OUP user menu

The MUSIC Risk score: a simple method for predicting mortality in ambulatory patients with chronic heart failure

Rafael Vazquez , Antoni Bayes-Genis , Iwona Cygankiewicz , Domingo Pascual-Figal , Lilian Grigorian-Shamagian , Ricardo Pavon , Jose R. Gonzalez-Juanatey , José M. Cubero , Luis Pastor , Jordi Ordonez-Llanos , Juan Cinca , Antoni Bayes de Luna
DOI: http://dx.doi.org/10.1093/eurheartj/ehp032 1088-1096 First published online: 24 February 2009


Aims The prognosis of chronic heart failure (CHF) is extremely variable, although generally poor. The purpose of this study was to develop prognostic models for CHF patients.

Methods and results A cohort of 992 consecutive ambulatory CHF patients was prospectively followed for a median of 44 months. Multivariable Cox models were developed to predict all-cause mortality (n = 267), cardiac mortality (primary end-point, n = 213), pump-failure death (n = 123), and sudden death (n = 90). The four final models included several combinations of the same 10 independent predictors: prior atherosclerotic vascular event, left atrial size >26 mm/m2, ejection fraction ≤35%, atrial fibrillation, left bundle-branch block or intraventricular conduction delay, non-sustained ventricular tachycardia and frequent ventricular premature beats, estimated glomerular filtration rate <60 mL/min/1.73 m2, hyponatremia ≤138 mEq/L, NT-proBNP >1.000 ng/L, and troponin-positive. On the basis of Cox models, the MUSIC Risk scores were calculated. A cardiac mortality score >20 points identified a high-risk subgroup with a four-fold cardiac mortality risk.

Conclusion A simple score with a limited number of non-invasive variables successfully predicted cardiac mortality in a real-life cohort of CHF patients. The use of this model in clinical practice identifies a subgroup of high-risk patients that should be closely managed.

  • Heart failure
  • Prognosis
  • Mortality
  • Natriuretic peptides


Chronic heart failure (CHF) is a major public health problem, with an increasing incidence and prevalence.12 The likelihood of survival may vary significantly among different subsets of patients with CHF. Several prognostic multivariable models have been developed,39 but all of them have some limitations for their use in current clinical practice: they relied on either peak oxygen consumption or invasive measures of cardiac function3 were validated during a hospitalization for heart failure (HF)46 or included only patients with systolic dysfunction.7 Furthermore, most of these models have not included a substantial proportion of patients taking contemporary evidence-based treatments, including angiotensin-converting enzyme (ACE)-inhibitors, beta-blockers, spironolactone, and angiotensin-receptor blockers. Finally, neither of the two recently reported models with participants from clinical trials included determination of natriuretic peptides nor Holter monitoring parameters.10,11

Current CHF guidelines prompted the development of newer prognostic models due to the great treatment improvement experienced over the past decade.1 The aim of this study was to develop a model for predicting mortality in an independent, non-clinical trial, outpatient CHF population using variables easily assessable in clinical practice.


Study population

The study population consisted of 992 patients with CHF enrolled in the MUSIC (MUerte Subita en Insuficiencia Cardiaca) study, a prospective, multicentre, longitudinal study designed to assess risk predictors of cardiac mortality and sudden cardiac death in ambulatory patients with CHF.

In the MUSIC study, patients were consecutively enrolled from the specialized HF clinics of eight University Hospitals between April 2003 and December 2004. All had symptomatic CHF (NYHA class II–III) and were treated according to institutional guidelines. This study included patients with either depressed (<45%) or preserved (≥45%) left ventricular ejection fraction (LVEF). The latter were included if they had HF symptoms and a prior hospitalization for HF or some objective signs of HF confirmed by chest X-ray (findings of pulmonary congestion) and/or echocardiography (abnormal LV filling pattern and LV hypertrophy). Patients were excluded if they had recent acute coronary syndrome or severe valvular disease amenable for surgical repair. Patients with other concomitant diseases expected to reduce life-expectancy were also excluded. The study protocol was approved by institutional investigation committees, and all patients signed informed consent.

