Usefulness of the Duke Sudden Cardiac Death Risk Score for Predicting Sudden Cardiac Death in Patients With Angiographic (>75% Narrowing) Coronary Artery Disease

Article Outline

• Abstract
• Methods
• Results
• Discussion
• Acknowledgment
• Appendix: Candidate Variables for Adjusted Models and Definitions
• References
• Copyright

The currently available sudden cardiac death (SCD) risk prediction tools fail to identify most at-risk patients and cannot delineate a specific patient's SCD risk. We sought to develop a tool to improve the risk stratification of patients with coronary artery disease. Clinical, demographic, and angiographic characteristics were evaluated among 37,258 patients who had undergone coronary angiography from January 1, 1985 to May 31, 2005, and who were found to have at least one native coronary artery stenosis of ≥75%. After a median follow-up of 6.2 years, SCD had occurred in 1,568 patients, 14,078 patients had died from other causes, and 21,612 patients remained alive. A Cox proportional hazards model identified 10 independent patient characteristic variables significantly associated with SCD. A simplified model accounting for 97% of the predictive capacity of the full model included the following 7 variables: depressed left ventricular ejection fraction, number of diseased coronary arteries, diabetes mellitus, hypertension, heart failure, cerebrovascular disease, and tobacco use. The Duke SCD risk score was created from the simplified model to predict the likelihood of SCD among patients with coronary artery disease. It was internally validated with bootstrapping (c-index = 0.75, chi-square = 1,220.8) and externally validated in patients with ischemic cardiomyopathy from the Sudden Cardiac Death Heart Failure Trial (SCD-HeFT) database (c-index = 0.64, chi-square = 14.1). In conclusion, the Duke SCD risk score represents a simple, validated method for predicting the risk of SCD among patients with coronary artery disease and might be useful for directing treatment strategies designed to mitigate the risk of SCD.

To improve sudden cardiac death (SCD) risk stratification among patients with coronary artery disease (CAD), we sought to develop a simple, validated risk score from a large population of patients with a broad range of SCD risk by incorporating clinical characteristics, angiographic findings, and ejection fraction (EF) values into a multivariable model.

Back to Article Outline

Methods 

The design of the Duke Databank for Cardiovascular Disease has been previously described.¹, ² In brief, demographic, clinical, angiographic, and laboratory information have been collected prospectively as a part of routine patient care for patients undergoing cardiac catheterization at Duke University Medical Center in Durham, North Carolina. The demographic data, including race, were self-reported. A history of clinical characteristics such as hypertension, diabetes, or cerebrovascular disease was determined by a physician or physician assistant after history taking, performing a physical examination, and reviewing the available medical records before the index catheterization. Patients with significant (≥75%) stenosis of at least one coronary artery underwent examination at 6 months and 1 year and annually thereafter by a mailed questionnaire with telephone back up. The National Death Index was used regularly to verify the vital status of nonresponders. The follow-up information was complete for 97.4% of those patients enrolled in the Duke Databank at all intervals.³

For the present study, we included patients aged ≥18 years who had undergone cardiac catheterization from January 1, 1985 to May 31, 2005, and who had a diameter stenosis of ≥75% of at least one native coronary artery (n = 39,052). We excluded patients who had no cause of death information available (n = 5) and those whose SCD risk was altered by previous implantable cardioverter-defibrillator (ICD) implantation (n = 1,789) for a final cohort of 37,258 patients.

Patient deaths were identified using the National Death Index, the Social Security Death Index, and reviews of annual mail and telephone surveys. The cause of death was classified using the following sources: telephone interviews with members of the patient's family, death certificates, National Death Index plus (including International Classification of Diseases, ninth revision, codes), and hospital discharge summaries. Independent committees consisting of 2 experienced data abstractors and a faculty cardiologist adjudicated the final cause of death after a review of all available data. Witnessed deaths occurring instantaneously or <60 minutes from symptom onset (n = 892), unobserved deaths occurring unexpectedly in a previously asymptomatic patient (n = 513), and deaths occurring after attempted resuscitation (n = 163) were classified as SCD. Other causes of death included definite, probable, or possible myocardial infarction, heart failure (HF), death during or after cardiac surgery, death during cardiac catheterization, death during interventional cardiac catheterization, other cardiac cause, vascular, trauma, noncardiac medical, noncardiac procedural, and unknown. The cause of death information was available for all deceased patients included in the present analysis (n = 15,646).

