Abstract
Aiming to optimize the diagnosis of pulmonary embolism (PE) in patients with acute exacerbation of chronic obstructive pulmonary disease (AECOPD), we conducted a retrospective study enrolling 185 AECOPD patients, of whom 90 were diagnosed with PE based on computed tomography pulmonary angiography (CTPA). Ten characteristic indicators and 27 blood indicators were extracted for each patient. First, we quantified the importance of each indicator for diagnosing PE in AECOPD using fuzzy rough sets (FRS) and selected the more important indicators to construct a support vector machine (SVM) diagnosis model called FRS-SVM. The performance of the proposed diagnosis model on the test sets was compared to that of the logistic regression model. The average accuracy and area under the curve (AUC) of the proposed model for the test sets in 10 independent trials were 94.67% and 0.944, respectively, compared to 80.41% and 0.809 for the logistic regression model. Thus, we validated the higher accuracy and stability of the FRS-SVM for PE diagnosis in patients with AECOPD. This model improved the prediction probability before CTPA and can be used in clinical practice to help doctors make decisions.
Introduction
Chronic obstructive pulmonary disease (COPD) is an independent risk factor for pulmonary embolism (PE) and is more pronounced in acute exacerbations of COPD (AECOPD) [Citation1]. However, the diagnosis of PE with coexisting AECOPD remains challenging due to their similar clinical symptoms. Computed tomography pulmonary angiography (CTPA) is the primary method used for the diagnosis of PE. However, the application of CTPA is limited by renal insufficiency and medical costs, which lead to delayed or missed PE recognition [Citation2–4]. Therefore, the prediction probability before performing CTPA requires improvement.
Wells and Geneva scores and elevated D-dimer levels are valuable in predicting PE and can improve the prediction probability before CTPA. However: (1) the accuracy of Wells and Geneva scores is poor for PE in patients with COPD [Citation5]. (2) The specificity of D-dimer level is low in AECOPD patients with PE [Citation6]. (3) Other blood and characteristic indicators, which are valuable in diagnosing PE in patients with AECOPD, should not be ignored [Citation7,Citation8], and several aspects of the patient’s state must be considered. However, dozens of indicators inevitably have the opposite characteristics, which interfere with accurate results. Therefore, there is a need to simplify complex indicators and extract the most effective indicators for PE diagnosis based on scientific technology.
Fuzzy rough sets (FRS) use the decision information system constructed by the attribute sets (features) and the decision sets (categories) to compute the dependence of each feature on the categories, and are suitable for indicator evaluation and reduction. The support vector machine (SVM) [Citation9] is a supervised learning algorithm that effectively recognizes subtle patterns in complex datasets [Citation10] and has significant advantages in lung diseases [Citation11,Citation12].
To extract the most sensitive indicators for PE in patients with AECOPD and improve the prediction probability before CTPA, this study proposed an optimized diagnostic algorithm using FRS and SVM, which achieved a better diagnostic performance in accuracy and AUC compared to those of the logistic regression diagnosis model.
Patients and methods
Study population
This retrospective study was conducted on a population of patients admitted to the hospital for AECOPD in the Department of Respiratory and Critical Care Medicine, The First Affiliated Hospital of Chongqing Medical University, between January 2015 and July 2021. Based on physician experience, all the included patients were suspected to have PE.
The inclusion criteria were: (1) age >40 years; (2) diagnosis of COPD based on the electronic medical records and lung function data, according to the Global Initiative for Chronic Obstructive Lung Disease (GOLD) diagnostic criteria stipulating a forced expiratory volume in one second (FEV1)/forced vital capacity (FVC) of <70% after bronchodilator administration. AECOPD is a stage in which respiratory symptoms, such as cough, wheezing, and sputum, show rapid worsening [Citation13]; (3) pulmonary arteries filled with defects or obstructions used as the diagnostic criteria to determine whether patients with AECOPD had PE [Citation14]. A radiology resident evaluated and reported all CT scans, and a radiology specialist conducted the review.