Study protocol

All subjects had clinical, 12-lead ECG, echocardiography, chest X-ray, 24 h Holter monitoring, and blood laboratory parameters performed at enrolment. Initial assessments were performed in 1054 patients; 58 were deemed ineligible for inclusion due to poor acoustic windows and 4 additional patients refused to give their consent for the study. The remaining 992 patients constitute the MUSIC study population.

Standard criteria were used for clinical variables. The parameter prior atherosclerotic vascular event (AVE) was computed as positive if the patient had any of the following: previous myocardial infarction, stroke, or lower limb ischaemia.

Conduction defects were defined as in the MUSTT trial.12 Specifically, intraventricular conduction delay (IVCD) was defined as QRS duration ≥0.11 s, but morphology differs from left bundle-branch block (LBBB) or right bundle-branch block (RBBB). As in the MUSTT trial, a combined variable was created to include patients presenting LBBB or IVCD.

Among Holter monitoring parameters, a combined variable was also used to identify patients who presented frequent (>240 in 24 h) ventricular premature beats (VPBs) and also non-sustained ventricular tachycardia (NSVT).

Left atrial (LA) diameter, left ventricular (LV) end-diastolic diameter, and LV mass were determined according to the previously published methods13,14 and were indexed to body surface area. Left ventricular ejection fraction was evaluated using the Simpson method.15 Mitral flow pattern was assessed as described previously and validated.16,17

Blood samples were processed as described previously.30 A cut-off value of >1.000 ng/L was used to dichotomize N-terminal pro-brain natriuretic peptide (NT-proBNP) results. Estimated glomerular filtration rate (eGFR) was computed by the MDRD method.18 Cardiac troponins (either I or T) were considered ‘positive’ if their values were greater than the corresponding 99th reference percentile.19

Follow-up visits were conducted on an outpatient basis every 6 months, for a median of 44 months (range 28–51), until June 2007. Information about deceased patients was obtained from medical records, patients' physicians, and relatives. In each case, the attempt was made to determine the nature of death. Total mortality was divided into non-cardiac and cardiac mortality, and the latter was further subdivided into pump-failure death (PFD) and sudden death (SD), using the following definitions:20,30 Death was defined as ‘sudden’ if it was: (i) a witnessed death occurring within 60 min from the onset of new symptoms, unless a cause other than cardiac was obvious; (ii) an unwitnessed death (<24 h) in the absence of pre-existing progressive circulatory failure or other causes of death; or (iii) a death during attempted resuscitation. Deaths occurring in hospitals as a result of refractory progressive end-stage HF were defined as PFD. Patients lost to follow-up (n = 11) were censored in survival analysis. Patients who underwent cardiac transplantation (n = 20, median time on the waiting list = 20 days) were defined as PFD at the time of surgery, according to previously published CHF studies.10,25

All previously mentioned parameters were prospectively collected using specially designed case record forms.

Statistical analysis

Variables previously reported to be associated with mortality in patients with CHF were identified (Table 1). As in the model based on the CHARM programme database (CHARM-model),11 we did not include variables containing information on concomitant or past medical or procedural treatments, as it is impossible to distinguish causality of treatment effects from selection bias. In addition, subjective variables as the NYHA functional class were excluded from multivariable analyses. Variables with >5% of missing data were also excluded (right ventricular dimension and function, LDL- and HDL-cholesterol, and thyroid hormones). For variables with <5% of missing data, median imputations were used, when necessary. The remaining variables, given in Table 1, were evaluated with the Cox proportional hazards model, using SPSS 15.1 for Windows. The model included only the main effects of the predictors, without any interactions term.

