Wayne Levy, MD 

Medscape Cardiology.  2006;10(2) ?2006 Medscape

Introduction

It is well known that cardiovascular disease (CVD) is the leading cause of death in the developed world and is rapidly becoming the leading cause worldwide because it will increase in low- and middle-income countries.^[1] The past 10-15 years saw many advances in the recognition and treatment of the various manifestations of CVD, ranging from hypertension to ischemic heart disease and cardiac arrhythmias. The one aspect of this overall disease state that has failed to show significant improvement on a population basis is heart failure (HF).^[2] Of equal importance, the incidence and the prevalence of HF have not decreased in more than 2 decades.^[3] In the United States alone, approximately 550,000 new HF cases are diagnosed each year,^[3] and almost 50% of patients discharged following HF hospitalization are rehospitalized within 6 months.^[4]

For individual HF patients and the clinicians who see them, however, these statistics must be counted as somewhat irrelevant. While the recent advances in medical science mean that the prognosis for the HF patient is no longer as poor as it once was, HF is still a condition from which the patient is not going to "recover." As a result, the single most important bit of information for almost all patients is knowing whether they have 1, 5, or 10 years to live.

Ideally, in order to answer that question, clinicians should have a way to enter a given set of parameters -- eg, clinical measures and actual or potential therapeutic regimens -- into a validated computation method and derive an answer that has been calculated for that individual patient, or at least a patient with the described characteristics. In other words, given the patient's diagnosis and the course of treatment prescribed, the result should be "X" -- and the answer needs to be in a format that patients and clinicians can understand. For our model, we focused on calculating an estimate of annual mortality and "years of expected survival."

There are several simple markers that can be used to estimate mortality risk in HF patients, including:

Existing models used to predict the risk of death or the need for urgent transplantation in HF all have features that may limit their applicability.^[5-10] The validity of these models, which is based on HF population data indicating whether the HF treatment has changed, has been poor, and most current models have not allowed the user to alter the estimate of mortality by adding (or subtracting) particular courses of HF therapy.

It is in this context that my colleagues and I recently published the Seattle Heart Failure Model (SHFM).^[11] Our purpose was to develop and validate a multivariate risk estimation model to calculate survival of HF patients that incorporates easily obtainable clinical and laboratory variables, plus applications of HF medications and/or devices.

Methods and Definitions

To create and validate the SHFM, we accessed the patient data from 6 clinical trials enrolling predominantly patients with left ventricular systolic HF. To develop the model, we used data from the Prospective Randomized Amlodipine Survival Evaluation (PRAISE-1)^[12] trial, and to validate the model we incorporated data from 5 other trials: Evaluation of Losartan in the Elderly II (ELITE-2),^[13] Valsartan Heart Failure Trial (Val-HeFT),^[14] University of Washington (UW),^[15,16] Randomized Etanercept North American Strategy to Study Antagonism of Cytokines (RENAISSANCE),^[17] and Italian Heart Failure Registry (IN-CHF),^[18] using the Cox proportional hazards model.

Model Development

To develop a prediction model, we had to identify which variables in the trial proved to be statistically significant and then derive the contribution of each to the final aggregate hazard ratio.

The hazard ratios for some of the most important HF medications and devices could not be effectively estimated on the basis of PRAISE-1 because of either widespread or rare use; these included angiotensin-converting enzyme (ACE) inhibitors, beta-blockers, angiotensin receptor blockers, aldosterone blockers, implantable cardioverter defibrillators (ICDs), biventricular pacing, and left ventricular assist devices . To overcome this apparent limitation, the benefit of adding these medications/devices to a patient's regimen was estimated by assessing outcomes from large, published randomized trials or meta-analyses, and for patients already on the medication/device, the impact of medication or device use on systolic blood pressure , NYHA class, and EF was similarly accounted for on the basis of their effects in large published trials ( Table 1 ). We took this approach to ensure that these hazard ratios were estimated on the basis of data from randomized clinical trials other than the validation trials' data.

Once we had created our model, we then used the patient data from the 5 other trial cohorts (N = 9942) to prospectively validate the model.

Patient Characteristics and Events

The SHFM score derived in PRAISE-1 was prospectively applied to each patient in the ELITE-2, UW, Val-HeFT, RENAISSANCE, and IN-CHF trials to provide individual estimates of survival at 1, 2, 3, and 5 years for each trial. Accuracy of the model across datasets was determined by comparing the mean 1-, 2-, and 3-year predicted survival vs actual survival. The discriminant ability of the SHFM was determined by the 1-year receiver operating characteristic (ROC) area under the curve for each dataset and for all datasets combined.