The exclusion criteria were: (1) active cancer, hematopoietic diseases, liver insufficiency, kidney dysfunction, gastrointestinal bleeding, asthma, pregnancy, or acute myocardial infarction; (2) limb fracture in the last month; (3) history of PE or deep vein thrombosis (DVT); (4) direct transfer to an intensive care unit; (5) prehospital diagnosis of PE and treatment with anticoagulants; and (6) incomplete clinical information.
All patients were divided into the study and control groups. The study group included 90 patients with AECOPD with PE while the control group included 95 patients with AECOPD without PE. The enrollment flowchart is shown in .
Figure 1. The enrollment flow chart of this study.
Abbreviations: AECOPD: acute exacerbations of chronic obstructive pulmonary disease; PE: pulmonary embolism; CTPA: computed tomography pulmonary angiography.
The ethics committee of the First Affiliated Hospital of Chongqing Medical University approved the study protocol. This retrospective study collected clinical data and did not interfere with the treatment plan. Therefore, patients did not provide informed consent.
Data collection
Characteristic indicators, including sex, age, smoking index (SMI), COPD disease course (COPD-DC), hypertension, diabetes, physical activity, symptoms of PE, and signs were collected. The symptoms included hemoptysis and chest pain, while the signs included lower extremity edema (LEE). The physical activity assessment was mainly based on the frequency of walking by the patient for >30 min in the past week, with a frequency of <3 times defined as physical activity reduction (PAR). Blood indicators such as hemoglobin level (Hb), platelet distribution width (PDW-SD), neutrophil count (NEUT), lymphocytes count (LYM), monocyte count (MONO), eosinophil count (EOS), basophil count (BASO), red-cell distribution width (RDW-CV), platelet-large cell ratio (P-LCR), prothrombin time (PT), activated partial thromboplastin time (APTT), thrombin time (TT), fibrinogen level (Fbg), fibrinogen degradation product levels (FDP), arterial partial pressure of carbon dioxide (PaCO2), partial pressure of arterial oxygen (PaO2), serum creatinine level (Scr), urate level (UA), albumin level (ALB); alanine transaminase level (ALT), aspartate transaminase level (AST), lactate dehydrogenase level (LDH), N-terminal pro-brain natriuretic peptide level (NT-proBNP), myoglobin (MYO), creatine kinase isoenzyme MB level (CK-MB), and cardiac troponin T level (cTnT), were collected within 12 h of admission. The hospitalization information consisted of the time and route of antibiotic use, glucocorticoid use time, ventilator use, and hospital stay.
Statistical analysis
All statistical analyses were performed using IBM SPSS Statistics for Windows, version 26.0 (IBM Corp., USA). Normality was tested using Shapiro–Wilk tests. p < 0.05 was considered statistically significant. For continuous variables with normal distribution, data are presented as means ± standard deviations (SD) and compared between groups using independent-sample t-tests. For continuous variables with non-normal distribution, median interquartile ranges (IQR) and Mann–Whitney U-tests were used. Categorical data are expressed as frequencies and percentages, and chi-square tests were used to compare the groups. The characteristic indicators and blood indicators (p < 0.05) in univariate analysis were analyzed by multivariate analysis using the “Forward” method.
PE diagnostic model
We constructed a diagnostic model for PE in patients with AECOPD using MATLAB 2018B. A decision information system S185 × 37 was constructed. All characteristic and blood indicators constituted the attribute set, while the diagnosis results served as the deciding set. The importance of each indicator was quantified using the FRS. Indicators with higher importance were extracted as inputs to the SVM. The patients were randomly divided into two groups: 93 patients were used to train the SVM model and 92 patients were used to verify the performance of the trained model. Moreover, a Gaussian kernel function was adopted to improve the nonlinear classification ability of the model. The accuracy, sensitivity, specificity, receiver operating characteristic (ROC) curves, and confusion matrix were compared between the proposed PE diagnostic model and the logistic regression model.
