Development of a multivariate model to predict significant coronary artery disease in Thai patients with left ventricular systolic dysfunction and determine the applicability of coronary angiography: a single-center, retrospective, case–control study

After approval of the study protocol by the Central Chest Institute of Thailand Institutional Review Board (approval No. 103/2557), we conducted a retrospective, observational, single-center, case–control study of Thai adult (at least 18 years old) patients with LV systolic dysfunction in our tertiary-care hospital. Data registered at the Central Chest Institute of Thailand from 240 patients with LV systolic dysfunction diagnosed between August 2000 and October 2014 were included. The study included data from patients with LV ejection fraction (LVEF) <50%. Data from patients were excluded if the patient had a history of CAD (stable CAD or acute coronary syndrome), history of revascularization (percutaneous coronary intervention (PCI) or coronary artery bypass grafting (CABG)), history of angina, regional wall motion abnormality from echocardiography, contraindication to antiplatelet therapy, more than a moderate degree of valvular heart disease [12] except mitral regurgitation from annular dilatation, history of current or former use of addictive drugs or substance use (e.g. amphetamines or cocaine), and stress cardiomyopathy, and if patients who refused CAG.

Data from patients were divided into 2 groups according to CAG results. These 2 groups were the group with ICM and the group with NICM. Predictor variables associated with CAD were collected. The outcome variables were the predictors that were found to be significantly associated with CAD in previous studies [1].

LV systolic dysfunction is defined as LVEF <50%. The definition of ICM has been modified from previous 2012 American College of Cardiology Foundation (ACCF)/AHA Guidelines for the Diagnosis and Management of Patients with Stable Ischemic Heart Disease [13] and Felker et al. [1], which included patients with LV systolic dysfunction and ≥50% stenosis of left main or ≥70% stenosis of proximal left anterior descending artery (LAD) or patients with ≥70% stenosis of 2 or more epicardial vessels (LAD, left circumflex artery, or right coronary artery).

Predictors of significant CAD that were included in the present study are age (>60 years at diagnosis); sex (male); smoking habit within 2 years of diagnosis; history of diabetes mellitus (DM; fasting blood sugar >126 mg/dL); hypertension (systolic blood pressure >140 mmHg, diastolic blood pressure >90 mmHg, or taking an antihypertensive drug); dyslipidemia (low-density lipoprotein (LDL) >160 mg/dL or triglyceride >150 mg/dL plus high-density lipoprotein (HDL) <40 mg/dL in men or HDL <50 mg/dL in women); family history of premature CAD (male relative <55 years old, female relative <65 years old); history of other atherosclerotic disease, such as peripheral arterial disease (ankle–brachial index <0.9 or known peripheral artery disease); ischemic stroke; pathologic Q wave in contiguous leads in any ECG; and body mass index (BMI) at the diagnosis >25 kg/m². Only available data were analyzed. All predictors are certain CAD risk factors from previous study [1].

The literature indicates that the overall prevalence of the predictors described earlier in the NICM group are 55% for age, 54% for sex (male), 56.6% for hypertension, 51% for DM, 20.8% for hyperlipidemia, and 51% for smoking [1]. We chose prevalence of hypertension (P₂ = 56.6%) for sample size calculation. We specified 0.05 for type I error, 0.20 for type II error with 80% power, and an odds ratio of 2.

P₂ = NICM group predictor exposure rate

P₁ = ICM group predictor exposure rate

From the formula:

The sample size (N) was calculated using the following formula:

The calculated sample size was data from 73 patients for each group. We increased the sample size by 10% to data from 80 patients for each group in case of missing or incomplete data and set the ratio of ICM to NICM as 1:2. Thus, overall sample size for this study was data from 240 patients (80 patients in the ICM group and 160 patients in the NICM group).

We used PASW Statistics for Windows, version 18.0 (SPSS) for all our analyses. Characteristics of patients from whom data were included are summarized using descriptive statistics, including frequencies and percentages for categorical variables. Continuous variables are reported as mean values, normally distributed variables as standard deviation, and non-normally distributed variables as median, minimum, and maximum. The normality of distribution of the variables was examined using a Kolmogorov–Smirnov test. Differences between normally distributed variables were investigated using a Student t test to compare mean values, and a χ² test was used to compare proportions between 2 groups.