View this table:
Table 1

Baseline characteristics in 992 heart failure patients in the MUSIC study

Demographic and clinical variablesEchocardiographic variables
 Age, years65 ± 12 LA size >26 mm/m2323 (32.6%)
 Gender, male718 (72.4%) LV end-diastolic diameter, mm/m232.8 ± 6.1
 Diabetes mellitus356 (35.9%) LV mass, g/m2162.4 ± 52.2
 History of hypertension565 (57.0%) LVEF, %37.0 ± 14.1
 History of dyslipemia494 (49.8%) LVEF <45%748 (75.4%)
 Current smoker104 (10.5%) LVEF ≤35%534 (53.8%)
 Ischaemic aetiology453 (45.7%) Restrictive filling pattern78 (7.9%)
 Prior AVE478 (48.2%)12-Lead ECG and 24-h Holter monitoring variables
 Prior CABG or PTCA256 (25.8%) Sinus rhythm703 (70.9%)
 Prior pacemaker135 (13.6%) Atrial fibrillation191 (19.3%)
 Body mass index, kg/m228.5 ± 4.5 Pacemaker rhythm98 (9.9%)
 NYHA class II778 (78.4%) Heart rate, b.p.m.70.1 ± 10.1
 NYHA class III214 (21.6%) QRS duration, ms125.5 ± 35.1
 Systolic blood pressure, mmHg127 ± 22 LBBB or IVCD485 (48.9%)
Radiographic variables Non-sustained VT352 (35.5%)
 Cardiothoracic ratio0.55 ± 0.07 Frequent VPBs (>240 VPBs in 24 h)480 (48.4%)
 Signs of pulmonary venous hypertension169 (17.0%) Non-sustained VT and frequent VPBs284 (28.6%)
Laboratory variablesMedications at presentation
 Haemoglobin, g/L137.2 ± 15.9 ACE inhibitor734 (74.0%)
 GGT >50 IU/L251 (25.3%) Angiotensin II receptor blocker165 (16.6%)
 Total cholesterol, mmol/L4.83 ± 1.07 Beta-blocker675 (68.0%)
 Creatinine, µmol/L109.1 ± 34.7 Spironolactone372 (37.5%)
 eGFR <60 mL/min/1.73 m2458 (46.2%) Loop diuretics721 (72.7%)
 Hyponatremia ≤138 mEq/L374 (37.7%) Digoxin298 (30.0%)
 NT-proBNP >1.000 ng/L452 (45.6%) Statins489 (49.3%)
 Troponin-positive158 (15.9%) Aspirin and/or clopidogrel463 (46.7%)
  • Qualitative data are presented as absolute frequencies and percentages and quantitative data as mean ± standard deviation.

  • AVE, atherosclerotic vascular events: myocardial infarction or stroke, or lower limb ischaemia; CABG, coronary artery bypass graft surgery; eGFR, estimated glomerular filtration rate (MDRD method); GGT, gamma-glutamil transpeptidase; IVCD, intraventricular conduction delay; LA, left atrium; LBBB, left bundle-branch block; LV, left ventricle; LVEF, left ventricular ejection fraction; NT-proBNP, amino-terminal pro-brain natriuretic peptide; NYHA class, New York Heart Association class; PTCA, percutaneous transluminal coronary angioplasty/stent; VPBs, ventricular premature beats; VT, ventricular tachycardia.

Primary outcome in this study is time to cardiac death. Secondary endpoints are time to death from any cause, PFD, and SD. For each endpoint, Cox proportional hazards models were built using a forward stepwise procedure with P < 0.05 as the inclusion/deletion criterion.

To calculate the MUSIC Risk score for each type of mortality, every variable in the multivariable model was multiplied by its β-coefficient, and the products were summed. A simplified and user-friendly version of the risk scores was developed in this way: The β-coefficients of the multivariable predictors were multiplied by 10 (11 in the SD risk score) and rounded to the nearest integer. Adding up the corresponding numbers of the multivariable predictors present in a particular patient results in scores of mortality, ranging from 0 to 40.