Results
Baseline characteristics of the study population
This study included 185 patients (143 men and 42 women; median age,76 years). Their baseline demographic characteristics are shown in . The age, sex, COPD-DC, SMI, hypertension, diabetes, chest pain, and LEE did not differ significantly between the two groups. However, PAR (60% vs. 40%, p = 0.016) and hemoptysis (85.7% vs. 14.30%, p = 0.045) were significantly higher in the study group compared to those in the control group.
Table 1. Baseline characteristics of participants with and without PE.
After admission, AECOPD patients with PE were more likely to be diagnosed with DVT (p < 0.001) and show long-term antibiotic use (p = 0.001), long-stay hospitalizations (p < 0.001), and more frequent ventilator use (p < 0.001). However, the duration of glucocorticoid use was longer in the control group than that in the study group (p = 0.008). The differences in the route of antibiotic use between the two groups were not statistically significant ().
Table 2. Hospitalization information of participants with and without PE.
Overall comparison of blood indicators
Compared to the control group, the study group showed significantly decreased PDW-SD, NEUT, ALB, and PaO2 levels. However, the FDP, D-dimer, ALT, AST, LDH, NT-proBNP, and cTnT levels were significantly higher ().
Table 3. Comparison of the blood indicators between the two groups.
The joint FRS-SVM PE diagnostic model
Indicator reduction based on FRS
The importance of each indicator was quantified using the FRS, as shown in . Five indicators, D-dimer, FDP, LDH, LEE, and PAR, were significantly higher than other indicators, implying that they were more important than other indicators for PE diagnosis in patients with AECOPD. Among them, D-dimer level was the most important indicator, in line with the guidelines. Thus, we excluded unimportant indicators, and the original 37 indicators were reduced to five important indicators.
Figure 2. The importance and reduction of indicators.
Abbreviations: AECOPD: acute exacerbations of chronic obstructive pulmonary disease; PE: pulmonary embolism; COPD-DC: COPD disease course; SMI: smoking index; LEE: lower extremity edema; PAR: the physical activity reduction; Hb: hemoglobin; PDW-SD: platelet distribution width; NEUT: neutrophils; LYM: lymphocytes; MONO: monocytes; EOS: eosinophils; BASO: basophils; RDW-CV: red-cell distribution width; P-LCR: platelet-large cell ratio; PT: prothrombin time; APTT: activated partial thromboplastin time; TT: thrombin time; Fbg: fibrinogen; FDP: fibrinogen degradation products; PaCO2: arterial partial pressure of carbon dioxide; PaO2: partial pressure of arterial oxygen; Scr: serum creatinine; UA: urate; ALB: albumin; ALT: alanine transaminase; AST: aspartate transaminase; LDH: lactate dehydrogenase; NT-proBNP: N-terminal pro-brain natriuretic peptide; MYO: myoglobin; CK-MB: creatine kinase isoenzyme MB; cTnT: cardiac troponin T.
PE diagnostic model based on the SVM
We used the five indicators as inputs and the corresponding diagnosis results as labels to train the SVM model. Considering the randomness, 10 trials were conducted. The classification confusion matrix was used to express the identification of each label in 10 trials (). The samples in the test set differed in each trial owing to randomness. In trial 5, only two AECOPD patients without PE were misdiagnosed. shows the ROC curves for the 10 trials. The highest and lowest AUCs were 0.982 and 0.908, respectively, with an average AUC value of 0.944. The diagnosis results are listed in , where the average accuracy of the 10 trials was 94.67%, the standard deviation was 0.022, the average sensitivity was 90.61%, and the average specificity was 98.75%. Therefore, the PE diagnostic algorithm proposed in this study achieved good diagnostic ability and stability in all aspects.
Figure 3. Classification confusion matrix of 10 trials based on the proposed model.