Variables found significantly associated with and affected by CAD in multivariate logistic regression were selected predictors with a base of P < 0.2 in univariate logistic regression, and a backward stepwise selection technique was used to find the best model. Interaction effect between variables was assessed using a Mantel–Hanzel test of homogeneity. Multicollinearity between independent variables was assessed using variance inflation factors (VIFs) before interpreting the final output. For all tests performed, a 2-tailed test result with P < 0.05 was considered as significant. The best model was multiple logistic regression (backward stepwise method).

The significant predictors of CAD were used to calculate CAD risk score. To validate the predictive logistic model, we generated a random sample from the original dataset for a larger database. The size of sample was drawn from the prevalence of CAD and sampling data without replacement. The cutoff score was considered the highest predictive accuracy of CAD for the model.

The cutoff score considered to have the highest predictive accuracy for CAD was identified using a receiver operating characteristic (ROC) curve and presented as sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), likelihood ratio (LR), and accuracy. For all tests performed, a 2-tailed P < 0.05 was considered as significant.

We included data from 240 patients, those with LV systolic dysfunction and who underwent CAG at our tertiary-care center from August 2000 to October 2014. CAG showed significant coronary artery stenosis in 80 patients and normal or nonsignificant stenosis in 160 patients. The mean age of patients was 55.7 ± 12.2 years. Male patients accounted for 70%, 53% of patients had hypertension, and 48% of patients had dyslipidemia. One-third of patients had DM, and 56% had smoked in the previous 2 years or were current smokers at the time of diagnosis. Baseline LVEF was 26.7 ± 8.6%, and LVEF at follow-up visit was 40.4 ± 13.7%. The time between the original and the follow-up LVEF was 755 ± 788 days (range, 6 to 3,782 days). Baseline characteristics are summarized in Table 1.

An initial univariate analysis was conducted. Predictors that were significantly high in those with ICM (P < 0.05) were age (>60 years), pathologic Q wave in contiguous leads, DM, and dyslipidemia. Predictors that were significantly high among those in the non-ICM group were BMI >25 kg/m² and atrial fibrillation. The test of homogeneity showed no interaction between variables. The VIF showed no evidence of a multicollinearity problem among independent variables. After the multivariate analysis, 5 predictors were identified to be significant for discrimination. Predictors that were found associated with increased probability of ICM (risk factors) were age (>60 years old), sex (male), and DM (P < 0.05). Predictors that were found associated with increased probability of insignificant coronary artery stenosis (protective factors) were BMI >25 kg/m² and left bundle branch block (LBBB) on ECG (P < 0.05) as shown in Table 2.

In the present study, we found that the suitable cutoff is 0 (Figure 1, area under the curve = 80.13% (95% confidence interval (CI) 79.35%–80.92%)). Therefore, there is a possibility to find significant coronary artery stenosis if the patient’s calculated CAD score is ê0. The sensitivity and specificity of this score are 60.5% (95% CI 58.9%–62.0%) and 85.1% (95% CI 84.3%–85.9%), respectively, with a PPV 66.8% (95% CI 65.3%–68.4%), NPV 81.3% (95% CI 80.1%–82.1%), LR+ 4.06 (95% CI 3.83–4.30), LR– 0.46 (95% CI 0.45–0.48), and overall accuracy 76.93% (Table 4).

Figure 1

Based on these 5 predictors, we can derive the original discriminating equation and the simplified version to predict the possibility of significant CAD in patients with LV systolic dysfunction as shown in Table 3.

ROC, receiver operating characteristic

We sought to identify predictors associated with significant CAD (stenosis) in patients with LV systolic dysfunction and create a simple equation to differentiate patients with ICM from those with NICM. We can use this classification to determine which patients would benefit from CAG. Predictors for ICM were age (>60 years), sex (male), and DM. Predictors for NICM were BMI >25 kg/m² and LBBB on ECG.

Well-known risk factors for CAD are older age, smoking, DM, hypertension, dyslipidemia, and family history of premature atherosclerotic disease [14, 15]. Advanced age is an important risk for CAD in both men and women, especially in patients aged >60 years [16]. For patients with LV systolic dysfunction, the mean age of patients with ICM (64 years) was significantly more than that of patients with NICM (55 years) [1]. We also found an association between increasing age and CAD. A similar association was found with DM, which is also associated with significant CAD.

