPER, Prospective Study of Pravastatin in the Elderly at Risk.

Description of cohorts

The Health Aging and Body Composition (Health ABC) study is a cohort of 1584 women and 1491 men (aged ≥70 years), randomly recruited among Medicare beneficiaries residing in Pittsburg, Pennsylvania, and Memphis, Tennessee (April 1997–June 1998).7

The Valutazione della PREvalenza di DIsfunzione Cardiaca asinTOmatica e di scompenso cardiaco (PREDICTOR) trial is a population study of a random sample of 5940 people from the Regional Population Registry of 4 Italian cities in the Lazio region (June 2007–January 2010). A total of 2001 people (33.7% of those invited), older than 65 years, provided written informed consent and were referred to 8 cardiology centers for clinical examination, blood tests, electrocardiography, and echocardiography.8, 9

PROSPER (Prospective Study of Pravastatin in the Elderly at Risk) is a randomized, double‐blind, placebo‐controlled trial designed to test the hypothesis that treatment with pravastatin compared with placebo would reduce the risk of cardiovascular and cerebrovascular events in a cohort of 2000 women and 2804 men older than 70 years, with a history of vascular disease or at high risk for developing vascular disease. From 1997 until 1999, patients were enrolled via 3 coordinating centers in Scotland, Ireland, and The Netherlands.10, 11

Validation cohort

ASCOT (Anglo‐Scandinavian Cardiac Outcomes Trial) is a prospective, randomized, open, blinded end point trial, which recruited 19 257 high‐risk hypertensive patients aged 40 to 80 years, in Scandinavian countries, the United Kingdom, and Ireland (1998–2000).12, 13

Definition of Events

Table 1 gives an overview of the definition of events within each study. In all 3 studies, incident HF was defined as HF hospitalization. HF hospitalizations in the Health ABC study were confirmed by reviewing medical records of the participants.14 In the PREDICTOR study, HF hospitalizations were defined by International Classification of Diseases, Ninth Revision, Clinical Modification code 428 as the main cause of hospital admission. For the PROSPER trial, incident HF was defined as HF hospitalizations. Hospitalization records were examined and defined based on a combination of symptoms and signs, including chest radiography with fluid congestion or echocardiogram with severely diminished left ventricular function.15

Validation cohort

ASCOT defined HF events based on clinical signs and symptoms, chest x‐rays and echocardiography, or diagnosis of HF by the attending physician.

Statistical Analysis

For database management and statistical analyses, SAS software version 9.3 (SAS Institute, Cary, NC, USA) was used. Continuous variables were expressed as mean±SD and categorical variables as proportions. Missing values for BMI, systolic blood pressure, diastolic blood pressure, heart rate, glucose, cholesterol, and creatinine were imputated from the regression slope on age after stratification for cohort and sex. For comparison of means and proportions, the large sample Z‐test and chi‐square test were used, respectively. Statistical significance was a 2‐sided P value of <0.05.

Time‐to‐event analysis was conducted using the Cox proportional hazard model. Follow‐up time in the Health ABC study was limited to the maximum follow‐up time of the other 2 included cohorts, ie, 5.5 years. Log‐linearity was checked by testing the functional forms of the covariable by the Kolmogorov‐type supremum test. Covariables were entered into the multivariable model according to a stepwise selection procedure with P values for covariables to enter and stay in the model set to P=0.15 and P=0.05, respectively. Sex and age were forced into all models and other covariables considered to be of potential relevance and offered to the model were BMI, systolic and diastolic blood pressure, heart rate, blood glucose, serum total, high‐density lipoprotein and low‐density lipoprotein cholesterol, serum creatinine, smoking and drinking, use of antihypertensive medication, and history of diabetes mellitus, cardiovascular, coronary artery disease, peripheral artery, and cerebrovascular disease. In a sensitivity analysis, specific drug classes of antihypertensive treatment were added to the Cox models. Antihypertensive drug classes included β‐blockers, inhibitors of the renin‐angiotensin‐aldosterone system (RAAS), vasodilators, and diuretics. RAAS inhibitors included angiotensin‐converting enzyme inhibitors and angiotensin II type 1 receptor blockers, while vasodilators included calcium channel blockers and α‐blockers. All models were additionally adjusted for study.