Abbreviations: AECOPD: acute exacerbations of chronic obstructive pulmonary disease; PE: pulmonary embolism.
Figure 4. ROC curves of 10 trials based on the proposed diagnostic model.
Abbreviations: AUC: area under the curve.
Table 4. The diagnosis results of 10 trials based on the proposed model.
Logistic regression model comparison
We used a logistic regression model for comparison. First, the results of the multivariate analysis indicated that D-dimer, NEUT, LDH, and PaO2 levels were significantly related to AECOPD patients with PE, with odds ratios of 2.180 (95% confidence interval [CI], 1.633–2.910; p < 0.001), 0.810 (95% CI, 0.716–0.917; p = 0.001), 1.005 (95% CI, 1.002–1.007; p < 0.001), and 0.968 (95% CI, 0.944–0.993; p = 0.013), respectively (). Further, taking these four indicators as input, in the same way, the improved dataset was arranged randomly, and 93 patients were selected to construct the logistic regression model, where the loss function was described by cross-entropy. The diagnostic results of the 10 trials of the logistic regression model are shown in . shows that the diagnosis of 10 trainings all showed convergence. The highest accuracy was 0.857, which was lower than the lowest accuracy of the proposed FRS-SVM model. Meanwhile, the standard deviation, 0.023, was higher than that of the proposed model, indicating the higher stability of the proposed model compared to the logistic regression model (). The AUC of the 10 trials in the logistic regression model was also lower (). Furthermore, the diagnostic information of the logistic regression model showed lower sensitivity and specificity ().
Figure 5. diagnosis results of 10 trials based on logistic regression model; (a) decreasing curves of cross entropy loss function; (b) the accuracies of test set; (c) ROC curves.
Abbreviations: ROC: receiver operating characteristic.
Table 5. Multivariate analysis of AECOPD with PE.
Table 6. The diagnosis results of 10 trials based on the logistic regression model.
Discussion
The treatments differed between the two groups of patients. Patients with AECOPD and PE more often required antibiotics and ventilator use, as well as fewer glucocorticoids. Timely identification of patients with PE and prompt initiation of anticoagulation treatment are vital for reducing the mortality risk. However, PE diagnosis in patients with AECOPD patients is difficult due to the similarities in clinical signs and symptoms between AECOPD and PE, which lead to diagnostic delays and additional mortality [Citation15]. For example, we observed no significant differences in chest pain and LEE between the two groups, consistent with the finding reported by Furcada [Citation5]. Thus, reliable tools are required, which provide valuable information for precise diagnostic decision-making and improve prediction probability before CTPA.
FRS is a mathematical tool used to address inaccurate and incomplete data, which uses approximate sets to describe effective information in the dataset. The main ideology of FRS is to process the actual value data and derive problem decision-making and classification rules by removing redundant attributes to maintain the classification ability and achieve attribute reduction; thus, this method is widely used in attribute reduction algorithms [Citation9]. Our study evaluated and reduced the indicators using the FRS and multivariate analysis, respectively. However, the multivariate analysis could not achieve the important sorting of multiple indicators, and the influence of the characteristic indicators was ignored. Therefore, FRS showed superiority in the choice of indicators.
Regarding the accuracy of the selected results of the FRS, previous studies demonstrated the relationships of the three blood indicators mentioned above to PE. D-dimer is a soluble degradation product of cross-linked fibrin under the action of the fibrinolytic system, thus representing coagulation and fibrinolytic system activation. Therefore, D-dimer level has a significantly negative predictive value in determining PE [Citation14,Citation16,Citation17]. LDH is a glycolytic enzyme widely present in the cytoplasm of skeletal muscle, heart, liver, and lung tissues [Citation18]; its levels change dynamically in the presence or absence of PE and are also associated with risk stratification [Citation19,Citation20]. FDP is the end-product of the coagulation cascade [Citation21]. Melvin et al. prospectively evaluated the diagnostic accuracy of FDP in patients with suspected PE. The results showed the potential clinical application value of FDP in excluding PE [Citation22]. LEE and PAR are considered independent risk factors for PE [Citation4].