From the European Guidelines for Cardiovascular Disease Prevention in Clinical Practice [8] that uses SCORE to determine risks for cardiovascular disease in the general population, the risk factor includes sex and the pooled cohort equation for ASCVD risk score from the ACC/AHA Guidelines on the Assessment of Cardiovascular Risk [9]. We also found a distinction between sexes in the present study. We found that male patients with LV systolic dysfunction have more significant CAD than female patients.

In the present study, patients with BMI >25 kg/m² were more likely to have NICM than ICM. This can be explained by data from previous studies, which indicate that although the best survival in the general population is in those with BMI 20–24.9 kg/m², for patients with chronic diseases, such as chronic kidney disease and chronic heart failure, overweight and obese patients have better survival rate. This phenomenon, known as the obesity paradox [17], was first described by Gruberg et al. [18]. They found that in patients with CAD undergoing PCI, the risk of in-hospital complications and cardiac death in obese patients is less than that in patients with normal BMI. Moreover, further study reveals that the risk of adverse events in patients with chronic heart failure is highest in patients with low BMI (<20 kg/m²) and the risk for cardiovascular mortality and hospitalization was lowest in patients who were overweight (BMI 25 to 29.9 kg/m²) [19]. The reason for the obesity paradox is unclear. There are 2 possible explanations. The first is that obese patients tend to receive guideline-based optimal medical therapy (i.e. statins, angiotensin-converting enzyme inhibitors, and β-blockers) more than nonobese patients. The second explanation is that the physicians tend to advise obese patients about lifestyle modifications, such as to quit smoking, eat a suitable diet, and exercise, more than those with normal BMI [20].

LBBB occurs when there are conduction defects at both the left anterior and left posterior fascicles. Normally, the left bundle branch is supplied by many branches of the coronary artery. Therefore, if a complete LBBB as a result of CAD occurs, patients must have large area of ischemia and have more than 1 coronary artery affected [21]. Thus, in those with LV systolic dysfunction, complete LBBB is usually a result of myocardial disease than CAD.

We found no significant differences in the prevalence of hypertension and dyslipidemia in the present study. The most likely explanation is that most patients with preexisting hypertension and dyslipidemia had received antihypertensive drugs and statins and advised to modify their lifestyle. Therefore, blood pressure and lipid level are well controlled, reducing cardiovascular risk. As seen from the calculation of 10-year risk of cardiovascular disease in SCORE from ESC and pooled cohort equation for ASCVD risk score from ACC/AHA [8, 9], predictors are systolic blood pressure and total cholesterol, HDL, or LDL levels. If the patients have been treated, blood pressure and lipid profile will decrease and cardiovascular risk will be reduced.

The limitations of our study include its retrospective and single-center design. Certain data regarding family history of premature cardiovascular disease and smoking were limited, but these were similar between the 2 groups. Another limitation is the method of internal validation, in that we used a random sample generated from the original data set. External validation with another data set was not conducted because of the limited population that met the inclusion criteria (all patients who have chest pain or LV regional wall motion abnormalities were excluded). The 240 patients in this study were all patients with LV systolic dysfunction who met the inclusion criteria diagnosed between August 2000 and October 2014 (14 years) in our center. Before 2000, the data were limited. An alternative sampling method such as a bootstrap validation approach by sampling individuals with replacement from the original sample might have been better. External validity is needed by having another set of samples independent of the sample from which the predictive model was derived. A further prospective multicenter study to validate the prediction score is warranted.

Because of the high specificity and NPV, this equation is suitable for excluding patients with NICM who should not undergo CAG or stress imaging. The score generated can be used to determine whether or not CAG should be performed. Patients with LV systolic dysfunction and a score <0 are not likely to benefit from CAG because there is little chance that it will detect significant CAD.

In patients with LV systolic dysfunction, predictors associated with ICM are age (>60 years), sex (male), and DM. Predictors associated with NICM are BMI >25 kg/m² and LBBB on ECG. The prediction score has modest sensitivity, but high specificity for predicting significant coronary stenosis and can be used to determine who will not benefit from CAG. However, further prospective multicenter study to validate the prediction score is warranted.