Discrimination of the Cox regression models was assessed by calculating the area under the receiver operating characteristic curve (AUC), through a method taking into account time dependence and censoring, as previously described by Chambles and Diao.16 Assessment of the calibration, which demonstrates the goodness of fit of the models, was performed with the Grønnesby‐Borgan χ² statistic. This test measures the level of agreement between observed versus predicted number of events, which were visually plotted across deciles of the predicted risk. SAS macros developed in the ARIC (Atherosclerosis Risk in Communities) study17 were used for these calculations. To internally validate our models, 100 bootstrap samples were generated, sampling subjects with replacement. We adjusted our AUC obtained from the internal validation for optimism. The external validation of our model was done in the ASCOT cohort.

Risk calculator

The formula below was used to calculate the risk $(\widehat{\mathrm{p}})$:

$$\widehat{\mathrm{p}}=1-{\mathrm{S}}_0{(\mathrm{t})}^{\text{exp}\left(\left[({\sum }_{\mathrm{i}=1}^{\mathrm{q}}{\mathrm{\beta }}_{\mathrm{i}}{\mathrm{w}}_{\mathrm{i},\mathrm{ref}})\right]-({\sum }_{\mathrm{i}=1}^{\mathrm{p}}{\mathrm{\beta }}_{\mathrm{i}}{\overline{\mathrm{X}}}_{\mathrm{i}})\right)}$$

where S₀(t) is average survival at follow‐up time t (t=5 years), q is the number of continuous risk factors, p is the total number of risk factors (continuous and categorical), β_i is the regression coefficient of the i‐th risk factor as estimated by the Cox regression model, w_i,ref is the reference value of the base category of the i‐th risk factor, $\overline{{\mathrm{X}}_{\mathrm{i}}}$ is the mean of the i‐th risk factor. We implemented this formula in Excel to develop an easy‐to‐use risk calculator (Figure S1).

Results

Study Population

A study population of 10 236 elderly people (51.3% women) with a mean age±SD of 74.5±3.7 years were included. Mean BMI±SD was 26.9±4.4 kg/m² (Table 2). Except for the distribution of smoking and drinking status and diastolic blood pressure, all variables listed in Table 2 differed significantly between persons with incident HF and those without.

Table 2.

Baseline Characteristics by Incident HF

	All	HF	No HF
No. of patients	10 236	470	9766
Mean of characteristics±SD
Age, y	74.5±3.7	75.3±3.7	74.5±3.6§
BMI, kg/m2	26.9±4.4	27.7±4.7	26.8±4.3§
Systolic blood pressure, mm Hg	146.9±22.7	149.5±24.9	146.8±22.5‡
Diastolic blood pressure, mm Hg	79.9±12.3	79.1±14.1	80.0±12.2
Heart rate, beats per min	66.7±11.5	68.9±12.5	66.6±11.5§
Blood glucose, mmol/L	5.60±1.60	5.92±2.13	5.58±1.57§
Cholesterol, mmol/L	5.50±0.97	5.37±0.98	5.51±0.97†
LDL cholesterol, mmol/L	3.54±0.89	3.44±0.90	3.54±0.89*
HDL cholesterol, mmol/L	1.33±0.40	1.29±0.41	1.33±0.38*
Serum creatinine, μmol/L	96.2±27.7	106.6±42.0	95.8±26.7§
No. with characteristics, %
Women	5248 (51.3)	212 (45.1)	5036 (51.6)†
Smokers	2052 (20.1)	100 (21.3)	1952 (20.0)
Consume alcohol	5592 (54.6)	237 (50.4)	5355 (54.8)
Hypertension	6577 (64.3)	334 (71.1)	6243 (63.9)†
Diabetes mellitus	1299 (12.7)	90 (19.2)	1209 (12.4)§
History of cardiovascular disease	3701 (36.2)	254 (54.0)	3447 (35.3)§
History of coronary artery disease	2606 (25.5)	209 (44.5)	2397 (24.5)§
History of peripheral artery disease	257 (2.5)	25 (5.3)	232 (2.4)§
History of cerebrovascular incident	856 (8.4)	55 (11.7)	801 (8.2)†
Use of antihypertensive medication	6596 (64.4)	355 (75.5)	6241 (63.9)§
Use of ACEIs	1697 (16.6)	121 (25.7)	1576 (16.1)§
Use of calcium channel blockers	2345 (22.9)	159 (33.8)	2186 (22.4)
Use of diuretics	3372 (32.9)	192 (40.9)	3180 (32.6)‡
Use of β‐blockers	2087 (20.4)	94 (20)	1993 (20.4)
Use of angiotensin receptor blockers	502 (4.9)	26 (5.5)	476 (4.9)

