Medical History for Prognostic Risk Assessment and Diagnosis of Stable Patients with Suspected Coronary Artery Disease

Abstract

Aims

To develop a clinical cardiac risk algorithm for stable patients with suspected CAD based upon angina typicality and CAD risk factors.

Methods and Results

Between 2004 and 2011, 14,004 adults with suspected CAD referred for cardiac imaging were followed: 1) 9,093 patients for CCTA (CCTA-1) followed for 2.0 years; 2) 2,132 patients for CCTA (CCTA-2) followed for 1·6 years, and 3) 2,779 patients for exercise myocardial perfusion scintigraphy followed for 5.0 years. A best-fit model from CCTA-1 for prediction of death or myocardial infarction (MI) was developed, with integer values proportional to regression coefficients. Discrimination was assessed using C-statistic. The validated model was also tested for estimation of the likelihood of obstructive CAD, defined as ≥50% stenosis, as compared to method of Diamond and Forrester (D-F).

Primary outcomes included all-cause mortality and non-fatal MI. Secondary outcomes included prevalence of angiographically obstructive CAD. In CCTA-1, best-fit model discriminated individuals at risk of death or MI (C-statistic 0·76). The integer model ranged from 3-13, and corresponded to 3-year death risk or MI of 0·25% to 53·8%. When applied to the CCTA-2 and MPS, the model demonstrated C-statistics of 0·71 and 0·77. Both best-fit (C=0·76, 95% CI 0·746-0·771) and integer model (C=0·71, 95% CI 0·693-0·719) performed better than D-F (C=0·64; 95% CI, 0·628-0·659) for estimating obstructive CAD.

Conclusions

For stable symptomatic patients with suspected CAD, we developed a history-based method for prediction of death and obstructive CAD.

INTRODUCTION

Medical history-based assessment of symptomatic stable patients with suspected coronary artery disease (CAD) has relied upon estimating the likelihood of obstructive CAD rather than the risk of clinical events, such as death or myocardial infarction (MI) (1-5). Recent data have challenged these diagnostic algorithms, and suggest that they overestimate the prevalence of obstructive CAD in patients referred for non-invasive and invasive testing (6-8). The absence of a well-validated pre-test method for determining risk of CAD events, as well as an accurate method for estimating the likelihood of obstructive CAD, may evoke overutilization of testing for individuals at low risk for incident clinical events or obstructive CAD (9). We sought to determine whether information acquired through medical history taking alone could predict risk of myocardial infarction or death and obstructive CAD in symptomatic stable individuals with suspected CAD.

METHODS

Study Participants

This study consisted of 3 distinct non-overlapping cohorts of symptomatic stable individuals with suspected but without prior history of CAD who were referred for non-invasive cardiac imaging between 2004-2011: 1) a development cohort of 9,093 patients from 8 centers and 6 countries referred for coronary CT angiography (CCTA-1) followed for death or MI; 2) a CCTA validation cohort of 2,132 patients from 4 centers and 4 countries referred for CCTA (CCTA-2) followed for death or MI; and 3) a validation cohort of 2,779 patients from 1 center referred for exercise stress myocardial perfusion scintigraphy (MPS) followed for death.

The CCTA-1 derivation cohort comprised patients referred for CCTA from Phase I of the CONFIRM registry (Coronary CT Angiography EvaluatioN For Clinical Outcomes: An InteRnational Multicenter Registry), which has been previously described (10). Among 12 participating sites, 4 were excluded due to lack of information related to symptoms or follow-up data for non-fatal MI, leaving 17,226 patients from 8 sites in 6 countries that constituted the derivation dataset for the clinical model. Sites included 3 from the United States (N=5727), and 1 each from Canada (N=2171), Germany (N=1785), Italy (N=1895), South Korea (N=4912), and Switzerland (N=736). In sequential order, we excluded patients with prior coronary revascularization or MI (n=1110); who were asymptomatic (n=4389); who did not have complete symptom and/or CAD risk factor collection (n=1959); who were referred for imaging from the emergency department or inpatient setting (n=604); or who were lost to follow-up (n=71). After exclusions, 9,093 patients comprised the final study dataset. The median follow-up time for CCTA-1 was 2.0 years (Interquartile Range [IQR] = 1.4 to 3.1 years).