For each outcome, the MUSIC Risk score was plotted against the probability of death, estimated by the corresponding Cox model. Model calibration and ability to separate populations of patients into different risk groups were evaluated by assessing predicted vs. actual (Kaplan–Meier) survival in deciles of their risk scores, and a correlation coefficient and standard error of the estimate (SEE) was calculated.10,11 The models’ discrimination abilities were determined by the c-index.11,21 The internal and external validity of the final predictive models was assessed by the bootstrap re-sampling technique.22 For each of 200 bootstrap samples, the model was re-fitted and tested on: (a) the original sample to obtain a bias-corrected estimate of predictive accuracy (internal validity);11,22 (b) the subgroup of patients omitted from the bootstrap sample (‘out-of-bag’—OOB—observations), as an estimate of the external validity of the model.23,24 Finally, low- and high-risk subgroups of CHF patients were identified using the cut-off value of 20, for any mortality score.

The authors had full access to the data and take responsibility for its integrity. All authors have read and agree to the manuscript as written.


Patient characteristics

The MUSIC study included 992 consecutive ambulatory patients with CHF (718 men and 274 women), aged 18–89 years (mean 65 ± 12). Most patients (78.4%) were in NYHA class II. Ischaemic aetiology of CHF was present in 45.7% of patients. Mean LVEF was 37.0 ± 14.1% (range 10–70%), and most patients (75.4%) presented LVEF <45%. The majority of patients (70.9%) was in sinus rhythm and 19.3% had atrial fibrillation. ECG showed LBBB or IVCD in 48.9% of patients. In 28.6% of Holter recordings, both frequent VPBs and NSVT were present. Among the blood test results, eGFR <60 mL/min/1.73 m2 was observed in 46.2% of cases, hyponatremia ≤138 mEq/L in 37.7%, NT-proBNP >1.000 ng/L in 45.6% and troponin-positive in 15.9% of patients. Detailed baseline characteristics of the study population are summarized in Table 1.

Predicting mortality

We recorded 267 (26.9%) deaths including 213 (21.5%) cardiac deaths and 54 (5.4%) non-cardiac deaths. Among cardiac deaths, 123 (12.4%) were PFD and 90 (9.1%) were SD. The final prognostic models for cardiac mortality, PFD, and SD are shown in Tables 2, 3, and 4, respectively. The total mortality model is not shown, as it was very similar to that of cardiac mortality, except for atrial fibrillation that was an independent predictor of the former but not the latter.

View this table:
Table 2

Univariable and multivariable predictors of cardiac mortality

VariableUnivariable hazard ratio (95% CI)Wald x2P-valueMultivariable hazard ratio (95% CI)Beta CoefficientWald x2P-Value
Demographic and clinical variables
 Age ≥65 years1.52 (1.15–2.01)8.50.004
Prior AVE1.49 (1.14–1.96)8.40.0041.36 (1.03–1.79)0.3064.70.030
 Diabetes mellitus1.33 (1.01–1.75)4.20.04
 Body mass index <25 kg/m21.44 (1.07–1.94)5.70.017
Echocardiographic variables
LA size > 26 mm/m23.66 (2.78–4.81)86.4<0.00052.50 (1.88–3.33)0.91639.0<0.0005
 LV end-diastolic diameter ≥ 33 mm/m22.14 (1.62–2.84)28.3<0.0005
 Grade 3 or 4 mitral regurgitation2.28 (1.64–3.18)23.8<0.0005
LVEF ≤35%2.49 (1.85–3.37)35.5<0.00051.69 (1.24–2.31)0.52510.90.001
 Restrictive filling pattern1.94 (1.29–2.90)10.30.001
12-Lead ECG and 24-h Holter monitoring variables
 Atrial fibrillation1.72 (1.27–2.34)12.2<0.0005
 Heart rate >80 b.p.m.1.61 (1.20–2.17)9.90.002
 QRS duration >120 ms1.61 (1.23–2.11)11.90.001
 LBBB or IVCD1.75 (1.33–2.30)15.8<0.0005
Non-sustained VT and frequent VPBs2.53 (1.93–3.31)45.7<0.00051.59 (1.20–2.11)0.46410.40.001
Laboratory variables
eGFR<60 mL/min/1.73 m22.41 (1.81–3.19)37.2<0.00051.55 (1.15–2.09)0.4378.20.004
Hyponatremia ≤138 mEq/L1.60 (1.22–2.09)11.70.0011.35 (1.03–1.77)0.2984.60.031
NT-proBNP > 1.000 ng/L0.24 (0.18–0.33)82.5<0.00052.15 (1.54–3.01)0.76620.0<0.0005
Troponin-positive2.67 (1.99–3.58)43.1<0.00051.74 (1.28–2.37)0.55412.4<0.0005
 Haemoglobin <120 g/L1.45 (1.02–2.07)4.20.04
 GGT >50 IU/L1.76 (1.33–2.34)15.7<0.0005
  • Abbreviations as in Table 1; P-values based on univariable and multivariable Cox proportional hazards models; multivariable predictors of mortality are represented in italic typeface.

