Prediction of cardiovascular risk in patients with chronic obstructive pulmonary disease: a study of the National Health and Nutrition Examination Survey database

The data were extracted from NHANES (2007–2012) database [21], a cross-sectional survey of the U.S. civilian. Information was collected via household interviews and standardized physical examinations in specially equipped mobile examination centers. A total of 3226 COPD adults aged 40 to 79 years were enrolled in this study, dividing into the training (n = 2351) and testing (n = 895) sets. The approval from the Institutional Review Board of Beijing Friendship Hospital, Capital Medical University was not required because the data accessed from NHANES were freely available.

COPD was confirmed based on the Medical Conditions Questionnaire (MCQ), including “Ever told you had emphysema” (MCQ160G) and “Ever told you had chronic bronchitis” (MCQ160K). The participants would be diagnosed as COPD if one of the two questions were answered yes. In addition, if the subjects underwent two pulmonary function measurements, the mean value of baseline 1st test spirometry-forced expiratory volume in the first 1.0s (SPXNFEV1) and bronchodilator 2nd test spirometry-forced expiratory volume in the first 1.0s (SPXBFEV1) was taken as FEV1, and the mean value of baseline 1st test spirometry-forced vital capacity (SPXNFVC) and bronchodilator 2nd test spirometry- forced vital capacity (SPXBFVC) was taken as FVC. If only one pulmonary function measurement was performed, the ratios of FEV1 (measured by SPXNFEV1) and FVC (measured by SPXNFVC) were calculated. FEV1 was predicted by gender, age and height in the whole population according to different races. The participants would be diagnosed as COPD when the actual FEV1 was less than 80% of the predicted value, and the actual FEV1/FVC was less than 70% of the predicted value.

CVD were determined respectively according to the questions “Ever told you had angina or heart failure” (MCQ160B), “Ever told you had heart attack” (MCQ160E), and “Has a doctor or other health professional ever told you that you had coronary heart disease” (MCQ160C). The individuals would be diagnosed as CVD if one of the three questions were answered yes.

Variables were extracted from the NHANES database containing demographic information, health-related characteristics, and healthcare-related characteristics. Demographic information was as follows: age, gender, ethnicity, education, and poverty income ratio. Health-related characteristics included general health, general health compared with last year, body mass index (BMI), smoking, anyone smoking inside the home, total smokers inside the home, cotinine (nicotine metabolites), 4-(methylnitrosamino)-1-(3-pyridyl)-1-butanol (NNAL, nicotine metabolites), sedentary activity per day, systolic blood pressure, diastolic blood pressure, high-density lipoprotein (HDL), glycosylated hemoglobin (HbA1c), family history of heart disease, and asthma. Healthcare-related characteristics covered healthcare place type, number of times received healthcare over past year, stayed overnight in the hospital due to illness last year, and number of overnight stays in the hospital due to illness.

All statistical analyses were performed using SAS software (version 9.4) and scikit-learn (version 0.23.1). Scikit-learn (version 0.23.1) is a Python-based open-source machine learning library and used to implement the random forest model. The quantitative data were respectively described as the median and quartile [M (Q1, Q3)] through Mann–Whitney U test. N (%) was used to express the categorical data using \({\upchi }^{2}\) test. Data distribution were adjusted with the sample weight of the Mobile Examination Center (MEC) to deal with the oversample of the data itself. Since the data missing ratio was less than 10%, the mean or mode of the weighted sample was directly used to filling. All COPD cases were divided into the training and testing sets. Variables with significant differences were into the multivariable logistic regression model using the training set, and then internal validation with the testing set were to assess the predictive effectiveness of the CVD risk among COPD patients. Similarly, a random forest model for the risk prediction of CVD was carried out using the training set, and then the testing set was used to internally validate the model performance. P < 0.05 was considered statistically significant.

A total of 3226 COPD cases were included form the NHANES (2007–2012). Of which, 2351 COPD patients identified through the spirometry data were included in the training set, while another 895 COPD patients identified by questionnaire data were included in the testing set. The characteristics of the training set were shown in Table 1. There were 428 patients in COPD & CVD group and 1923 cases in COPD group, respectively. Compared with COPD group, the patient’s age (Z = − 2097.465, P < 0.001), male ratio (χ² = 13.470, P < 0.001), education (χ² = 8.905, P = 0.012), and poverty income ratio (Z = − 648.776, P < 0.001) were higher in COPD & CVD group. The overall health status of patients in COPD group was better than that in COPD & CVD group (χ² = 57.185, P < 0.001), while the BMI (Z = 1084.647, P < 0.001), smoking ratio (χ² = 16.890, P < 0.001), sedentary activity per day (Z = − 894.344, P < 0.001), systolic blood pressure (Z = 34.157, P < 0.001), HDL (Z = − 1256.720, P < 0.001), HbA1c (Z = 1539.918, P < 0.001), and family history of heart disease ratio (χ² = 20.298, P < 0.001) were lower in COPD group than those in COPD & CVD group. The number of times received healthcare over past year (χ² = 53.250, P < 0.001) and number of overnight stays in the hospital due to illness (χ² = 501.298, P < 0.001) were less in COPD group than that in COPD & CVD group.

The proportion of patients with chronic bronchitis or emphysema was shown in Fig. 1. The results indicated that the COPD & CVD group patients had higher proportion of chronic bronchitis (19.45% vs 8.29%, χ² = 16.689, P < 0.001) and emphysema (21.88% vs 9.96%, χ² = 15.207, P < 0.001) than those in COPD group. In addition, there was no statistical difference in the distribution of angina, heart attack, heart failure, and coronary between males and females in the COPD & CVD group.

The multivariate Logistic regression analysis was carried out to assess the determinants of CVD risk in COPD cases (Table 2). The results showed that age (OR 1.073, 95% CI 1.054 to 1.092), BMI (OR 1.025, 95% CI 1.003 to 1.048), HbA1c (OR 1.192, 95% CI 1.073 to 1.323), family history of heart disease (OR 2.665, 95% CI 1.79 to 3.967), and stayed overnight in the hospital due to illness last year (OR 2.314, 95% CI 1.543 to 3.551) were determinants for the risk of CVD among COPD patients. In addition, elevated HDL (OR 0.417, 95% CI 0.255 to 0.681) and females (OR 0.537, 95% CI 0.384 to 0.751) were associated with a reduced risk of CVD in COPD patients.

The Logistic regression model was established based on these determinants using the training set. The receiver operating characteristic (ROC) curves of the predictive model were plotted in Fig. 2. The area under curve (AUC) of the model was 1.000. The result of the internal validation with the test set showed that the AUC was 0.741, suggesting the model could be used to predict CVD risk in COPD patients.

Totally 1024 decision trees were used in the random forest analysis, the maximum number of sampling features was 4, and the rest were set using default parameter. The important variables-associated with CVD risk were NNAL, HbA1c, HDL, age, gender, diastolic blood pressure, cotinine, poverty income ratio, BMI, systolic blood pressure, and sedentary activity per day. The importance of variables was listed in Fig. 3.

The ROC curves of the random forest model were displayed in Fig. 4. The AUCs of this model and the internal validation were 1.000 and 0.948, respectively. It was indicated that the random forest model was performed well predictive effectiveness in predicting CVD risk among COPD patients, which may be used to guide the clinical practice.