Predicting survival in heart failure: a risk score based on 39 372 patients from 30 studies (2024)

Article history

Abstract

See page 1391 for the editorial comment on this article (doi:10.1093/eurheartj/ehs363)

Introduction

Heart failure (HF) is a major cause of death, but prognosis in individual patients is highly variable. Quantifying a patient's survival prospects based on their overall risk profile will help identify those patients in need of more intensive monitoring and therapy, and also help target appropriate populations for trials of new therapies.

There exist previous risk models for patients with HF.^1–8 Each uses a single cohort of patients and hence their generalizability to other populations is questionable. Each model's development is from a limited cohort size, compromising the ability to truly quantify the best risk prediction model. Also most models are restricted to patients with reduced left-ventricular ejection fraction (EF), thus excluding many HF patients with preserved EF.

The Meta-analysis Global Group in Chronic Heart Failure (MAGGIC) provides a comprehensive opportunity to develop a prognostic model in HF patients, both with reduced and preserved EF. We use readily available risk factors based on 39 372 patients from 30 studies to provide a user-friendly score that readily quantifies individual patient mortality risk.

Methods

The MAGGIC program's details are documented previously.⁹ Briefly, we have individual patient data from 31 cohort studies (six randomized clinical trials and 24 observational registries). Here one registry is excluded since it had only median 3-month follow-up. The remainder comprised 39 372 patients with a median follow-up of 2.5 years (inter-quartile range 1.0–3.9 years), during which 15 851 patients (40.2%) died. Thirty-one baseline variables were considered as potential predictors of mortality (Table1).

The Coordinating Centre at the University of Auckland assembled the database for 29 studies. The London School of Hygiene and Tropical Medicine team the added in the CHARM trial data. The online Appendix lists the MAGGIC investigators (SupplementarySupplementary Data).

In 18 studies, a preference was for rounding the EF to the nearest 5%. In these studies, such rounded values were re-allocated within 2.5% either side using a uniform distribution.

Statistical methods

Poisson regression models were used to simultaneously relate baseline variables to the time to death from any cause, with study fitted as a random effect. Since mortality risk is higher early on, the underlying Poisson rate was set in three time bands: up to 3 months, 3–6 months, and over 6 months. Models were built using forward stepwise regression with inclusion criterion P < 0.01.

For binary and categorical variables, dummy variables were used. Quantitative variables were fitted as continuous measurements, unless there was a clear evidence of non-linearity, e.g. body mass index, EF, and creatinine. Also two highly significant statistical interactions were included in the main model: the impact of age and systolic blood pressure both depend on EF.

Each variable's strength of contribution to predicting mortality was expressed as the z statistic. The larger the z the smaller the P-value, e.g.: z values 3.29, 3.89, 5.32, and 6.11 are associated with P-values 0.001, 0.0001, 0.0000001, and 0.000000001, respectively.

Missing values are handled by multiple imputations using chained equations.^10,11 This method has three steps. First, for each variable with missing values, a regression equation is created. This model includes the outcome and follow-up time, in this case the Nelson–Aalen estimator (as recommended by White and Royston¹⁰), an indicator variable for each study and other model covariates. For continuous variables, this is a multivariable linear regression, for binary variables, a logistic regression, and for ordered categorical variables, an ordinal logistic regression. Once all such regression equations are defined, missing values are replaced by randomly chosen observed values of each variable in the first iteration. For subsequent iterations, missing values are replaced by a random draw from the distribution defined by the regression equations. This was repeated for 10 iterations, the final value being the chosen imputed value. This is similar to Gibbs sampling.¹²

This entire process was repeated 25 times, thus creating 25 imputed data sets. The next step was to estimate the model for each of these data sets. Finally, the model coefficients are averaged according to Rubin's rule.¹³ This ensures that the estimated standard error of each averaged coefficient reflects both between and within imputation variances, giving valid inferences.

We converted the Poisson model predictor to an integer score, which is then directly related to an individual's probability of dying within 3 years. A zero score represents a patient at lowest possible risk. Having grouped each variable into convenient intervals, the score increases by an integer amount for each risk factor level above the lowest risk. Each integer is a rounding of the exact coefficient in the Poisson model, making log rate ratio 0.1 equivalent to 1 point.

The data were analysed using Stata version 12.1 statistical package.

Results

This report is based on 39 372 patients from 30 studies: six were randomized controlled trials (24 041 patients) and 24 were registries (15 331 patients). Supplementary material online Table S1 describes each of the 30 studies. Overall, 15 851 (40.2%) patients died during a median follow-up of 2.5 years. The six largest studies (DIAMOND,¹⁴ DIG,¹⁵ CHARM,¹⁶ and ECHOS¹⁷ trials and IN-CHF¹⁸ and HOLA¹⁹ registries) contributed 75.8% of patients and also 75.8% of deaths.

There were 31 baseline variables available for inclusion in prognostic models. Table1 provides their descriptive statistics for patients still alive and patients who died during follow-up.