The study population includes data from the Health ABC (Health Aging and Body Composition), PREDICTOR (Valutazione della PREvalenza di DIsfunzione Cardiaca asinTOmatica e di scompenso cardiac), and PROSPER (Prospective Study of Pravastatin in the Elderly at Risk) studies. Data are reported as mean±SD or number (percentage). ACEIs indicates angiotensin‐converting enzyme inhibitors; HDL, high‐density lipoprotein; LDL, low‐density lipoprotein.

Significance for the comparison between patients developing heart failure (HF) and those not developing HF: *P≤0.05; ^† P≤0.01; ^‡ P≤0.001; ^§ P≤0.0001.

After a median follow‐up time of 3.5 years (5th to 95th percentile: 2.3–5.5 years), 470 incident HF cases were present. Overall sex‐ and age‐adjusted incidence of HF was 12.1 (95% CI, 11.0–13.2) per 1000 person‐years. In persons younger than 75 years, the sex‐adjusted incidence rate was 9.8 (95% CI, 8.6–11.1) per 1000 person‐years, while the sex‐adjusted incidence rate for HF in persons older than 75 years was 15.3 (95% CI, 13.5–17.3) per 1000 person‐years. The sex‐adjusted cumulative incidence per age group is shown in Figure 2.

Figure 2.

Sex‐adjusted cumulative heart failure (HF) incidence per age group.

The Cox regression analysis, corrected for study, showed that higher age, BMI, smoking, diabetes mellitus, history of coronary artery disease, use of antihypertensive medication, higher systolic blood pressure, heart rate, and serum creatinine were significantly associated with an increased risk of HF onset. Hazard ratios for a 1‐unit increase in risk factor are given in Table 3.

Table 3.

Hazard Ratios Relating HF Outcome to Risk Factors

Study Population (Health ABC, PREDICTOR, PROSPER)
Risk Factor	Hazard Ratio	95% CI	P Value
Female	0.86	0.71–1.04	0.12
Age, per y	1.08	1.05–1.11	<0.0001
BMI, per kg/m2	1.03	1.01–1.06	0.002
Smoking (0,1)	1.84	1.46–2.32	<0.0001
Diabetes mellitus (0,1)	1.41	1.12–1.79	0.004
History of coronary artery disease (0,1)	2.49	2.06–3.01	<0.0001
Use of antihypertensive medication (0,1)	1.65	1.33–2.06	<0.0001
Systolic blood pressure, per mm Hg	1.009	1.005–1.013	<0.0001
Heart rate, per beats per min	1.02	1.01–1.03	<0.0001
Serum creatinine, per μmol/L	1.005	1.004–1.006	<0.0001

Hazard ratios are expressed per 1‐unit increase in continuous risk factors. All analyses are corrected for study. Health ABC indicates Health Aging and Body Composition; HF, heart failure; PREDICTOR, Valutazione della PREvalenza di DIsfunzione Cardiaca asinTOmatica e di scompenso cardiac; PROSPER, Prospective Study of Pravastatin in the Elderly at Risk.

