Clinical Factors Related to Morbidity and Mortality in High Risk Heart Failure Patients: The GUIDE-IT Predictive Model and Risk Score

Abstract

Background

Most heart failure (HF) risk scores have been derived from cohorts of stable HF patients and may not incorporate up to date treatment regimens or deep phenotype characterization that change baseline risk over the short term and long term follow up period. We undertook the current analysis of participants in the GUIDE IT (Guiding Evidence-Based Therapy Using Biomarker Intensified Treatment) Trial to address these limitations.

Methods

The GUIDE IT study randomized 894 high risk patients with HFrEF (EF ≤ 40%) to biomarker-guided treatment strategy vs usual care. We performed risk modeling using Cox proportional hazards models and analyzed the relationship between 35 baseline clinical factors and the primary composite endpoint of cardiovascular (CV) death or HF hospitalization, the secondary endpoint of all-cause mortality, and the exploratory endpoint of 90 day HF hospitalization or death. Prognostic relationships for continuous variables were examined and key predictors were identified using a backward variable selection process. Predictive models and risk scores were developed.

Results

Over a median follow-up of 15 months, the cumulative number of HF hospitalizations and CV deaths was 328 out of 894 patients (Kaplan-Meier (KM) event rate 34.5% at 12 months). Frequency of all-cause deaths was 143 out of 894 (KM event rate 12.1% at 12 months). Outcomes for the primary and secondary endpoints between strategy arms of the study were similar. The most important predictor that was present in all three models was the baseline natriuretic peptide level. Hispanic ethnicity, low sodium and high heart rate were present in 2 of the 3 models. Other important predictors included the presence or absence of a device, NYHA class, HF duration, black race, co-morbidities (sleep apnea, elevated creatinine, ischemic heart disease), low blood pressure and a high congestion score. Risk models using readily available clinical information are able to accurately predict short term and long term cardiovascular events and may be useful in optimizing care and enriching patients for clinical trials (Clinicaltrials.gov NCT01685840).

Introduction

High-risk heart failure (HF) patients continue to pose a major burden on healthcare systems with increased rates of hospitalizations and readmissions, long lengths of stay, excessive costs, and high mortality rates. In addition, the prevalence of such high-risk HF patients continues to increase as overall survival improves in this patient population. There is a need to understand this population with better risk stratification, in order to provide optimal resource allocation. Stratifying the risk in this population can also identify appropriate patients for clinical trials focused on therapies that may improve long-term outcomes.

Several risk scores have been developed over the years mostly in stable chronic patients suffering from HF with reduced ejection fraction (HFrEF), which have performed variably well in the general population with limited adoption. Importantly, prior models were developed before the high use of guideline directed evidence based therapies, routine natriuretic peptide level ascertainment, and globally applied performance measures. (1, 2, 3, 4) A risk assessment algorithm ideally should integrate all clinically relevant, appropriately validated variables. Few have included important biomarker parameters such as natriuretic peptides in a high risk population treated in contemporary fashion. (5, 6, 7, 8, 9) Widely used models have not performed well in predicting short term outcomes. (10, 11, 12)

The recently completed Guiding Evidence-Based Therapy Using Biomarker Intensified Treatment (GUIDE-IT) trial (Clinicaltrials.gov NCT01685840) included 894 well characterized higher risk patients with HFrEF, randomized to receiving guideline-based therapy to suppress natriuretic peptide levels versus usual care with carefully assessed clinical outcomes of HF hospitalization or death, and death alone. (13, 14) The goal of this analysis was to develop a simple prediction tool to identify patients at highest risk for short term and long term morbidity and mortality, and who can be more carefully followed or potentially receive the most beneficial therapies.