The CCTA-2 validation cohort consisted of patients from Phase II of CONFIRM, detailing identical elements to Phase I CONFIRM and prospectively designed specifically to serve as a validation dataset to Phase I CONFIRM findings (10). Among the 5 participating sites, 1 was excluded due to lack of follow-up information for MI, resulting in 3,996 patients from the United States (N=1016), Canada (N=636), South Korea (N=1404) and Austria (N=940). In sequential order, we excluded patients with prior coronary revascularization or MI (n=250); who were asymptomatic (n=1072); who did not have complete symptom and/or CAD risk factor collection (n=542); who were referred for imaging from the emergency department or inpatient setting (n=0); or who were lost to follow-up (n=0). 2,132 patients comprised the dataset, and are herein referred to as CCTA-2. The median follow-up time for CCTA-2 was 1·6 years (IQR = 0·8 to 2·8 years).

The MPS validation cohort was referred for exercise myocardial perfusion scintigraphy (MPS) at the Cedars-Sinai Medical Center and followed for all-cause mortality. Excluded patients include those undergoing prior coronary revascularization or MI or who underwent pharmacologic stress testing (n=6405); who were asymptomatic (n=1095); who did not have complete symptom and/or CAD risk factor collection (n=425); who were referred for imaging from the emergency department or inpatient hospital setting (n=1903); or who were lost to follow-up (n=0). 2,779 comprised the final study dataset and are herein referred to as the MPS validation cohort. The median follow-up time for MPS was 5·0 years (IQR 3·5 to 6·4 years).

Coronary Artery Disease Risk Factor and Symptom Type

Prior to testing, we performed medical histories to ascertain the presence of categorical CAD risk factors, which were defined in accordance to accepted guidelines. Hypertension was defined by a history of blood pressure of ≥140/90 mmHg and/or treatment with anti-hypertensive medications (11). Diabetes mellitus was defined by a fasting glucose of 126 mg/dl or greater and/or use of insulin or oral hypoglycemic agents (12). Dyslipidemia was defined in accordance with National Cholesterol Education Program Adult Treatment Panel III guidelines or by treatment with lipid lowering medication (13). Current smoking status was defined by active smoking within 3 months of presentation. Family history of CAD was defined as MI or cardiac death in a first-degree relative (14).

Symptoms were ascertained through interview by a physician or health professional, or by written questionnaire. Chest pain was categorized according to criteria for angina pectoris (3, 15). Patients with typical angina experienced substernal jaw, and/or arm pressure-like pain that occurred with exertion and resolved within 15 minutes of rest and/or use of nitroglycerin. Patients with atypical angina experienced 2 of these characteristics. Patients with non-anginal chest pain experienced 1 or none of these characteristics. Dyspneic patients without chest pain were categorized as having typical angina, in accordance to their prognostic risk (16). From these data, the pre-test likelihood of obstructive CAD was calculated by the method of Diamond and Forrester (3).

Follow-up and Event Ascertainment

All study individuals for CCTA-1 and CCTA-2 were followed for an endpoint of death or non-fatal myocardial infarction, and for an endpoint of death for MPS. Follow-up procedures were approved by all study centers’ institutional review boards. Death status for non-US centers was gathered by clinical visits, telephone contacts and questionnaires sent by mail; with verification of all reported events by hospital records or direct contact with a patient’s attending physician. Death status for US centers was ascertained by query of the Social Security Death Index or by scripted in-person or phone interview by an experienced physician and/or nurse study investigator.

General Statistical Analyses

Categorical variables are displayed as frequencies and percents, while, continuous variables are described as mean ± standard deviations, or medians with interquartile ranges. Variables were compared with chi-squared statistic for categorical variables and by Student’s unpaired t-test or Wilcoxon non-parametric test where appropriate for continuous variables. A two-tailed p value <0.05 was considered statistically significant. Analyses were conducted using SAS version 9.2 (SAS Institute, www.sas.com, Cary, NC).

Development of a Medical History-Based Model for Adverse Clinical Events

Information from the CCTA-1 cohort was used to develop a clinical prediction model for death or MI (CCTA-1) or death (MPS) based on CAD risk factors and symptom type. The best overall clinical prediction algorithm was fit using Cox proportional hazards models. Eight variables were evaluated for risk prediction, and were restricted to those that could be easily obtained from a standard cardiac medical history: age, gender, diabetes, hypertension, dyslipidemia, family history of CAD, current smoking status and symptom type. Symptom type was categorized dichotomously in accordance to prognostic relevance for individuals referred for non-invasive imaging as typical angina or dyspnea versus atypical angina or non-anginal pain (16). The final best-fit Cox model was selected by applying backwards stepwise regression, examining the - 2loglikelihood, and minimizing Bayes’ Information Criterion (BIC) (17). Interactions between age and all risk factors were examined.