Model Assessment

The AUC for discriminating HF risk based on our Cox model was 0.706. After bootstrapping, we corrected the AUC for overoptimism and obtained an AUC of 0.702. Observed versus expected numbers of HF according to deciles of the 5‐year risk show good agreement (Grønnesby‐Borgan χ² 7.9; P=0.54). Figure 3 shows the calibration of the Cox model by plotting deciles of HF risk.

Figure 3.

Calibration is shown. Observed (open bar) vs expected (solid bar) number of heart failure events across deciles of 5‐year heart failure risk.

We assessed calibration among the individual studies. Grønnesby‐Borgan χ² statistics for PREDICTOR, Health ABC, and PROSPER were 2.9 (P=0.97), 5.1 (P=0.83), and 16.2 (P=0.06), respectively. Figure 4 shows the observed versus expected number of events for each individual study.

Figure 4.

Calibration per individual study. Observed (open bar) vs expected (solid bar) number of heart failure events across deciles of 5‐year risk for the Health ABC (Health Aging and Body Composition), PREDICTOR (Valutazione della PREvalenza di DIsfunzione Cardiaca asinTOmatica e di scompenso cardiac), and PROSPER (Prospective Study of Pravastatin in the Elderly at Risk) studies, which were used to develop the heart failure risk function. ASCOT (Anglo‐Scandinavian Cardiac Outcomes Trial) was used as an external validation cohort.

The risk function was externally validated using data from ASCOT. In 13 121 participants (mean age 63.3 years, 23.0% women), 191 HF events (HF hospitalization or death) occurred over a median follow‐up of 5.6 years (5th–95th percentile: 3.4–6.6). The AUC was 0.692 when we applied our risk function to this cohort.

Sensitivity Analyses

We examined the effect of antihypertensive treatment on incident HF in more detail by using specific classes of antihypertensive treatment: β‐blockers, RAAS inhibitors (including angiotensin‐converting enzyme inhibitors and angiotensin II type 1 receptor blockers), vasodilators (calcium channel blockers and α‐blockers), and diuretics. Results (Table 4) show that RAAS inhibitors, diuretics, and vasodilators were significantly associated with risk of incident HF, while use of β‐blockers were not.

Table 4.

Hazard Ratios Relating HF Outcome to Risk Factors, Including Classes of Antihypertensive Treatment

Study Population (Health ABC, PREDICTOR, PROSPER)
Risk Factor	Hazard Ratio	95% CI	P Value
Female	0.87	0.71–1.05	0.15
Age, per y	1.08	1.05–1.11	<0.0001
BMI, kg/m2	1.03	1.01–1.05	0.007
Smoking (0,1)	1.84	1.46–2.32	<0.0001
Diabetes mellitus (0,1)	1.36	1.07–1.72	0.012
History of coronary artery disease (0,1)	2.52	2.08–3.06	<0.0001
Use of β‐blockers (0,1)	1.03	0.81–1.30	0.84
Use of RAAS inhibitors (0,1)	1.55	1.27–1.90	<0.0001
Use of vasodilators (0,1)	1.35	1.13–1.67	0.003
Use of diuretics (0,1)	1.37	1.13–1.67	0.002
Systolic blood pressure, per mm Hg	1.009	1.005–1.013	<0.0001
Heart rate, per beats per min	1.02	1.01–1.03	<0.0001
Serum creatinine, per μmol/L	1.005	1.004–1.006	<0.0001

Hazard ratios are expressed per 1‐unit increase in continuous risk factors. All analyses are corrected for study; the analysis in the patient cohort was further corrected for randomization group. Health ABC indicates Health Aging and Body Composition; HF, heart failure; PREDICTOR, Valutazione della PREvalenza di DIsfunzione Cardiaca asinTOmatica e di scompenso cardiac; PROSPER, Prospective Study of Pravastatin in the Elderly at Risk; RAAS, renin‐angiotensin‐aldosterone system.