METHODS

Patient Cohort

The GUIDE-IT trial design and outcomes has been previously reported. (13, 14) The GUIDE-IT study was a multicenter randomized controlled study that tested the strategy of augmented guideline-based therapy to suppress N-terminal pro-B type natriuretic peptide (NT-proBNP) concentrations to <1000 pg/mL versus usual care. The primary endpoint was the composite endpoint of time to HF hospitalization or CV death. The secondary endpoint of interest was all cause mortality. A clinically and policy relevant exploratory endpoint of 90 day HF hospitalization or death was examined. Members of an independent endpoint committee were blinded to treatment allocation. Inclusion criteria have been previously describe and briefly, were as follows: age≥ 18 years, HF event or treated with diuretics in prior 12 months, documented left ventricular LVEF ≤ 40% by any method within 12 months prior to randomization, and BNP > 400 pg/mL, or NT-proBNP > 2000 pg/mL in the 30 days prior to randomization (13). Exclusion criteria were as follows: clinical diagnosis of acute coronary syndrome or cardiac revascularization within 30 days, cardiac resynchronization therapy (CRT) within prior 3 months or current plans to implant CRT device, severe stenotic valvular disease, anticipated heart transplantation or LV assist device implantation within 12 months, chronic inotropic therapy, complex congenital heart disease, renal replacement therapy, non-cardiac terminal illness with expected survival less than 12 months, women who are pregnant or planning to become pregnant, inability to comply with planned study procedures, and enrollment or planned enrollment in another clinical trial. Clinical endpoints including all-cause hospitalization, cardiovascular-specific hospitalization, all-cause mortality, and cardiovascular-specific mortality were ascertained through endpoint committee adjudication.

Statistical Methods

Baseline characteristics were summarized as frequency and percentages for categorical variables, and medians with 25^th and 75^th percentiles for continuous variables. For the composite, primary endpoint of HF hospitalization or CV death, secondary endpoint of all-cause death and the exploratory endpoint of 90-day rehospitalization plus mortality, we developed a set of 35 candidate variables for possible model inclusion (Appendix table). The exploratory endpoint of 90-day rehospitalization plus mortality was examined because of the importance in the Center for Medicare and Medicaid Services (CMS) directed bundled care initiatives for HF and the associated financial penalties to hospitals for excessive short term readmission and mortality rates. The candidate variables represented a spectrum of baseline characteristics, demographics, medical history, physical exam, laboratory values, and biological markers such as NT-pro-BNP. For variables that were not 100% complete, chained equations multiple imputation was performed to create 25 data sets with no missing data. A complete list of the variables included in the imputation process are described in Supplemental Table 1. All missing data was assumed to be missing at random (MAR). The congestion score was defined as the sum of jugular venous pressure (cm, ordinal range: 0–3), hepatomegaly (ordinal range: 0–1), peripheral edema (ordinal range: 0–3), rales (ordinal range: 0–3), ascites (ordinal range: 0–1), and orthopnea (ordinal range: 0–4). The final ordinal score ranges from 0 – 14.

Multivariable Cox proportional hazards models were developed using backward selection at the 0.1 level of significance for variable inclusion in each of the 25 data sets. All continuous variables were assessed for linearity with respect to each outcome. For those that violated the assumption, natural cubic spline transformations were performed prior to the variable selection process. Variables that were chosen in at least 20 of the 25 imputed datasets were included in the final predictive models. Parsimonious risk models were then constructed by clinical guidance and by decreasing the significance level for variable inclusion such that only 5–7 variables were chosen.

The predictor variable relationships with the respective clinical outcomes were described by averaging the Cox model parameter estimates, and corresponding χ² statistics and c-indices, across the 25 imputed data sets. Internal validation was performed using 200 bootstrap samples to produce bias-corrected c-indices for each simplified model.

Model calibration for the simplified models was assessed by comparing 1 year or 90 day predicted risk and Kaplan-Meier (KM) event rates. The KM event rates at 1 year or 90 days were stratified according to deciles of predicted risk, and calibration was assessed using the D’Agostino-Nam Goodness-of-Fit test. (15) A significant p-value indicates that the model may be poorly calibrated. Calibration was also graphically assessed by plotting 1 year or 90 day KM event rates versus the predicted event rates by decile of predicted risk.

All analyses were performed by the Duke Clinical Research Institute (Duke University, Durham, NC) using SAS version 9.4 (SAS Institute, Inc., Cary, NC).