View this table:
Table 3

Univariable and multivariable predictors of pump-failure death

VariableUnivariable hazard ratio (95% CI)Wald x2P-valueMultivariable hazard ratio (95% CI)Beta coefficientWald x2P-value
Demographic and clinical variables
 Age ≥65 years1.52 (1.05–2.20)5.00.025
 Body mass index <25 kg/m21.84 (1.27–2.68)10.20.001
Echocardiographic variables
LA size >26 mm/m23.70 (2.58–5.30)50.8<0.00052.50 (1.73–3.62)0.91723.6<0.0005
 LV end-diastolic diameter ≥33 mm/m22.19 (1.51–3.17)17.1<0.0005
 Grade 3 or 4 mitral regurgitation2.61 (1.71–3.98)19.9<0.0005
LVEF ≤35%2.42 (1.64–3.58)19.6<0.00051.75 (1.18–2.61)0.5617.60.006
 Restrictive filling pattern1.87 (1.09–3.21)5.20.023
12-Lead ECG and 24-h Holter monitoring variables
 Atrial fibrillation1.79 (1.20–2.67)8.20.004
 Heart rate >80 b.p.m.2.02 (1.39–2.95)13.6<0.0005
 QRS duration >120 ms1.77 (1.23–2.53)9.70.002
 LBBB or IVCD1.43 (1.00–2.04)3.90.048
 Non-sustained VT and frequent VPBs2.37 (1.66–3.37)22.6<0.0005
Laboratory variables
eGFR<60 mL/min/1.73 m23.01 (2.04–4.42)31.3<0.00051.79 (1.19–2.69)0.5847.90.005
Hyponatremia ≤138 mEq/L1.90 (1.33–2.71)12.6<0.00051.60 (1.12–2.29)0.4736.80.009
NT-proBNP > 1.000 ng/L0.18 (0.12–0.28)59.2<0.00052.87 (1.80–4.57)1.05319.6<0.0005
Troponin-positive3.14 (2.16–4.58)35.8<0.00052.09 (1.41–3.08)0.73613.7<0.0005
 Haemoglobin <120 g/L1.86 (1.21–2.86)7.90.005
 GGT >50 IU/L2.23 (1.56–3.20)19.1<0.0005
  • Abbreviations as in Table 1; P-values based on univariable and multivariable Cox proportional hazards models; multivariable predictors of mortality are represented in italic typeface.