Heart Failure Risk Calculator

Figure S1 is a calculator to compute a person's HF risk. Using the calculator, we obtain an HF risk of 2.58% for a 71‐year‐old woman, with a BMI of 26.4 kg/m², nonsmoker, with diabetes mellitus, no history of coronary artery disease, who needs to use antihypertensive medication, with a systolic blood pressure of 150 mm Hg, heart rate of 62 beats per minute, and serum creatinine of 151.2 μmol/L. The calculation is done as described below:

$$\begin{array}{cc}\hfill ({\mathrm{\beta }}_1& \times \text{age}+{\mathrm{\beta }}_2\times \text{sex}+{\mathrm{\beta }}_3\times \text{body mass index}+{\mathrm{\beta }}_4\hfill \\ \hfill & \times \text{smoking}+{\mathrm{\beta }}_5\times \text{diabetes}+{\mathrm{\beta }}_6\hfill \\ \hfill & \times \text{history of coronary artery disease}+{\mathrm{\beta }}_7\hfill \\ \hfill & \times \text{antihypertensive treatment}+{\mathrm{\beta }}_8\hfill \\ \hfill & \times \text{systolic blood pressure}+{\mathrm{\beta }}_9\times \text{heart rate}+{\mathrm{\beta }}_{10}\hfill \\ \hfill & \times \text{serum creatinine})\hfill \end{array}$$

Value of the risk factor of the individual

$$\begin{array}{cc}\hfill \sum _{\mathrm{i}=1}^{\mathrm{p}}{\mathrm{\beta }}_{\mathrm{i}}{\mathrm{X}}_{\mathrm{i}}=& 0.0748\times (71)-0.149\times (1)+0.0329\times (26.4)\hfill \\ & +0.61\times (0)+0.346(1)+0.914\times (0)+0.503\hfill \\ & \times (1)+0.0089\times (150)+0.0203\times (62)\hfill \\ & +0.0050\times (151.2)=10.2242\hfill \end{array}$$

Mean value of the risk factors

$$\begin{array}{cc}\hfill +\sum _{\mathrm{i}=1}^{\mathrm{p}}{\mathrm{\beta }}_{\mathrm{i}}{\overline{\mathrm{X}}}_{\mathrm{i}}=& 0.050\times (73.4)-0.019\times (0.50)+0.66\hfill \\ & \times (0.11)+1.03\times (0.17)+0.53\times (0.52)\hfill \\ & +0.018\times (137.0)+0.024\times (67.1)\hfill \\ & +0.078\times (5.79)+0.0044\times (90.2)\hfill \\ & =10.2419\hfill \end{array}$$

5‐year estimated survival=S₀(t)= 0.9738

$$\mathrm{p}=1-{\mathrm{S}}_0{(\mathrm{t})}^{\text{exp}({\sum }_{\mathrm{i}=1}^{\mathrm{p}}{\mathrm{\beta }}_{\mathrm{i}}{\mathrm{X}}_{\mathrm{i}}-{\sum }_{\mathrm{i}=1}^{\mathrm{p}}{\mathrm{\beta }}_{\mathrm{i}}{\overline{\mathrm{X}}}_{\mathrm{i}})}=1-0.{9738}^{\text{exp}(10.2242-10.2419)}=0.0258=2.58\%$$

Discussion

In this study, we developed a multivariable risk prediction tool to be used in daily clinical practice. The developed risk calculator estimates the risk of new‐onset HF in older people. Our stepwise approach identified several routinely available risk factors such as smoking, blood pressure, heart rate, diabetes mellitus, and creatinine levels, as well as nonmodifiable factors such as sex and age.