RESULTS

From 2013–2016, 894 patients were enrolled in the GUIDE-IT study at 45 centers in the United States and Canada. Outcomes for the primary and secondary endpoints between strategy arms of the study were similar. There were no major differences in baseline characteristics among treatment groups. The median follow-up was 15 months. Three hundred twenty-eight patients (37%) experienced the primary endpoint, and 143 (16%) experienced the secondary endpoint of all-cause mortality. (14)

Table 1 shows the clinical characteristics for the patients who had an HF hospitalization or CV death and those who were censored. The median age was 63 years; 32% were women; blacks comprised 37% and Hispanics 7% of participants; the median LVEF was 24% and NT-pro-BNP level of 2880 pg/ml. Patients experiencing the primary composite endpoint had a numerically higher median baseline age, were more likely to be black or Hispanic, had higher NYHA class level, more comorbidities, higher natriuretic peptide levels, and a higher congestion score, amongst other differences (Table 1).

Table 1.

Baseline Characteristics by Patient HF Hosp or CV Death Status

		HF Hospitalization or CV Death
Characteristic	N Missing	Censored	Event	P-value
Age	0 (0.00%)	62, 53–71	64, 53–72	0.77
Male Sex	0 (0.00%)	378 (66.8%)	230 (70.1%)	0.67
Ethnicity	0 (0.00%)	27 (4.8%)	31 (9.5%)	0.002
African American/Black	23 (2.64%)	179 (32.4%)	145 (45.6%)	0.0001
Caucasian/White	23 (2.64%)	337 (60.9%)	153 (48.1%)	0.0001
Weight	8 (0.90%)	85, 72–100	85, 72–106	0.17
BMI	21 (2.41%)	28, 24–33	29, 25–35	0.02
Duration of Heart Failure (months)	196 (28.08%)	6, 1–43	48, 7–96	<.0001
Any Hospitalization for Heart Failure	1 (0.11%)	407 (72.0%)	270 (82.3%)	0.0001
ICD or Pacemaker	1 (0.11%)	194 (34.3%)	201 (61.3%)	<.0001
History of Ischemic Heart Disease	1 (0.11%)	251 (44.4%)	196 (59.8%)	<.0001
Myocardial Infarction	1 (0.11%)	140 (24.8%)	111 (33.8%)	0.009
Ventricular Tachycardia/Fibrillation	2 (0.22%)	82 (14.5%)	78 (23.8%)	0.003
Peripheral Arterial Vascular Disease	1 (0.11%)	52 (9.2%)	42 (12.8%)	0.16
Stroke	1 (0.11%)	52 (9.2%)	43 (13.1%)	0.04
Hypertension	1 (0.11%)	434 (76.8%)	272 (82.9%)	0.10
Diabetes mellitus	1 (0.11%)	231 (40.9%)	179 (54.6%)	0.0001
COPD	1 (0.11%)	108 (19.1%)	85 (25.9%)	0.01
Chronic Liver Disease	1 (0.11%)	20 (3.5%)	12 (3.7%)	0.93
Depression treated with Medication	1 (0.11%)	83 (14.7%)	58 (17.7%)	0.07
6-minute walk distance	148 (19.84%)	312, 209–392	242, 157–343	<.0001
LVEF at Baseline	0 (0.00%)	25, 20–30	23, 20–30	0.67
Baseline NT-pro-BNP	0 (0.00%)	2355, 1280–4813	3405, 1849–6504	<.0001
NYHA Class at Enrollment	13 (1.48%)			<.0001
1		47 (8.4%)	12 (3.7%)
2		319 (57.2%)	128 (39.6%)
3		186 (33.3%)	172 (53.3%)
4		6 (1.1%)	11 (3.4%)
Heart Rate	7 (0.79%)	75, 66–86	78, 69–87	0.01
Systolic Blood Pressure	5 (0.56%)	116, 102–130	111, 100–126	0.03
Diastolic Blood Pressure	5 (0.56%)	70, 62–80	69, 60–78	0.04
Atrial Fibrillation at Baseline	13 (1.48%)	98 (17.6%)	47 (14.5%)	0.35
SpO2	107 (13.60%)	97, 96–99	97, 95–99	0.008
S3 Auscultation	26 (3.00%)	43 (7.9%)	38 (11.8%)	0.05
Jugular Venous Pressure	57 (6.81%)	109 (20.7%)	93 (29.9%)	<.0001
Peripheral Edema	11 (1.25%)	140 (25.2%)	105 (32.1%)	<.0001
Congestion Score	6 (0.67%)	2, 1–3	3, 1–4	<.0001
Potassium	36 (4.20%)	4, 4–5	4, 4–5	0.14
Sodium	33 (3.83%)	139, 137–141	138, 136–141	0.006
BUN	47 (5.55%)	20, 13–30	25, 17–40	<.0001
Creatinine	31 (3.59%)	1, 1–2	1, 1–2	<.0001
Cholesterol	808 (90.38%)	154, 131–184	133, 110–162	0.16
Lymphocytes	756 (84.56%)	22, 16–30	19, 15–23	0.02