Using Poisson regression models for patient survival with forward stepwise variable selection, adjusting for study (random effect) and follow-up time (higher mortality rate in early follow-up), we identified 13 independent predictor variables (Table2). All were highly significant P < 0.002, and most were overwhelmingly significant, i.e. P < 0.0001.

Table3 lists the extent of missing data for these 13 variables. A multiple imputation algorithm (see Methods) was used to overcome this problem. Consequently, all results are based on average estimates across 25 imputed data sets.

For continuous variables, potential non-linearity in the prediction of survival was explored, as were potential statistical interactions between predictors. Hence the associations of EF, body mass index, and serum creatinine with mortality risk were, respectively, confined to EF <40%, body mass index <30 kg/m², and serum creatinine <350 µmol/L. The mortality association of increased age was more marked with higher EF, whereas the inverse association of systolic blood pressure with mortality became more marked with lower EF.

Figure1 displays the independent impact of each predictor on mortality risk. The impact of age (which varies with EF) is particularly strong, and hence is shown on a different scale to the other plots.

From the risk coefficients given in Table2, an integer score has been created (Figure2). For each patient, the integer amounts contributed by the risk factor's values are added up to obtain a total integer score for that patient. The bell-shaped distribution of this integer risk score for all 39 372 patients is shown in Figure3. The median is 23 points and the range is 0–52 points, with 95% of patients in the range of 8–36 points. The curve in Figure3 relates a patient's score to their probability of dying within 3 years. For instance, scores of 10, 20, 30, and 40 have 3-year probabilities 0.101, 0.256, 0.525, and 0.842, respectively. Table4 details the link between any integer score and the probabilities of dying within 1 year and 3 years.

Figure4 shows mortality over 3 years for patients classified into six risk groups. Groups 1–4 comprise patients with scores 0–16, 17–20, 21–24, and 25–28, respectively, approximately the first four quintiles of risk. To give more detail at higher risk, groups 5 and 6 comprise patients with scores 29–32 and 33 or more, approximately the top two deciles of risk. The marked continuous separation of the six Kaplan–Meier curves is striking: the 3-year % dead in the bottom quintile and top decile is 10 and 70%, respectively.

Regarding model goodness-of-fit, Figure5 compares observed and model-predicted 3-year mortality risk across the six risk groups. In the bottom two groups, the observed mortality is slightly lower than that predicted by the model, but overall the marked gradient in risk is well captured by the integer score.

Tables5 and 6 show two separate models for patients with reduced and preserved left-ventricular function (EF <40 and ≥40%, respectively). For most predictors, the strength of mortality association is similar in both subgroups. However, the impact of age is more marked and the impact of lower SBP is less marked in patients with preserved left-ventricular function, consistent with the interactions in the overall model.

In this meta-analysis of 30 cohort studies, we explored between-study heterogeneity in mortality prediction. From fitting separate models for each study, we observe a good consistency across studies re the relative importance of the predictors (data not shown). We have also repeated the model in Table2, now fitting study as a fixed effect (rather than a random effect). This reveals substantial between-study differences in mortality risk not explained by predictors in our model. However, a comparison of the seven randomized trials with the 23 patient registries reveals no significant difference in their mortality rates.

Discussion

This study identifies 13 independent predictors of mortality in HF. Although all have been previously identified, the model and risk score reported here are the most comprehensive and generalizable available in the literature. They are based on 39 372 patients from 30 studies with a median follow-up of 2.5 years, the largest available database of HF patients. Also, we include patients with both reduced and preserved EF, the latter being absent from most previous models of HF prognosis.

Given the wide variety of different studies included, with a global representation, the findings are inherently generalizable to a broad spectrum of current and future patients. Conversion of the risk model into a user-friendly integer score accessible by the website www.heartfailurerisk.org facilitates its use on a routine individual patient basis by busy clinicians and nurses.

All 13 predictors in the risk score should be routinely available, though provision will be made in the website for one or two variables to be unknown for an individual. Note, the ‘top five’ predictors age, EF, serum creatinine, New York Heart Association (NYHA) class, and diabetes are important to know. The inverse association of EF with mortality is well established, and as previously reported,⁹ in above 40% there appears no further trend in prognosis. We included serum creatinine rather than creatinine clearance or eGFR. The latter involve formulae that include age, which would artificially diminish the huge influence of age on prognosis.

We confirm the association of body mass index with mortality,²⁰ but with a cut-off of 30 kg/m², above which there appears no further trend. While others report heart rate as a significant predictor of mortality,²¹ we find that once the strong influence of beta blocker use is included, heart rate was not a strong independent predictor. A modest association of ACE-inhibitor and/or angiotensin-receptor blockers (ARB) use with lower mortality was highly significant, though many of our cohorts were established before ARBs were routinely available.