The differences in baseline characteristics between groups were assessed with Pearson chi-square tests for categorical variables and Wilcoxon log-rank tests for continuous variables. Continuous measures were described using the median and twenty-fifth and seventy-fifth percentiles, and categorical variables are reported as percentages. A p value of <0.05 was considered significant for all statistical test results.

The cumulative probability of SCD was calculated using the Kaplan-Meier method. The unadjusted rates of SCD were reported at various prespecified points. A Cox proportional hazard modeling procedure was performed for the interval-to-SCD, censoring those who had died from other causes and those remaining alive. The Cox model's assumption of linearity of dependent variables relative to the outcome was tested, and continuous and ordinal variables that were in violation were transformed as needed to satisfy this model requirement. The candidate variable list included covariables that were statistically significant in univariable models, in addition to those known to have clinical relevance for SCD. The final pool of candidate variables and their definitions are included in the Appendix. EF data were missing and imputed for 122 patients with SCD (7.8%) and 4,402 patients without SCD (12.3%), with a multiple imputation procedure using the Statistical Analysis Systems procedure “PROC MI” (SAS Institute, Cary, North Carolina).⁴ To incorporate appropriate variability across imputations, missing values were replaced with a set of plausible values after 5 imputations were generated and compared. Because of the relatively small number of missing values and the arbitrary missing pattern, a Markov Chain Monte Carlo was used.

Using a significance level of 0.05, the full Cox model's forward stepwise selection results consisted of EF, number of diseased coronary arteries, diabetes, hypertension, history of HF, history of cerebrovascular disease, history of tobacco use, renal insufficiency, age, and history of myocardial infarction. To improve the clinical usefulness of the multivariable model and the derived risk score, the full stepwise model was simplified by excluding all variables with a chi-square of ≤10 in a stepwise fashion. The discriminative capacity of the full and simplified multivariable models were assessed using the area under the receiver operating characteristic curve as an index of model performance (c-index).⁵, ⁶ A calibration plot comparing the expected and observed outcomes was created to assess the ability of the simplified model to predict probabilities across all ranges of risk. The reliability of the simplified multivariable model was evaluated using a bootstrapping technique. One hundred samples were withdrawn from the original data set with replacement, and the stepwise selection process was repeated for each sample. The models developed in the bootstrap samples were then tested in the original data set to assess the internal validity of the c-index. The discriminative capacity of the simplified multivariable model was then externally validated by calculating a c-index for prediction of SCD in patients with ischemic cardiomyopathy who were randomized to the control group (no ICD implantation and no antiarrhythmic therapy) in the Sudden Cardiac Death Heart Failure Trial (SCD-HeFT).⁷ The Duke SCD risk score was developed from the simplified multivariable model by assigning weighted scores to the predictive variables according to the model's variable coefficients. All statistical modeling was done using Statistical Analysis Systems, version 8.2 (SAS Institute). The institutional review board at Duke University Medical Center approved the present analysis, and the requirement for individual informed consent was waived.

Back to Article Outline

Results 

The baseline characteristics, stratified by patient outcome, are listed in Table 1. The risk of SCD accrued gradually with time, as did the risk of death from other causes (Figure 1). The risk of death from non-SCD causes was proportionally greater than the risk of SCD throughout the follow-up period. During a median follow-up of 6.2 years (interquartile range 2.5 to 11), 1,568 patients died from SCD (4.2%), 14,078 patients died from other causes (37.8%), and 21,612 patients remained alive (58.0%). The median interval from the index cardiac catheterization to SCD was 4.5 years (interquartile range 1.4 to 8.1), and the median interval from the index cardiac catheterization to non-SCD death was 4.8 years (interquartile range 1.4 to 9.0).