View this table:
Table 4

Univariable and multivariable predictors of sudden cardiac death

VariableUnivariable hazard ratio (95% CI)Wald x2P-valueMultivariable hazard ratio (95% CI)Beta coefficientWald x2P-value
Demographic and clinical variables
Prior AVE2.05 (1.33–3.16)10.60.0012.15 (1.40–3.32)0.76712.10.001
Echocardiographic variables
LA size >26 mm/m23.61 (2.37–5.50)35.7<0.00052.59 (1.65–4.06)0.95117.0<0.0005
 LV end-diastolic diameter ≥33 mm/m22.09 (1.36–3.22)11.20.001
 Grade 3 or 4 mitral regurgitation1.87 (1.09–3.21)5.20.023
 LVEF ≤35%2.60 (1.63–4.14)16.0<0.0005
 Restrictive filling pattern2.03 (1.10–3.73)5.20.023
12-Lead ECG and 24-h Holter monitoring variables
 Atrial fibrillation1.64 (1.02–2.63)4.20.042
LBBB or IVCD2.34 (1.50–3.63)14.1<0.00051.90 (1.21–2.99)0.6447.80.005
Non-sustained VT and frequent VPBs2.77 (1.83–4.19)23.4<0.00051.89 (1.23–2.90)0.6358.40.004
Laboratory variables
 eGFR <60 mL/min/1.73 m21.81 (1.19–2.76)7.70.006
NT-proBNP > 1.000 ng/L0.34 (0.22–0.52)23.4<0.00051.82 (1.14–2.92)0.6016.20.012
 Troponin-positive2.10 (1.31–3.38)9.40.002
  • Abbreviations as in Table 1; P-values based on univariable and multivariable Cox proportional hazards models; multivariable predictors of mortality are represented in italic typeface.

The 20 parameters displayed in Table 2 were associated with increased cardiac mortality, but only 8 of them were independent predictors in multivariable analysis: prior AVE, LA size >26 mm/m2, LVEF ≤35%, NSVT and frequent VPBs, eGFR <60 mL/min/1.73 m2, hyponatremia ≤138 mEq/L, NT-proBNP >1.000 ng/L, and troponin-positive.

As displayed in Table 3, PFD was independently associated with 6 from 18 univariable predictors: LA size >26 mm/m2, LVEF ≤35%, eGFR <60 mL/min/1.73 m2, hyponatremia ≤138 mEq/L, NT-proBNP >1.000 ng/L, and troponin-positive.

Finally, SD was independently associated with 5 from 12 univariable predictors: prior AVE, LA size >26 mm/m2, LBBB or IVCD, NSVT and frequent VPBs, and NT-proBNP >1.000 ng/L (Table 4).

Predicting an individual's risk

The models presented in Tables 24 can be used to predict any individual's risk of each endpoint. The simplest method to calculate the risk score for each mode of death is by adding up the simplified β-coefficients (as whole numbers) presented in Figure 1. The obtained risk score may be used to estimate survival using the curves displayed in Figure 2. Figure 2 shows the relationship between risk score and estimated probability of mortality within 44 months of follow-up.

Figure 1

Risk scores for each type of mortality. This figure summarizes the independent predictors of the multivariable models shown in Tables 24. Scores are calculated for each patient by adding together the points corresponding to his/her risk predictors. Scores >20 are associated with high risk of mortality. Horizontal bars in the figure represent the relative weight of each predictor. Abbreviations as in Table 1.

Figure 2

Probability of death in chronic heart failure patients followed for 44 months, as a function of the risk score.

Performance of the models

For the entire cohort of 992 patients, the predicted vs. actual 44-month survival rates were: 72.1% vs. 71.8%, 77.1% vs. 76.7%, 86.2% vs. 85.2%, and 91.7% vs. 89.6%, for the total, cardiac, PFD, and SD models, respectively. The correlation between predicted and actual survival by deciles of the risk scores was 0.99, SEE = ±3% (Figure 3). The discrimination abilities of the models were rather strong, with c-indices of 0.76, 0.78, 0.80, and 0.77, for total mortality, cardiac mortality, PFD, and SD models, respectively. After bootstrap sampling, c-indices were 0.77, 0.78, 0.80, and 0.78, respectively, using the re-fitted model on the original sample (internal validation), and 0.75, 0.78, 0.78, and 0.74 on the OOB observations (external validation).

Figure 3

Predicted vs. actual survival by deciles of the risk score, for each outcome.

The main model of this study (the cardiac mortality model) was also tested in the subgroups of patients with preserved (≥45%) and depressed (<45%) LVEF with c-indices of 0.80 and 0.77, respectively.