Risk Factors

The risk factors identified were in line with expectations and have been previously shown to be associated with incident HF.18, 19, 20 Older individuals had a higher risk for incident HF, explained by the age‐related cardiac structural and functional changes. The Framingham Study showed that per decade, the risk of HF increased by ≈24% in women and by 37% in men.21 Several studies examined the effect of smoking in older adults and found a higher risk for HF in both current and former heavy cigarette smokers.22, 23 Elevated systolic blood pressure is associated with increased risk of cardiovascular events and death in the general population.24 In older people and high‐risk patients, higher systolic blood pressure was a consistent risk factor. Resting heart rate is another variable with potential predictive capacity in HF and cardiovascular disease development. A recent study showed that resting heart rate was independently associated with incident HF in men during follow‐up and this association was found to be significant in men with both single and repeated heart rate measurement over time.25 These findings are replicated in our analyses as well. We also showed that diabetes mellitus were associated with increased HF risk. Hyperglycemia and insulin resistance are known to have unfavorable effects on the heart by increasing left ventricular mass index and wall thickness and promoting myocardial fibrosis and left ventricular dysfunction, which may be caused by increased oxidative stress and cardiac autonomic neuropathy.26, 27 In fact, incident HF was among the most frequent events among other cardiovascular events in recently published diabetes cardiovascular outcome trials on new oral hypoglycemic drugs in patients with type 2 diabetes mellitus.28, 29, 30 In addition, in a subgroup of the Health ABC study, in persons without known diabetes mellitus, glucose was a strong predictor of incident HF.31 Subsequently, serum creatinine was identified as a risk factor for incident HF. Creatinine is widely used to measure kidney function, and kidney damage has been associated with increased HF risk.32 Our results further reveal that persons who are taking antihypertensive treatment are at higher risk for incident HF. Previous studies showed that treatment with antihypertensive medication was associated with an increased residual cardiovascular risk.33, 34, 35, 36 It seems that the cardiovascular risk associated with hypertension cannot be totally diminished by intake of medication. Probably, patients taking hypertensive treatment at entry already had a higher cardiovascular risk profile, independent of blood pressure control. Looking into specific classes of antihypertensive treatment, we identified vasodilators, diuretics, and RAAS inhibitors as being associated with higher HF risk. This counterintuitive observation might be explained by reverse epidemiology: persons who need to take these drugs are most likely the ones who are already at higher risk for HF. Indeed, persons with symptoms of angina pectoris or previous coronary heart disease, who already have an increased HF risk, are often under treatment with vasodilators.37 Similarly, persons with severe hypertension, treated with diuretics, are at higher risk. Moreover, physicians might prescribe these medications to patients with suspected HF, or with mild symptoms, before the diagnosis of overt HF is made.

Prediction Models

Rather than assessing single risk factors, risk prediction models allow the identification of a broader segment of the population. Previously, risk models have been developed to predict HF risk, including mostly US and European cohorts.19 Most HF prediction models were described as mathematical equations, with only 2 US studies providing a point‐based risk score.14, 21 We provide an easy‐to‐use risk calculator in Excel to allow simple risk calculation for clinicians and facilitate its use in daily practice. Several HF prediction models assessed the discriminatory ability of B‐type natriuretic peptide or N‐terminal pro–B‐type natriuretic peptide on top of traditional risk factors.38, 39, 40, 41, 42 Overall, a modest improvement in discrimination was observed.40, 41, 42, 43 This does not mean that biomarkers could not be useful in further refining the identification of at‐risk patients. Adding biomarker information should be the next step to further categorize persons initially identified at high risk, based on our available clinical risk parameters. The HOMAGE consortium remains dedicated to discovering biomarkers that could lead to new insights and improve risk stratification, mechanistic profiling, and early detection of HF. Data of the Health ABC study were previously used to develop an HF risk score.14 We pooled the Health ABC data with data from the European PREDICTOR and PROSPER studies, which also include elderly people. Inclusion of both US and European persons adds to the generalizability of the developed risk score in elderly persons.

Limitations

Limitations of our study include the inability to systematically distinguish between participants with incident HF according to the type of left ventricular dysfunction (systolic versus diastolic). Recently, a large study including 4 population‐based cohorts identified differences in risk factors predicting HF with preserved ejection fraction and HF with reduced ejection fraction.44 Although age, BMI, systolic blood pressure, previous myocardial infarction, and antihypertensive treatment were identified as risk factors for both HF subtypes, male sex, smoking, left ventricular hypertrophy, left bundle branch block, and diabetes mellitus were found to only predict HF with reduced ejection fraction.20

Next, the diagnosis of HF is less clear and specific then, for example, myocardial infarction. The HF end point in our cohort was based on nonfatal HF hospitalization, which implicates that HF diagnosed at an outpatient clinic or by general practitioners was not included. Moreover, differences in HF hospitalization definition between the cohorts occur due to factors such as hospital admissions policy, treating physicians, adherence to guidelines for therapy, and support by dedicated HF nurses.