Table 1. Baseline clinical, demographic, and angiographic characteristics

Variable	Alive (n = 21,612)	SCD (n = 1,568)	Noncardiac Death (n = 5,175)	Cardiac/Non-SCD (n = 8,903)
Patient presentation
Age (years)	59(51,68)	61(53,69)	67(60,73)	67(59,74)
Systolic blood pressure (mm Hg)	134(119,153)	130(112,146)	130(115,150)	130(112,150)
Body mass index (kg/m2)	27(25,31)	27(24,30)	26(23,29)	26(23,30)
White race	81.5%	82.0%	84.6%	82.8%
Medical history
Hypertension	56.1%	62.7%	57.2%	63.6%
Diabetes mellitus	21.5%	32.5%	28.3%	33.7%
Heart failure	12.3%	28.1%	21.4%	33.0%
Tobacco use	59.1%	68.6%	69.1%	58.2%
Previous myocardial infarction	47.4%	64.6%	53.5%	60.8%
Cerebrovascular disease	6.6%	13.3%	12.8%	15.1%
Peripheral vascular disease	6.2%	15.8%	15.1%	16.6%
No heart failure	86.9%	71.8%	78.3%	66.8%
NYHA class I	3.1%	8.0%	5.2%	8.1%
NYHA class II	4.0%	6.1%	5.3%	7.3%
NYHA class III	4.0%	7.2%	6.7%	9.0%
NYHA class IV	2.1%	6.8%	4.5%	8.7%
Coronary artery disease duration (mo)	2.7(0.2,37.8)	25.3(0.8,101.0)	11.3(0.5,79.6)	20.9(0.6,94.0)
Renal insufficiency	0.8%	1.3%	2.1%	1.4%
Previous percutaneous coronary intervention	3.9%	3.8%	4.1%	3.4%
Liver disease	0.3%	0.2%	0.4%	0.2%
Chronic obstructive pulmonary disease	3.5%	5.4%	10.0%	6.0%
Metastatic cancer	0.1%	0.1%	0.4%	0.1%
Solid tumor	1.5%	0.5%	2.1%	1.1%
Coronary artery bypass grafting	9.6%	13.1%	9.4%	13.6%
Cardiac catheterization details
Acute myocardial infarction at index catheterization	25.9%	23.8%	20.1%	25.0%
Any myocardial infarction within 6 weeks of index catheterization	38.6%	43.2%	36.6%	42.8%
Angiographic findings
No. of diseased coronary arteries (%)
1	49.8%	28.5%	38.3%	28.4%
2	26.3%	29.6%	28.9%	28.2%
3	23.4%	41.2%	32.1%	42.9%
Ejection fraction⁎ (%)	56.4%(47.1,64.0)	44.0%(33.0,55.0)	53.0%(43.0,62.0)	46.4%(34.0,57.8)

Continuous measures are presented as medians (twenty-fifth, seventy-fifth percentiles), and categorical variables are presented as percentages.

NYHA = New York Heart Association; SCD = sudden cardiac death.

• 
  • View Large Image
  • Download to PowerPoint

• Figure 1. 

  Event curves. Kaplan-Meier event curves stratified by SCD versus all other outcomes.

When all candidate variables were considered simultaneously in a Cox proportional hazards analysis, 10 variables remained as significant predictors of SCD (Table 2). After eliminating the variables with chi-square values of ≤10, the 7 variables retained in the simplified Cox proportional hazards model were EF, number of diseased coronary arteries, diabetes, hypertension, HF, cerebrovascular disease, and tobacco use (Table 3). The simplified Cox model retained 97% of the overall prognostic information from the full multivariable model (evaluated as a ratio of the summed Wald chi-square statistics from the simplified model compared to the full model). The mean bootstrap validated c-index value of the simplified model was 0.75. The model calibration curve is shown in Figure 2. When the simplified model was externally validated in the patients with ischemic cardiomyopathy randomized to the control arm of the SCD-HeFT registry (n = 748), the global chi-square was 14.1 and the mean c-index value was 0.64.

Table 2. Full Cox proportional hazards sudden cardiac death (SCD) model

Global likelihood ratio chi-square = 1,124.23, c-index = 0.75.

Table 3. Simplified Cox proportional hazards sudden cardiac death (SCD) model

Global likelihood ratio chi-square = 1,104.2; bootstrap validation c-index = 0.75.

Variables with Wald chi-square <10 were eliminated.

• 
  • View Large Image
  • Download to PowerPoint

• Figure 2. 