Cardiovascular disease history (e.g. myocardial infarction, angina, stroke, atrial fibrillation, LBBB) was considered in our model development. What mattered most was the time since first diagnosis of HF, best captured by whether this exceeds 18 months. Besides the powerful influence of diabetes, the other disease indicator of a poorer prognosis was prevalence of COPD. Previous myocardial infarction, atrial fibrillation, and LBBB were not sufficiently strong independent predictors of risk to be included in our model.

For patients with reduced and preserved EF, we developed separate risk models (Tables5 and 6). Nearly all predictors display a similar influence on mortality in both subgroups. Two exceptions are age (better prognosis of preserved EF compared with reduced EF HF is more pronounced at younger ages) and systolic blood pressure, which have a stronger inverse association with mortality in patients with reduced EF. These two interactions are incorporated into the integer risk score, as displayed in Figure1.

Our meta-analysis of 30 cohort studies enables exploration of between-study differences in mortality risk. Separately, for each of the 10 largest studies, we calculated Poisson regression models for the same 13 predictors. Informal inspection of models across studies shows a consistent pattern to be expected, given there are no surprises among the selected predictors.

An additional model, with study included as a fixed effect (rather than a random effect), reveals some between-study variation in mortality risk not captured by the predictor variables. This may be due to geographic variations or unidentified patient-selection criteria varying across registries and clinical trials, though overall patients in registries and trials appear at similar risk. Also, calendar year may be relevant since improved treatment of HF may enhance prognosis in more recent times. We will explore these issues in a subsequent publication.

The integer risk score gives a very powerful discrimination of patients' mortality risk over 3 years, and also has excellent goodness-of-fit to the data across all 30 studies combined (Figures3 and 4). Specifically, the score facilitates the identification of low-risk patients, e.g. score <17 has an expected 90% 3-year survival, and very high-risk patients, e.g. score ≥33 has an expected 30% 3-year survival.

We recognize some limitations. In combining evidence across multiple studies, we inevitably encountered substantial missing data (Table3), with a few variables (e.g. body mass index, HF duration) missing in some entire cohorts. To overcome this problem, we used sophisticated computer-intensive multiple imputation methods. In addition, we have checked the robustness of our overall findings for each predictor by separate analyses within each cohort where full data for that predictor were available.

Conventional good practice seeks to validate a new risk score on external data. That is important when a risk score arises from a single cohort in one particular setting, especially when that cohort has limited size. Here, the circ*mstances are different. We have a global meta-analysis of 30 cohorts with the largest numbers of patients and deaths ever investigated in HF. We found an internal consistency across studies in risk predictors, but inevitably found between-cohort differences in mortality risk not attributable to known risk factors, probably due to geographic variations and differing patient-selection criteria. Thus, no single external cohort can provide a sensible, generalizable validation of our risk model. We feel that internal validation found across studies is sufficient.

There exist several other risk scores for predicting survival in HF.^1–8 Best known is the Seattle Heart Failure Model.¹ It was developed from a small database, 1125 patients in the PRAISE clinical trial,²² confined to patients with severe HF: NYHA class III B or IV and EF ≤30%. Such patients account for <20% of patients in our meta-analysis. Thus the robustness, applicability, and generalizability of the Seattle model are somewhat limited. Some variables in the Seattle model, e.g. serum sodium and haemoglobin, were not found to be independent predictors for inclusion in our model. Also, the Seattle model does not include diabetes, body mass index, and serum creatinine, well established risk factors in HF. A recently developed predictive model for survival is from the 3C-HF Study,² but its relatively small size and only 1 year follow-up is limiting.

Any new risk score's success depends on the patient variables available for inclusion. Current knowledge of biomarkers in HF is inevitably ahead of what data are available across multiple cohort studies. For instance, natriuretic peptide level markedly influences prognosis in HF,^8,23 but could not be included in our model. In principle, its inclusion would enhance further the excellent prognostic discrimination we achieved with routinely collected long-established predictors. The risk score is most applicable for patients at a stable point in their disease, the short-term impact of acute HF events being a separate matter.

In conclusion, the risk score developed here on a huge database of 30 cohort studies provides a uniquely robust and generalizable tool to quantify individual patients' prognosis in HF. The simplified integer score, accessible by the website www.heartfailurerisk.org makes findings routinely usable by busy clinicians. Such immediate awareness of a patient's risk profile is of value in determining the most appropriate management and treatment of their HF.

Supplementary material

Supplementary material is available at European Heart Journal online.

Funding

K.K.P. is supported by the Heart Foundation of New Zealand. The work was supported by grants from the New Zealand National Heart Foundation, the University of Auckland, and the University of Glasgow. R.N.D. holds the New Zealand Heart Foundation Chair in Heart Health.

Conflict of interest: none declared.

References

Published on behalf of the European Society of Cardiology. All rights reserved. © The Author 2012. For permissions please email: journals.permissions@oup.com

Issue Section:

Clinical Research > Chronic heart failure

Comments

1 Comment

Email alerts

See also

Citing articles via

More from Oxford Academic

Having trouble contacting the network. Please try again in a moment or two.