Finally, our risk score was developed in elderly patients at risk for vascular disease. Particularly in this population, aggravating factors related to rehospitalizations such as intercurrent infections, atrial fibrillation, anemia, or hyperthyroidism could play a role. We did not take into account these factors, since we focused on first hospitalization for HF and aimed to provide a risk score based on readily available clinical risk factors, to facilitate its use in daily practice.

Strengths

The strengths of our study are that we included persons from well‐characterized populations, with an adequate follow‐up period. Risk factors included are routinely available and our estimates are derived from a large sample. The risk function calculator is easy to use by clinicians or nurses and could be useful as a tool in risk factor modification screening within a public health or clinical setting. We assessed the validity of our HF risk model in the ASCOT cohort, showing good discrimination and calibration.

According to a recent systematic review, there are no recommendations in current guidelines for using risk prediction models for HF or studies assessing the effect of such models in routine clinical practice.19 Research into a practical application of our risk prediction scores remains necessary to give a definitive answer to whether improvement of outcome can be expected by using the risk tool. Indeed, for example, collaborative care in patients at risk for developing HF based on B‐type natriuretic peptide screening reduced the combined rates of left ventricular systolic dysfunction, diastolic dysfunction, and HF.45

Conclusions

An increase in HF over the next decades will translate to more than a 200% increase in direct medical costs of care.46 To mitigate mortality and healthcare costs, effective prevention of HF that can offer early correction of predisposing conditions and risk factors in susceptible persons is required. Our HF risk model allows reliable identification of high‐risk populations, contributing to a more targeted and cost‐effective implementation of preventive measures for HF.

Appendix

We would like to thank all contributing investigators.

HOMAGE‐Investigators

Inserm—Alexandre Mebazza, Florence Pinet, Anne Pizard, Philippe Rouet, Faiez Zannad; IT—Catherine Clusel; EDDH—Stephanie Grosjean; ACS Biomarker—Heico Breek, Joost Leenders; FIMA—Javier Diez; UCD—Kenneth McDonald; HULL—Andrew Clark; UM—Stephane Heymans; IRFMN—Roberto Latini, Serge Masson; MHH—Thomas Thum; KUL—Nicholas Cauwenberghs, Ljupcho Efremov, Lutgarde Thijs, Lotte Jacobs, Tatiana Kuznetsova, Augustine Odili, Jan A. Staessen, Fang‐Fei Wei, Wen‐Yi Yang, Zhen‐Yu Zhang; LSHTM—Tim Collier; UG—Christian Delles, Naveed Sattar; MOS—Harald Mischak; Stony Brook—Javed Butler; KU—Mamas Mamas; TATAA—Jens Bjorkman; Charité—Burkert Pieske; for Health ABC—Stephen Kritchevsky, Anne Newman, Lampros Papadimitrious; for PREDICTOR—Alessandro Boccanelli, Marina Davoli, Gian Francesco Mureddu; for PROSPER—Brendan Buckley, Ian Ford, Wouter Jukema, David J. Stott; and for ASCOT—Bjorn Dahlof, Neil Poulter, Peter Sever.

Sources of Funding

This work is supported by the European Union: HEALTH‐F7‐305507 HOMAGE (EU FP7 305507 http://www.homage-hf.eu). The European Research Council Advanced Researcher Grant‐2011‐294713‐EPLORE and the Fonds voor Wetenschappelijk Onderzoek Vlaanderen, Ministry of the Flemish Community, Brussels, Belgium (G.0881.13 and G.088013), currently support the Studies Coordinating Centre in Leuven.

Disclosures

None.

Supporting information

Figure S1. Heart failure risk calculator.

(J Am Heart Assoc. 2017;6:e005231 DOI: 10.1161/JAHA.116.005231.)28465299