Once variables were selected for the final Cox model, an integer-based model was developed to predict the 3-year probability of death or MI. The integer-based model was created by transformation of the Cox regression coefficients from the best-fit model. To maximize ease of use, risk score values were scaled and transposed such that age points related to the decade of a patient’s age (e.g., a 5 for patients aged 50-59) and each CAD risk factor present added an additional point to the total score.

Validation of the Medical History-Based Model

External validation was conducted for the best-fit model using the CCTA-2 and MPS cohorts. For comparison, a best-fit Cox regression model was constructed based on the same candidate variables as in the CCTA-1 model but using data from the validation cohorts. Similar to the CCTA-1 model, backwards stepwise regression was used to create the final best-fit model. For MPS, we adjusted the predicted risk based upon ratio of risk of death to risk of death or MI from the CCTA-1 best-fit model. The accuracy of the CCTA-1 best-fit model was examined using discrimination and calibration. Discrimination was evaluated by C-statistic, while calibration was described by Nam and D’Agostino’s modification of the Hosmer-Lemeshow goodness of fit procedures (18). We further examined discrimination and calibration of the integer-based model (Table 1). Observed versus predicted risk was computed based on categories defined by deciles of predicted risk, and compared by plotting predicted and actual event rates within each decile.

Table 1.

Risk Factor	Categories	Points
Age
	18-39	3
	40-49	4
	50-59	5
	60-69	6
	Greater than 70	7
Sex
	Male	1
	Female	0
Symptom
	Non-Typical	0
	Typical	1
Diabetes
	Non-Diabetic	0
	Diabetic	1
Hypertension
	Normotensive	0
	Hypertensive	1
Family History of CAD
	No Family
	History	0
	Family History	1
Current Smoking
	Non-Smoker	0
	Current Smoker	1

Discrimination of the Medical History-Based Model for Obstructive Coronary Artery Disease

We evaluated the ability of the integer-based risk model to discriminate between individuals with and without obstructive CAD, as defined by a ≥50% luminal diameter stenosis in any coronary artery ≥1·5 mm in diameter. Obstructive CAD was determined from CCTAs from the CCTA-1 and CCTA-2 cohorts, but was not available within the MPS validation cohort. CCTAs were performed and interpreted by Level III experts in an intent-to-diagnose fashion (19, 20).

We applied the integer-based model from CCTA-1 to estimate the probability of obstructive CAD in CCTA-2. For obstructive CAD, the area under the receiver operator characteristics curve of the integer-based model was directly compared to the method of Diamond and Forrester. Further, we used the aggregate data from CCTA-1 and CCTA-2 to determine the rate of obstructive CAD by integer values from the developed model.

RESULTS

Clinical Characteristics of the Study Cohorts

Baseline characteristics of the CCTA-1 development and CCTA-2 and MPS validation cohorts are listed in Table 2. During follow-up, 65 deaths and 155 MIs occurred in the CCTA-1 cohort, while 14 deaths and 90 MIs occurred in the CCTA-2 validation cohort. In the MPS validation cohort, 51 deaths occurred during follow-up.

Table 2. Baseline characteristics of the study groups.

Characteristic	CCTA-1	CCTA-2	P-value (CCTA-1 vs. CCTA-2)	MPS	P-value (CCTA-1 vs. MPS)
N	9093	2132	-	2779
Age, mean 26 [years]	56·5 (11·9)	59·3 (12·3)	<0·0001	55·9 (11·6)	0·0098
Male Sex	50·0%	53·9%	0·0012	54·8%	<0·0001
Classic Symptoms	41·5%	52·7%	<0·0001	28·5%	<0·0001
Diabetes	13·5%	34·2%	<0·0001	10·2%	<0·0001
Hypertension	49·3%	60·5%	<0·0001	48·3%	0·3289
Dyslipidemia	55·8%	41·7%	<0·0001	56·2%	0·7156
Family History	31·3%	30·8%	0·5994	17·7%	<0·0001
Current Smoker	16·1%	21·7%	<0·0001	5·6%	<0·0001
Median follow-up (IQR)	2·0 (1·4-3·1)	1·6 (0·8-2·8)	<0·0001	5·0 (3·5-6·4)	<0·0001
D-F Pre-test LK
Low (<15)	22·8%	12·9%	<0·0001	8·9%	<0·0001
Intermediate (15-85)	66·5%	60·2%		85·1%
High (>85)	10·7%	26·9%		6·0%

Clinical Prediction Model Derivation

In univariate Cox regression applied to CCTA-1, several CAD risk factors and symptom type predicted risk of death or MI (Appendix). Despite its lack of significance in univariate models, gender was forced into multivariate models given its accepted clinical importance. As no significant interactions were identified, the final best-fit Cox model for prediction of death or MI included age, gender, symptom type, diabetes, hypertension, family history of CAD and smoking status. The BIC value for the best-fit Cox model was 3554·4, which represented a substantial improvement over the null model BIC of 3716.6. The β coefficients, standard errors, hazard ratios and p-values for each of the covariates in the best-fit Cox model are shown in Table 3. The C-statistic for this model was 0·76.