  Calibration plot. Calibration plot of simplified multivariable model comparing predicted and observed probabilities of SCD-free survival.

We generated a risk score nomogram from the simplified multivariable model to predict the risk of SCD for an individual patient. The method for calculating a Duke SCD risk score is provided in Table 4. For an individual patient, the probability of SCD can be estimated by calculating the point total according to the presence or absence of the risk factors (Table 4) and then correlating that point total with the estimated 1-, 3-, 5-, 7-, or 10-year incidence of SCD (Table 5).

Table 4. Calculate the Duke sudden cardiac death (SCD) risk score by adding points for all risk factors

Risk Factor	Points
EF (%)
1–5	56
6–10	51
11–15	46
16–20	42
21–25	37
26–30	32
31–35	28
36–40	23
41–45	19
46–50	14
51–55	9
56–60	5
>61	0
No. of diseased arteries
1	0
2	6
3	13
Cerebrovascular disease	8
Heart failure	7
Diabetes	7
Hypertension	6
Tobacco use	4

EF = ejection fraction; SCD = sudden cardiac death.

Table 5. Correlate the Duke sudden cardiac death (SCD) risk score with the estimated 1-, 3-, 5-, 7-, or 10-year SCD risk

SCD = sudden cardiac death.

The observed patient outcomes stratified by the baseline Duke SCD risk score quartile are listed in Table 6. The proportion of patients with SCD and death by other causes increased across the SCD risk score quartiles. Assuming ICD implantation reduces the risk of tachyarrhythmic death by 60% in at-risk patients,⁸ the predicted number needed to treat to prevent one SCD using a prophylactic ICD implantation strategy decreased incrementally across the SCD risk quartiles. In the greatest SCD risk quartile, 18 patients would have required prophylactic ICD therapy to prevent one SCD.

Table 6. Distribution of patient outcomes by Duke sudden cardiac death (SCD) risk score quartile

Duke SCD Risk Score	Patients (n)	Alive (%)	SCD (%)	Other Deaths (%)	NNT to Prevent One SCD
0–14	9,570	7,404(77.4%)	138(1.4%)	2,028(21.2%)	118
15–23	9,230	6,189(67.1%)	265(2.9%)	2,776(30.1%)	57
24–36	9,534	5,191(54.4%)	403(4.2%)	3,940(41.3%)	38
37–101	8,924	2,828(31.7%)	762(8.5%)	5,334(59.8%)	18

Median follow-up 6.2 years.

NNT = number needed to treat; SCD = sudden cardiac death.

Back to Article Outline

Discussion 

We used the predictive information from a multivariable analysis of a large database of patients with significant CAD to develop a simple, clinical risk score for the estimation of long-term SCD risk. By combining the predictive power of multiple clinical, demographic, and angiographic variables, the Duke SCD risk score is the first validated tool capable of prospectively identifying patients with CAD and preserved or mildly depressed EF who are at an increased risk of SCD.

Risk models used for patient risk stratification should meet a set of standards before their general application to patient care.⁶ First, the patient cohort from which a model has been derived should be representative of the population to which it will be applied. The Duke SCD risk score was created in a large patient cohort with obstructive CAD. The ability of the risk score to predict SCD in patients without CAD remains untested; therefore, the Duke SCD risk score should only be applied to patients with significant CAD. Second, the model must provide accurate prediction across a wide range of patient risk. The Duke SCD risk score provided accurate prediction across a broad range of patient risk (Figure 2). Finally, the model should be validated either internally using a bootstrap technique or externally in a separate, but similar, patient cohort. The simplified model used to generate the Duke SCD risk score was validated internally using the bootstrapping technique, but because no patient cohort with characteristics similar to the patients followed in the Duke Databank was available for external validation, we used the portion of the control population in the SCD-HeFT registry with ischemic cardiomyopathy. These patients differed significantly from our derivation cohort with respect to the frequency of previous HF (100% in SCD-HeFT vs 19% in our derivation cohort) and EF (median EF value of 25% in SCD-HeFT vs 54% in our derivation cohort).