Censored column contains patients who were lost to follow-up or did not have an event in the follow-up period.

P-value from the Cox proportional hazards model. All continuous variables were fit with a natural cubic spline to correct for any potential issues of non-linearity.

Predictors of Events in the GUIDE-IT Study

The rate of events in this cohort is shown in (Table 2). There was a high rate of HF hospitalizations or CV death, despite the use of guideline directed evidence based therapies. The intermediate predictive model for the primary endpoint is presented in (Appendix Table 3). The key prognostic factors in the parsimonious model associated with increased risk of the outcome included presence of implantable devices, lower serum sodium levels, higher NT-pro-BNP levels, higher NYHA class, sleep apnea, higher creatinine levels, Hispanic ethnicity, black race, and higher heart rate (Table 3).

Table 2.

Number of Events and Kaplan Meier Event Rate

Outcome	Events/Randomized Patients	Event Rate (%)
HF Hospitalization and CV Death[1]	328/894	34.5
HF Hospitalization and All-cause Death[1]	344/894	38.0
CV Death[1]	110/894	9.9
All-cause Death[1]	143/894	12.2
90-day Death or Re-hospitalization[2]	123/894	14.2

^[1]KM Rate at 1 year.

^[2]KM Rate at 90 days.

The number of events presented is over the entire follow-up period of the study (median follow-up = 15 months).

Table 3.

Multivariable Predictive Model for CV death or HF hosp (Parsimonious Model)

Predictor	HR (95% CI)	Chi-Square	P-value
ICD or Pacemaker	2.02 (1.61 – 2.54)	36.850	<.001
Sodium	0.94 (0.91 – 0.97)	14.651	<.001
Log of NT-pro-BNP	1.25 (1.10 – 1.41)	12.600	<.001
NYHA Class IV*	4.34 (1.90 – 9.91)	12.101	<.001
NYHA Class III*	2.01 (1.11 – 3.64)	5.262	0.022
NYHA Class II*	1.36 (0.75 – 2.47)	1.048	0.306
Sleep apnea	1.48 (1.15 – 1.89)	9.621	0.002
Log of Creatinine	1.67 (1.20 – 2.32)	9.453	0.002
Hispanic or Latino	1.84 (1.24 – 2.72)	9.295	0.002
Black	1.45 (1.14 – 1.84)	9.211	0.002
Heart Rate	1.01 (1.00 – 1.02)	6.281	0.012

^*Reference is NYHA Class I.

^ǂReported model selected from the following list of predictors: NYHA class, prior ICD or pacemaker implantation, black race, Hispanic ethnicity, NT-pro-BNP, sodium, history of sleep apnea, heart rate, creatinine, potassium, history of ischemic heart disease, diastolic BP, depression handled with medication and atrial fibrillation at baseline. Model was selected using backward selection with a significance level of 0.005 as the criterion for variable inclusion.

Uncorrected C-index: 0.707

Optimism-Corrected C-index: 0.694

Calibration (D’agostino, Nam test): p-value=0.33

Baseline Survival at 1 year [S₀(1 year)]: 0.8787

The intermediate predictive model for all-cause mortality is presented in (Appendix Table 4). Several factors from the primary endpoint were also highly prognostic in the mortality model. Significant factors in the parsimonious model included ischemic etiology of HF, longer HF duration, lower diastolic blood pressure, and higher heart rate (Table 4). Congestion score and NT-proBNP were the most important predictors of cardiovascular death or heart failure hospitalization by 90 days. The intermediate predictive model for the 90-day HF hospitalization or CV death outcome is shown in (Appendix Table 5) and the parsimonious model in (Table 5).