Table 3. Multivariable Cox Proportional Hazards Models for Prediction of 3-Year Risk of Death or Myocardial Infarction (3-year survival = 0.9711).

		Multivariate Best-Fit Cox Model
Parameter	Effect	β	HR (95% CI)	SE	P value
Age	Age in Years	0.06590	1.07 (1.05-1.08)	0.007	<0.0001
Sex	Male vs. Female	0.38154	1.46 (1.11-1.93)	0.140	0.0065
Symptom	Typical vs. Non-Typical	0.51894	1.68 (1.28-2.21)	0.141	0.0002
Diabetes	Yes vs. No	0.33708	1.40 (1.01-1.94)	0.165	0.0412
Hypertension	Yes vs. No	0.4861	1.63 (1.21-2.18)	0.150	0.0012
Family history of CAD	Yes vs. No	0.51748	1.68 (1.26-2.23)	0.145	0.0004
Smoking status	Current vs. Not Current	0.56197	1.75 (1.25-2.45)	0.171	0.0010

Discrimination and Calibration of the Medical History-Based Prediction Model

Table 4 presents the C-statistics for the best-fit and integer-based Cox model for the CCTA-1, CCTA-2 and MPS validation cohorts. For MPS, for which only death was available during follow-up, probabilities were determined by multiplication of the risk of death or MI by 0·405, which represented the ratio of the 3-year risk of death (1·17%) to the 3-year risk of death or MI (2·89%) in CCTA-1. Discrimination of the best-fit model performed well in both CCTA-2 and MPS, achieving similar discrimination to the CCTA-1 best-fit model. Figure 1 presents the decile plots for assessment of calibration of the CCTA-1 model. Integer models were based upon covariates categorized in binary fashion except for age, which was categorized by range as 18-39 years, 40-49 years, 50-59 years, 60-69 years, and ≥70 years (Table 1).

Table 4. C-Statistic of Prediction Models for 3-Year Risk of Death or Myocardial Infarction.

	CCTA-1 Derivation Cohort	CCTA-2 Validation Cohort	MPS Validation Cohort*
Best-fit Cox Model	0.76 (0.758, 0.762)	0.71 (0.707, 0.713)	0.77 (0.766, 0.774)
Integer-Based Model	0.73 (0.729, 0.733)	0.66 (0.659, 0.666)	0.76 (0.756, 0.764)

^*Risk of death alone

Figure 1.

Calibration Plots for (A) CCTA-1 Development, (B) CCTA-2 Validation and (C) MPS Validation Cohort

Integer scores were calculated by the following formula: Total points = [Age 40-49 years = 4; 50-59 years = 5; 60-69 years = 6; ≥70 years = 7) + Gender (1=male, 0=female) + Diabetes (1=diabetic; 0=non-diabetic) + Hypertension (1=hypertensive; 0=normotensive) + Family history of CAD (1=yes, 2=no).

Discrimination of the Medical History-Based Model for Estimating the Likelihood of Obstructive Coronary Artery Disease

The AUC of the best-fit CCTA-1 model (0·76; 95% CI 0·746 to 0·771) was higher than the Diamond-Forrester (0·64; 95% CI 0·628 to 0·659) methods for discriminating individuals with versus without obstructive CAD (p<0·0001). Across the range of integer values, the likelihood of obstructive CAD ranged from 0 to 82·4% (Table 5).

Table 5. 3-year risk of death or MI and likelihood of obstructive CAD (CCTA-1).

Total Points	Risk of Death or MI over 3 years (%)	Likelihood of obstructive CAD
3	0.25	0
4	0.46	1.4
5	0.76	3.4
6	1.26	5.5
7	2.53	13.2
8	4.53	21.3
9	8.03	31.0
10	15.13	43.2
11	23.29	52.5
12	34.95	82.4
13	53.81

Total Points are the range of integer-based model values. Prevalence of obstructive CAD was the observed rates of obstructive CAD for each value of the integer-based model within the CCTA-1 cohort. Risk score values equal the total of each CAD risk factor as a single point plus points related to the decade of a patient’s age (e.g., 5 for patients aged 50-59):fr

Total points = [Age 40-49 years = 4; 50-59 years = 5; 60-69 years = 6; ≥70 years = 7) + Gender (1=male, 0=female) + Diabetes (1=diabetic; 0=non-diabetic) + Hypertension (1=hypertensive; 0=normotensive) + Family history of CAD (1=yes, 2=no).