Accurate SCD risk estimation facilitates the development of informed, cost-effective preventive strategies to reduce the risk of SCD across broad patient populations. To make the Duke SCD risk score useful for the discrimination of patients likely to benefit from therapies specifically designed to treat SCD, the analysis was designed to test the ability of candidate variables to discriminate death by SCD from all other possible outcomes, including survival and death from other causes. Patients at high risk of SCD using the Duke SCD risk score could be at considerable risk of death from other causes, and appropriate consideration of all clinical variables, especially patient co-morbidities, should be used in determining the appropriate therapeutic strategies for specific patients. The Duke SCD risk score has several potential advantages over a single dichotomous EF cutoff for the prediction of SCD. First, the Duke SCD risk score provides detailed risk estimates rather than a binary distinction of high or low risk. These accurate risk assessments could potentially facilitate more informative discussion between clinicians and patients on the risks and benefits of strategies designed to prevent SCD, including the implantation of prophylactic ICDs. Second, by capturing the predictive power of multiple variables, the Duke SCD risk score can identify high-risk patients with a relatively preserved EF, potentially improving the outcomes for a group of patients not currently eligible for ICD implantation. Randomized controlled investigations of prophylactic ICD implantation for the primary prevention of SCD might be warranted in this high-risk group. Finally, the Duke SCD risk score might identify patients with significant CAD and a depressed EF but no other risk factors, who have a relatively low cumulative risk of SCD and who might not obtain cost-effective benefits from prophylactic ICD implantation.⁹ The ability of the Duke SCD risk score to improve the cost-effectiveness of prophylactic ICD implantation should be studied prospectively.

Several limitations of the Duke SCD risk score should be acknowledged. First, the electrocardiographic and medication data were incomplete in the Duke Databank for Cardiovascular Disease and thus could not be included in the list of candidate variables. The addition of these variables to the multivariable model might have provided some incremental predictive power to the Duke SCD risk score. Second, the EF values were missing for 12% of patients included in the analysis and were imputed to limit the potential confounding that would have occurred by excluding these patients from the analysis. Finally, we used a rigorous method to classify the deaths as SCD versus non-SCD; however, because no universally accepted definition for SCD is available, the Duke SCD risk score should be evaluated in other databases with different methods of classifying SCD events.

Back to Article Outline

Acknowledgment 

We thank Charles B. McCants, BS, and Judith A. Stafford, MS, for their efforts in the maintenance of the Duke Databank for Cardiovascular Disease.

Back to Article Outline

Appendix: Candidate Variables for Adjusted Models and Definitions 

Patient characteristic variables: age (years); New York Heart Association functional class (I–IV); body mass index; carotid bruits; race (as determined by patient or family); gender; Charlson's co-morbidity index (a weighted index that takes into account the number and severity of a variety of co-morbid conditions).

Catheterization variables: elective cardiac catheterization (defined as stable cardiac function in the days or weeks before the procedure and the catheterization procedure could have been delayed without risk of compromised cardiac outcome); urgent cardiac catheterization (defined as unstable cardiac function requiring catheterization during the same hospitalization to minimize risk of additional cardiac deterioration); emergent cardiac catheterization (defined as ongoing ischemic dysfunction, including angina at rest despite maximum medical or intra-aortic balloon pump therapy, acute evolving ST-segment elevation myocardial infarction within 24 hours of catheterization, pulmonary edema requiring intubation, or cardiogenic shock requiring immediate catheterization to minimize additional deterioration of cardiac function).

Angiographic variables: ejection fraction; number of diseased coronary arteries; angiographic mitral insufficiency grade.