Table 4.

Multivariable Predictive Model for all-cause death (Parsimonious Model)

Predictor	HR (95% CI)	Chi-Square	P-value
Log of NT-pro-BNP	1.86 (1.56 – 2.22)	46.712	<.001
Sodium (mmol/L)	0.91 (0.87 – 0.95)	16.172	<.001
History of Ischemic Heart Disease	2.01 (1.39 – 2.91)	13.618	<.001
HF Duration (in years)	1.05 (1.02 – 1.08)	12.789	<.001
Diastolic Blood Pressure	0.97 (0.96 – 0.99)	12.398	<.001
Heart Rate	1.02 (1.01 – 1.03)	9.113	0.003

^ǂReported model selected from the following list of candidate predictors: NT-pro-BNP, diastolic blood pressure, sodium, heart rate, SpO₂, depression handled with medication, third heart sound, 6-minute walk distance, heart failure duration, history of ischemic heart disease, age, and congestion score. Model was selected using backward selection with a significance level of 0.01 as the criterion for variable inclusion.

Uncorrected C-index: 0.765

Optimism-Corrected C-index: 0.757

Calibration (D’agostino, Nam test): p-value=0.68

Baseline Survival at 1 year [S₀(1 year)]: 0.9414

Table 5.

Multivariable Predictive Model for CV death or HF Hosp at 90 days (Parsimonious Model)

Predictor	HR (95% CI)	Chi-Square	P-value
Congestion	1.23 (1.14 – 1.33)	26.82	<.001
Sleep Apnea	1.92 (1.32 – 2.80)	11.62	<.001
Log of NT-pro-BNP	1.35 (1.12 – 1.63)	10.10	0.001
Ethnicity	2.21 (1.32 – 3.69)	9.12	0.003

^ǂReported model selected from the following list of candidate predictors: history of ischemic heart disease, ethnicity, sleep apnea, NT-pro-BNP, potassium, congestion score, creatinine, heart rate, NYHA class, and atrial fibrillation at baseline. Model was selected using backward selection with a significance level of 0.01 as the criterion for variable inclusion.

Uncorrected C-index: 0.729

Optimism-Corrected C-index: 0.722

Calibration (D’agostino, Nam test): p-value=0.14

Baseline survival at 90 days [S₀(90 days)]: 0.9049

The only variable present across all three risk models was the NT-proBNP variable. In the parsimonious models, for every 1-unit increase in the natural log of natriuretic peptide level, there was a 25% increased hazard of the primary endpoint, an 86% increased hazard of all-cause mortality, and a 43% in the 90 day composite endpoint.

Having an implantable device (ICD, pacemakers, CRT, or combination) was predictive of the composite endpoint possibly due to the fact that those patients had higher median baseline NT-pro-BNP levels and were more often of Class III-IV HF status (Appendix Tables 7 and 8. The sensitivity analysis of the parsimonious model performance without this variable demonstrated continued good discrimination (c-index change: 0.697 no correction, 0.685 corrected).

Risk Score

As shown in Figures 1a and 1b, the primary endpoint and all-cause death endpoint models showed good calibration (D’Agostino-Nam test: p=0.30; p=0.37, respectively) and good discrimination with an optimism-corrected c-index of 0.69 for the primary endpoint and 0.76 for the mortality endpoint. The 90-day endpoint showed similarly good discrimination (c-index = 0.72). However, it showed less optimal calibration due to the fewer number of events in the first 90 days as compared to the primary endpoint (Supplemental Figure 1).

Figures 1a and 1b.

Calibration plots according to deciles of predicted risk for a) HF hospitalization or CV death and b) all cause death. Error bars depict the upper and lower 95% confidence bounds for the observed event rates.

The MAGGIC integer risk score for mortality was calculated in our cohort. We found that this risk score discriminates well (c-statistic: 0.70). It is not surprising that the c-statistic is lower than the c-statistic from our parsimonious model for all-cause death (0.76) since the MAGGIC score was derived on a different cohort of subjects.