Results

In multivariate analyses of the PRAISE-1 data, the following patient data were associated with increased mortality:

On the basis of the multivariate analyses, renal function, body mass index, WBC count, heart rate, and diabetes were not considered independent predictors of survival.

The resulting calculated daily diuretic dose per kilogram of body weight was the most powerful multivariate predictor of mortality (x²=77), with an 18% increase in risk of death for every 80 mg of furosemide. Therefore, a patient with NYHA class II HF who required 240 mg/day of furosemide had a risk of mortality similar to that of an NYHA class III HF patient who did not require diuretics.

Application of the Model

When application of the model to the PRAISE-1 data was evaluated, the overall "fit" of the model was highly significant and quite accurate. The 1- and 2-year survival rates predicted by the model for the entire cohort were 73.4% and 56.7%, respectively, vs actual survival rates of 74.3% and 56.0%, and the correlation between predicted and actual survival by deciles was 0.97 (SEE ?5%).

The Seattle HF model was then applied prospectively to the datasets from the ELITE-2, Val-HeFT, UW, RENAISSANCE, and IN-CHF trials. The predicted vs actual survival percentages and their correlations are shown in Table 2 .

The 6 datasets were then combined into 1 dataset ( Table 3 ).

The overall fit by deciles was very good (r = 0.98, SEE = ? 3), but with a slight underestimation of mortality, which was most evident at 2 and 3 years for lower-risk patients. The 1-year ROC was 0.73.

A unique feature of the model is its ability to estimate how the addition of medications/devices may alter mortality. For example, triple drug therapy (ACE inhibitor, beta blocker, and aldosterone blocker) will reduce predicted annual mortality in a patient with NYHA class III HF by ~65%, from ~20% to 8%. However, most patients and clinicians will have difficulty understanding what that means for an individual patient. Using the model results in the answer that triple drug therapy would add ~4 years of life (predicting ~8 vs ~4 years, on vs off therapy). As another example, an ICD is reported to decrease annual mortality from ~8% to ~6%, whereas the SHFM translates an ICD as conferring an additional 1.6 years of life.

Patients and clinicians alike can easily understand that if you apply these therapies, you will live on average 4 (or 1.6) years longer, whereas a change in "risk of death" from 20% to 8% is much more difficult to comprehend.

Moreover, the SHFM may be even more important in low-risk patients. For example, adding a beta-blocker to the regimen of an HF patient with a predicted annual mortality of 4% would decrease the annual mortality to ~2.6% per year. However, the SHFM reveals that this would mean adding 2.9 years of life (12 years vs 14.9 years). It is easy for clinicians -- and even patients -- to think that a ~1.4% decrease in annual mortality is not really worth trying to achieve, but ~3 years of life is harder to ignore.

The figure below illustrates a calculation shown on the Web-based version of the SHFM.

Figure 1. 

As can be seen, implanting a CRT-D in a patient with NYHA class III HF will add ~2 years of life (6.2 vs 8.1 years). However, in addition, the interactive model is designed to allow selection of devices only if the patient meets current CMS criteria for the device. Another positive feature of the model is that it may allow identification of high-risk HF patients who would benefit from LVAD therapy, prior to their becoming inotrope dependent or developing cardiac cachexia.
BiV-ICD = biventricular pacemaker-implantable cardioverter defibrillator; CMS = Centers for Medicare and Medicaid Services; LVAD = left ventricular assist device

     

It is quite likely that adding more variables, eg, BNP, will improve the accuracy of the model. (A challenge with including BNP is the difficulty in obtaining access to HF databases with BNP levels measured using commercially available assays.) Similarly, peak oxygen consumption may add prognostic significance, and certainly is relevant for cardiac transplant programs that are trying to determine the appropriate time for listing for transplantation.

Conclusion and Further Refinements

The SHFM provides an estimate of 1- to 5-year expected survival for HF patients with predominantly systolic dysfunction, using simple clinical variables that are readily available to all clinicians.

The calculation of estimated survival includes 14 continuous variables and 10 categorical values, which makes it impractical for computation by hand. However, both Web-based and Palm calculator versions have been developed, providing clinicians with a convenient interactive method of calculating estimated survival and predicting the effects on survival of adding or subtracting medications or devices to or from a patient's regimen. This information can be accessed at http://www.seattleheartfailuremodel.org/.