Historical variables: chronic obstructive pulmonary disease; previous percutaneous coronary intervention; previous coronary artery bypass grafting; tobacco use (including smokeless tobacco); solid tumor; metastatic cancer; connective tissue or rheumatoid diseases; acute or evolving myocardial infarction within 3 days of index catheterization (defined by elevation of troponin T or I, creatinine kinase-MB, or total creatinine kinase greater than the myocardial infarction decision limit with accompanying ST-segment changes or ischemic symptoms); angina (defined as pain or discomfort in the chest, epigastrium, neck, or jaw consistent with a cardiac etiology); heart failure (defined as previous symptoms of volume overload, including paroxysmal nocturnal dyspnea, orthopnea, cardiac rales, exertional fatigue, peripheral edema, or chest X-ray diagnosis of pulmonary edema occurring outside the setting of an acute coronary syndrome and ≥2 weeks before the current date of admission); myocardial infarction (defined as a previously documented episode of ST-segment elevation or non–ST-segment elevation myocardial infarction ≥8 days before the enrollment admission); renal insufficiency (defined as serum creatinine level >1.9 mg/dl or the need for chronic hemodialysis); dyslipidemia (defined as total cholesterol >200 mg/dl, low-density lipoprotein >130 mg/dl, high-density lipoprotein <30 mg/dl, triglycerides >150 mg/dl, or need for current cholesterol medical therapy); hypertension (defined as any history of hypertension diagnosed and/or treated by a physician with medication or diet and exercise or a systolic blood pressure >140 or a diastolic blood pressure >90 on ≥2 occasions, or currently taking antihypertensive mediations); diabetes (defined as any previous physician diagnosis with treatment with medication or diet or a fasting blood glucose >140 mg/dl on ≥2 blood sample measurements); peripheral vascular disease (defined as claudication either at exertion or at rest, or amputation for arterial insufficiency, or previous vascular surgery for occlusive disease, or positive noninvasive or invasive test results in any arterial bed other than the carotid); cerebrovascular disease (defined by any history of one of the following: unresponsive coma >24 hours, history of stroke, reversible ischemic neurologic deficit, or transient ischemic attack, noninvasive or invasive carotid test demonstrating >75% occlusion, or previous carotid artery surgery or stent placement); moderate liver disease (defined as previous physician diagnosis of liver disease without life-threatening clinical sequelae); severe liver disease (defined as previous physician diagnosis of liver disease with life-threatening clinical sequelae); mild cardiac valvular disease (defined by previous physician diagnosis of any valvular lesion without clinical sequelae).

Back to Article Outline

References 

1. Hlatky MA, Califf RM, Harrell FE, Lee KL, Mark DB, Pryor DB. Comparison of predictions based on observational data with the results of randomized controlled trials of coronary artery bypass surgery. J Am Coll Cardiol. 1988;11:237–245
   • View In Article
   • Abstract
   • Full-Text PDF (629 KB)
   • CrossRef
2. Mark DB, Nelson CL, Califf RM, Harrell FE, Lee KL, Jones RH, et al. Continuing evolution of therapy for coronary artery disease (Initial results from the era of coronary angioplasty). Circulation. 1994;89:2015–2025
   • View In Article
   • MEDLINE
3. Newby LK, Lapointe NM, Chen AY, Kramer JM, Hammill BG, DeLong ER, et al. Long-term adherence to evidence-based secondary prevention therapies in coronary artery disease. Circulation. 2006;113:203–212
   • View In Article
   • CrossRef
4. Heymans MW, Buuren SV, Knol DL, van Mechelen W, de Vet HCW. Variable selection under multiple imputation using bootstrap in a prognostic study. BMC Med Res Methodol. 2007;7:33
   • View In Article
   • CrossRef
5. Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology. 1982;143:29–36
   • View In Article
   • MEDLINE
6. Ohman EM, Granger CB, Harrington RA, Lee KL. Risk stratification and therapeutic decision making in acute coronary syndromes. JAMA. 2000;284:876–878
   • View In Article
   • MEDLINE
   • CrossRef
7. Bardy GH, Lee KL, Mark DB, Poole JE, Packer DL, Boineau R, et al. Amiodarone or an implantable cardioverter-defibrillator for congestive heart failure. N Engl J Med. 2005;352:225–237
   • View In Article
   • CrossRef
8. Packer DL, Bernstein R, Wood F, Boehmer J, Carlson M, Frantz RP, et al. Impact of amiodarone versus implantable cardioverter defibrillator therapy on the mode of death in congestive heart failure patients in the SCDHeFT trial. Heart Rhythm. 2005;2(S38):AB20–AB22
   • View In Article
9. Owens DK, Sanders GD, Heidrenreich PA, McDonald KM, Hlatky MA. Effect of risk stratification on cost-effectiveness of the implantable cardioverter defibrillator. Am Heart J. 2002;144:440–448
   • View In Article
   • Abstract
   • Full Text
   • Full-Text PDF (132 KB)
   • CrossRef