Risk Tool

Figure 1 show the variables required for input to the risk tool, depending on which risk model is chosen. The resulting output appears to the right and at the bottom of the screen. After new patient details are entered, the risk will update automatically.

DISCUSSION

This analysis from the GUIDE-IT Trial clinical database represents contemporary risk prediction models using natriuretic peptides, and carefully selected clinical characteristics in a high-risk population with HFrEF; the cohort included recently hospitalized patients, and those in an ambulatory state with high levels of baseline natriuretic peptide. In this cohort of patients, there continues to be a high risk of morbidity and mortality. Simple clinical predictors available to clinicians at the bedside can discriminate a wide range of risks in this patient population for the risk of HF hospitalization plus CV death, all-cause death, and 90-day HF hospitalization or death rates.

One of the most important predictors that was present in all three models was the log of NT-pro-BNP. This is an important variable that not only is associated with prediction of all cause death, the composite of CV death and HF hospitalization, but HF hospitalization alone. (5, 16, 17). While it has been well known that NT-pro-BNP has important prognostic information, the confirmation in this high-risk population independent of ethnicity and race, many comorbidities and the congestion score provides additional important information to the clinician. (18) The systematic sampling of natriuretic peptide is highly variable across practitioners in this country, although it does possess Class I recommendation, Level of Evidence A for prognosis. (19, 20) Despite this fact, application has been variable and at times limited. (21) The findings in this study should help promote the routine use of ascertaining natriuretic peptides to understand the risk in HF patients at presentation.

Unique predictors of cardiovascular outcomes in this patient population included black race and Hispanic or Latino ethnicity. Both of these clinical characteristics have not been previously shown to be important independent risk factors in a high-risk heart failure population until recently. There is considerable literature supporting the fact that black patients and Hispanic patients may have increased risk for hospitalization.(22) However, in a carefully conducted clinical trial, the presence of these characteristics as risk factors independent of other important variables is unique.

Comorbidities remain important contributors to risk in this patient population. Comorbidities that were significant in the model included a history of sleep apnea, hyponatremia, renal dysfunction, and ischemic heart disease. The findings of this study emphasized the importance of screening and addressing comorbidities. (23)

High heart rate and low blood pressure emerged as important risk factors in our analysis. Increases in heart rate and low blood pressure have been associated with increased risk of morbidity and mortality. (18, 24) The importance of identifying elevated heart rate and treating appropriately with more aggressive application of beta blocker therapy or ivabradine remains an important strategy.

The counter-intuitive finding that implantable devices were associated with a worse outcome only in the composite endpoint model may be explained by the fact that sicker patients at baseline had devices, and therefore were associated with higher event rates. The independent contribution of this variable above other degrees of illness likely reflects the higher NYHA Class, higher natriuretic peptide levels, and longer duration of disease but more investigation is needed. The fact that the variable did not present in the mortality or 90 day composite endpoint model suggests its predictive value is based on the late term HF hospitalizations.

The simple 90 day short term model, which demonstrated good prediction (c-index 0.74), may have a particular advantage in patients at time of discharge, identifying high risk patients who may drive readmission and mortality penalties for hospitals. The congestion score was the most important predictor, highlighting the importance of severe congestion. The clinical utility of these risk prediction models includes identifying patients for consideration of advanced therapies in very high risk patients (MCS, transplant, or palliative care), and enhanced surveillance for short term risk management. These predictive models are indeed valuable with c-indices comparable to the CMS 30-day hospitalization risk prediction model associated with hospital penalties in the U.S., with fewer variables.

The development of risk models for HFrEF patients at high risk for hospitalization may be helpful in evaluating prognosis at the time of presentation, although many models are too complex. By developing simple risk scores based on the most important prognostic factors, we can provide important information in real time on how to counsel patients at risk and allocate therapies for different patients. For example, in this population we see that we can stratify risk of one-year mortality from as low as < 5% to as high as > 70%. Clearly, the use of advanced therapies such as temporary and durable VADs, cardiac transplantation, and palliative care may be opportunities for therapies in these patients. This analysis is novel in the following manner: different demographics from previous publications in the setting of contemporary baseline therapy, ethnic subgroups, and the duration of follow up in which the natriuretic peptide is predictive. In addition, the careful phenotypic characterization of a high risk HFrEF population, short term and long term modeled outcomes, and the use of a standardized natriuretic peptide assay provide unique characteristics of the study. In summary, the GUIDE-IT risk scores offer several important and unique advantages over similar models and risk scores:

1. Natriuretic peptides were a focus of the clinical trial and were carefully collected throughout the study.

2. There was an emphasis to enrich the recruitment of women, blacks, and Hispanic patients, and therefore sufficient information is available to understand their heightened risk in this clinical data set.