Low- and high-risk subgroups of patients

The whole cohort of 992 patients was divided into two subgroups of high- and low-risk patients using the cut-off value of 20 for each risk score. For the cardiac mortality model, the low-risk subgroup includes 662 patients (66.7%) with an average mortality of 11%, while in the high-risk subgroup (n = 330, 33.3%), mortality reached 47%. For the SD model, the low-risk subgroup consists of 682 patients (68.8%) with an average mortality of 5%, while in the high-risk subgroup (n = 310, 31.2%), mortality was around 20%. Similar information for the PFD and total mortality models are summarized in Figure 4.

Figure 4

Predicted vs. actual mortality in high and low subgroups of risk, for each outcome. High-risk subgroups include only one-third of patients, with a four-fold mortality risk.


Our study documented that a simple score, based on non-invasive variables, predicts mortality in a large ambulatory population of CHF patients, with the full spectrum of LV systolic function, as seen in everyday clinical practice.

Cardiac mortality risk score should be regarded as the main result of the present analysis. Risk scores to predict PFD and SD are based on fewer events and less predictors, and should not be used directly; on the contrary, they should only be used after the cardiac mortality risk score, to evaluate if a high-risk patient (>20 points in the cardiac mortality risk score) is prone to PFD or SD. On the other hand, the total mortality risk score, although based on more events and predictors, includes a heterogeneous subgroup of non-cardiac deaths. Nevertheless, total and cardiac mortality risk scores are so similar that their results in most patients are almost identical.

Predicting an individual's risk in daily clinical practice requires only adding up the points of the predictors present in that patient to calculate the cardiac mortality risk score. The scores obtained in this way may be translated into probabilities of death (Figure 2). This procedure provides the most exact estimations of survival. To further simplify this process, a quick and easy estimation of the patient's risk profile (high vs. low) may be done by just checking if the cardiac mortality score is >20. The rationale for this simplification is that many clinical decisions are taken on the basis that the patient is known to be at high risk of an event, without calculating the exact probability of that event to happen. A cardiac mortality score >20 identifies a small subgroup of patients (one-third of the entire cohort) with a four-fold cardiac mortality risk. This high-risk subgroup of patients should be managed closely by specialized HF units, whereas most low-risk patients require a less intensive follow-up.

Several prognostic CHF models have been published previously.311 Among other aforementioned limitations, older models39 did not include a substantial number of patients with optimized medical therapy according to current guidelines,12 and/or were derived in hospitalized patients46 and/or included only patients with systolic dysfunction.7 More recently, the Seattle Heart Failure Model (SHFM)10 and the CHARM-model11 have been developed in participants from clinical trials; these two models did not include current biochemical markers, such as natriuretic peptides or cardiac troponins. Moreover, both models10,11 comprise such a huge number of parameters that require complex computations for predicting an individual's risk. Therefore, although they are very valuable for research purposes, these shortcomings limit their use in daily clinical practice. Finally, previously published prognostic models on ambulatory CHF relied upon variables such as the NYHA functional class.7,911 As stated in the ACC/AHA 2005 CHF guidelines, the NYHA functional classification reflects a subjective assessment and can change frequently over short periods of time.1 Furthermore, NYHA functional class is difficult to establish among patients with limited physical activity, co-morbidities or advanced age, circumstances that are very common among CHF patients.1 For all these reasons, we intentionally excluded this parameter from multivariable analyses, in order to develop a model exclusively based on objective variables.