DISCUSSION

We developed and validated a method for identifying stable individuals with suspected CAD who are at risk of death or myocardial infarction. Our aim was to create a parsimonious model that incorporates CAD factors that can be derived solely from a medical history. We demonstrated the effectiveness of this method for measures of risk stratification, discrimination, and calibration. Further, the model exhibited superior discriminatory performance when tested against the Diamond-Forrester method for determining pre-test likelihood of obstructive CAD.

One potential function of this model may be to serve as an effective “gatekeeper” to identify individuals who are at sufficiently low risk of incident death or MI that the yield from further testing may be low. At present, more than 10 million CAD imaging tests are performed annually in the United States alone at high direct and indirect costs, with rates of test normalcy significantly higher today as compared to only a decade ago (9,21). As an example, in the largest CCTA study to date, more than 42% of individuals were identified as having no CAD (22). Similarly, the rates of test normalcy for MPS have dramatically increased over time, with a prevalence of >90% in a study of 39,515 patients (21).

In this study, those with an integer value ≤7—which represented nearly half of the study population—only 1·0% experienced an adverse clinical event, and only 10% of these individuals possessed obstructive CAD. These findings may convey a greater sense of prognostic and diagnostic certainty towards avoidance of testing for individuals whose risk and prevalence of disease is sufficiently low to preclude the need for further evaluation. Conversely, this model may be used to identify individuals whose risk is high, and who may benefit from further testing. While representing only 3% of the combined cohort, 9% of individuals with an integer score >9 experienced an event and 54% possessed obstructive CAD. The need for testing is apparent in this group, and in accordance with societal guidance documents (23).

To date, no robust medical history-based prediction model has been available for identifying stable symptomatic patients at risk for death or MI. In its absence, clinicians have relied upon estimates of the likelihood of an individual possessing obstructive CAD (3). These methods have served as the cornerstone for assessment of patients with suspected CAD for the last 40 years. Nevertheless, these methods were validated against invasive angiographic and pathologic correlates; and their application demonstrates a 3-fold overestimation of CAD in patients referred for non-invasive imaging that suggests the need for contemporary revision (6). In the present study, the developed medical history-based model was superior to the Diamond-Forrester method for discrimination of individuals with versus without obstructive anatomic CAD. In this regard, this model may serve a dual role not only as a prognostic instrument but also as a diagnostic tool.

This study is not without limitations. We studied patients referred for CCTA and exercise MPS—given their common performance and similarities for indications considered appropriate for use—with our study findings robustly applicable to both modalities (24,25). Yet, whether the model can be employed for patients being considered for exercise treadmill testing, other imaging tests, or in patients not considered for any testing remains unknown. Second, patient-reported histories may have resulted in ascertainment bias, and CAD risk factor profiles may have been more accurate with adjunctive confirmation by vital sign or laboratory values. One advantage of the present study is the applicability to practicing physicians who routinely solicit histories in the office setting. Third, the prediction model included all-cause rather than cardiac-specific death. In doing so, this study is disencumbered from information bias but may result in reduced prognostic specificity. Finally, we examined stable symptomatic individuals referred for outpatient imaging, and caution should be taken to apply the model to patients in the emergency department or inpatient setting.

CONCLUSIONS

We developed and validated a simple medical-history based method that predicts risk of death and myocardial infarction in symptomatic stable patients with suspected CAD. Further, this method exhibits superior performance to traditional methods for identifying individuals with obstructive CAD.

TRANSLATIONAL PERSPECTIVE.

A simple and easy-to-develop model derived solely from a patients’ medical history is useful in identifying stable individuals with suspected CAD who are at risk of death or myocardial infarction. This history-based model may potentially serve a useful purpose as an effective “gatekeeper” towards avoidance of testing for individuals whose risk and prevalence of disease is sufficiently low to preclude the need for further evaluation. Conversely, this model may be used to identify individuals whose risk is high, and who may benefit from further testing.

ABBREVIATIONS

CAD

coronary artery disease

MPS

myocardial perfusion scintigraphy

MI

Myocardial infarction

CCTA

coronary CT angiography

D-F

Diamond and Forrester

CONFIRM

Coronary CT Angiography EvaluatioN For Clinical Outcomes: An InteRnational Multicenter Registry

BIC

Bayes’ Information Criterion