3. This study represents a cohort of patients with a broad base of presentations, including both recently acute decompensated HF and ambulatory patients with high natriuretic peptides with an LVEF of < 40%.

4. The scores incorporate readily available information that is available at the bedside to clinicians, and the importance of natriuretic peptide as a prognostic variable re-emphasizes the value of collecting this information for prognostication and optimization of clinical care.

This analysis has several limitations. The present models were developed in the cohort of high-risk heart failure patients both hospitalized and high risk ambulatory heart failure with systolic dysfunction. Further study is needed to understand whether these models would be equally informative to patients with HF and preserved EF or generally lower risk patients. While there are several important predictive models in the ambulatory HF population including MAGGIC, HF-ACTION, the Seattle HF Model, and others, no model has performed well with this phenotype characterization across the short term and long term follow –up periods. (1, 11, 25) Therefore, we believe the development of predictive models and risk scores in this population represents an important advance in knowledge. Additionally, because of a number of distinctive variables incongruous with other datasets, external validation of our risk scores in a similar cohort could not be performed. “While we were able to conduct internal validation through bootstrapping techniques, and apply the GUIDE-IT data to the MAGGIC predictive model with excellent performance, we were not able to validate the model in an external dataset. We performed a variable by variable comparison of the GUIDE-IT dataset versus the PROTECT dataset, the ESCAPE dataset, and the HF-ACTION dataset, and found that greater than 20% of the variables were not available or incongruent, including standardization of the natriuretic peptide assays. In addition, the outcome variables were ascertained at different time points (i.e. 180 days vs 1 year). We therefore felt that the ability to conduct external validation would be incomplete and misleading.” The entire variable selection process was not carried out for each bootstrap sample, since that can only be done if the entire variable selection process was automated. In this analysis, there was clinical input in the variable selection process.

This trial represents a relatively young cohort of patients, and therefore the very elderly and frail patients are less represented. On the other hand, this is one of the largest cohorts of Hispanics, blacks, and women included in a HF clinical trial, and therefore the characterization of special populations and their importance to predictive models is important. Most of the 35 candidate variables for the models had < 10% missing data, with 4 of the 35 variables having more than 10% missing data. Multiple imputation was used to handle missing data.

Conclusions

High risk patients with HFrEF have an elevated rate of morbidity and mortality despite a high use of guideline directed evidence-based therapies. Simple, easily obtained clinical characteristics and laboratory values are important in determining the risk of HF hospitalization and death, both short term and long term. The use of these predictive risk scores may be a powerful tool for helping the clinician in determining risk, allocating treatment, identifying patients for clinical trials, and providing more aggressive follow up and surveillance if needed.

Figures 2a and 2b.

Screenshot of web based tool for risk prediction. A) Representation of patient with NT-pro-BNP of 1000 pg/mL, creatinine of 50 mg/dL, sodium of 135 mmol/L, heart rate of 79, and reference categories for the other variables. B) Representation of patient with NT-pro-BNP of 1000 pg/mL, heart failure duration of 7 months, ischemic heart disease, sodium of 127 mmol/L, heart rate of 97, and diastolic blood pressure of 74.

Abbreviations

CHARM

candesartan in heart failure-assessment of reduction in mortality and morbidity

CRT

cardiac resynchronization therapy

CV

cardiovascular

EF

ejection fraction

GUIDE-IT

guiding evidence based therapy using biomarker intensified treatment

LVEF

left ventricular ejection fraction

MAGGIC

meta-analysis global group in chronic heart failure

NT-proBNP

N- terminal prohormone of brain natriuretic peptide

PHQ-2

patient health questionnaire-2

VAD

ventricular assist device