Several of the predictors included in the cardiac mortality model have been traditionally used to estimate prognosis in CHF, such as LVEF, eGFR, or hyponatremia. Age was not independently associated with mortality, probably due to the inclusion of more powerful risk factors in the model. Recently, natriuretic peptides have been associated with both SD and non-sudden cardiac death,20,26 and LA enlargement has also been acknowledged as a precise surrogate of cardiovascular mortality and SD.2730 These two predictors are the only parameters present in the four risk scores for each type of mortality. Another biochemical parameter, high serum troponin, has been recently associated with higher mortality in HF patients, in the absence of ischaemic events; the mechanism of this ongoing subclinical myocyte injury in patients with CHF remains unclear.31 The results of our study support this prognostic role of troponin in CHF patients, as this variable was an independent predictor of cardiac and total mortality, as well as PFD. Among ECG parameters, ‘LBBB or IVCD’ was associated with all types of mortality on univariable analyses but only with SD on multivariable analyses. Other conduction defects such as RBBB were not associated with increased mortality. These findings are similar to those of the MUSTT trial.12 Previous reports found a high risk of SD among CHF patients with NSVT.32,33 Owing to the high prevalence of this arrhythmia among CHF patients, we used the combined parameter ‘NSVT and frequent VPBs’, which resulted in an independent predictor of SD, cardiac mortality, and total mortality.


The MUSIC cardiac mortality risk score was derived in a cohort of CHF patients with predominant systolic CHF (75% of cases); however, the proportion of patients with preserved LVEF in our study is the same as in the CHARM-model.11 The SHFM was developed and validated mainly in patients with systolic HF, but its accuracy was similar in the IN-CHF population, in whom one-third of patients had an EF of ≥40%.10 Finally, we demonstrated that the performance of the cardiac mortality model was equally good in the preserved and depressed LVEF populations, with c-indices of 0.80 and 0.77, respectively. Our model was not validated in a separate data set, because there are no other similar contemporary cohorts of CHF patients in which the 10 multivariable predictors summarized in Figure 1 have been determined. On the other hand, the relatively small number of SDs (n = 90) precluded splitting our cohort into training and validation samples. However, the bootstrap re-sampling technique ruled out any ‘over-optimism’ in the predictive discrimination with similar c-indices, using the original and re-fitted models.11,21,22 Furthermore, using the re-fitted model on the OOB observations after bootstrapping resulted in almost equally good c-indices, which is nowadays considered as a reliable estimation of external validation and a good argument against the existence of significant overfitting of the data.23,24 As previously stated, PFD and SD models are based on fewer events than the cardiac mortality risk score; therefore, the former are more prone to overfitting than the latter. However, the PFD and SD models are only secondary endpoints of the MUSIC study, while the number of events of the primary endpoint (cardiac mortality, n = 213) is clearly enough to avoid significant overfitting in an 8-predictor model, built from 20 candidate variables. Nevertheless, further studies are needed with validation of the MUSIC cardiac mortality risk score in other CHF cohorts to confirm its value as a generalizable clinical prediction tool.

In conclusion, this independent, non-interventional study in an outpatient CHF population showed the ability to predict mortality in CHF patients using a simple score including a limited number of predictors that can be easily applied in clinical practice. None of these predictors require invasive or expensive procedures, and all are unequivocally objective variables. The use of this model identifies a subgroup of high-risk patients who should be closely managed.


Grant no. G03/078 from the Instituto de Salud Carlos III, Madrid, Spain.

Conflict of interest: none declared.


Scientific advisors: We thank Drs Eugene Braunwald, Günter Breithardt, Nabil El-Sherif, Valentin Fuster, Sidney Goldstein, Juan Carlos Kaski, Arthur J. Moss, Gaetano Thiene, and Wojciech Zareba for their support and advice during the entire project and the critical review of the manuscript.

Clinical investigators participating in MUSIC Study: Valme Hospital: Juan Leal del Ojo, Antonio Fernández, Dolores García-Medina; Santiago de Compostela Hospital: Pilar Mazón; Son Dureta Hospital: Miquel Fiol, Carlos Fernández; Arrixaca Hospital: Mariano Valdés; Gregorio Maranon Hospital: Roberto Muñoz, Jesús Almendral, Marta Domínguez; Joan XXIII Hospital: Alfredo Bardají, Pilar Valdovinos; Insular Las Palmas Hospital: Vicente Nieto, Ricardo Huerta. Sant Pau Hospital: Agustina Castellví-Grisó, Maite Domingo, Mariana Noguero.


View Abstract