Subsequently, we compared the PE diagnosis effect of the proposed model to that of a logistic model. The average accuracy, standard deviation of accuracy, average AUC, average sensitivity, and average specificity of the proposed model for 10 trials were 94.67%, 0.022, 0.944, 90.61%, and 98.75%, respectively, compared to 80.41%, 0.023, 0.809, 73.37%, and 88.80% for the logistic regression method. A cross-sectional study evaluating the diagnostic performance of traditional risk stratification scores reported the sensitivity and specificity, respectively, for the Wells of 24% and 90%, and 59% and 43% for the Geneva score [Citation5]. Thus, the proposed model showed better PE diagnosis ability and robustness.
Our results provide important evidence that the machine model based on blood and characteristic indicators can substantially improve the prediction probability of PE in patients with AECOPD and, thereby, quickly categorize them. However, few studies have been conducted in this area. Prospective studies should be conducted to explore the authenticity and practicability of this model. Moreover, clinical testing is required to improve the machine model.
In this study, 90 (49%) of the 185 patients were considered to have PE. The incidence of PE in AECOPD varied among different studies owing to the study design, inclusion and exclusion criteria, and ethnicity. For example, the characteristics may have differed between the people selected for the study sample and those who did not participate in this study, leading to selection bias. In this study, according to physician experience, all included patients were suspected of having PE; thus, a large proportion of AECOPD patients without PE were excluded, which led to increased incidence. Moreover, this study had a small sample size, which might have overestimated the actual incidence of PE.
This study had several limitations. First, this was a single-center retrospective study with small sample sizes, some bias, and hidden confounders. Second, this study considered only 37 indicators; however, some studies have revealed that inflammation (determined by CRP and ESR) may be responsible for the increased incidence rates of PE in patients with COPD. Future prospective studies should include more samples and indicators to prospectively evaluate the real-world performance of this model.
Conclusions
This study quantified the importance of each blood indicator and characteristic indicator for diagnosing PE in patients with AECOPD. The results showed that D-dimer, FDP, LDH, LEE, and PAR levels were the most important indicators. Furthermore, the proposed diagnosis model combining FRS and SVM achieved better diagnostic performance and significantly improved the prediction probability before CTPA.
Credit authorship contribution statement
Rui Yu: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Writing - Original Draft; Xianghua Kong: Supervision, Formal analysis; Youlun Li: Resources, Supervision, Conceptualization.
Declaration of interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Funding
The author(s) reported there is no funding associated with the work featured in this article.
References
- Chen WJ, Lin CC, Lin CY, et al. Pulmonary embolism in chronic obstructive pulmonary disease: a population-based cohort study. COPD. 2014;11(4):438–443. DOI:10.3109/15412555.2013.813927
- Wang J, Ding YM. Prevalence and risk factors of pulmonary embolism in acute exacerbation of chronic obstructive pulmonary disease and its impact on outcomes: a systematic review and meta-analysis. Eur Rev Med Pharmacol Sci. 2021;25(6):2604–2616.
- Swan D, Hitchen S, Klok FA, et al. The problem of under-diagnosis and over-diagnosis of pulmonary embolism. Thromb Res. 2019;177:122–129. DOI:10.1016/j.thromres.2019.03.012
- Dentali F, Pomero F, Micco PD, et al. Prevalence and risk factors for pulmonary embolism in patients with suspected acute exacerbation of COPD: a multi-center study. Eur J Intern Med. 2020;80:54–59. DOI:10.1016/j.ejim.2020.05.006
- Maritano Furcada J, Castro HM, De Vito EL, et al. Diagnosis of pulmonary embolism in patients with acute exacerbations of chronic obstructive pulmonary disease: a cross-sectional study. Clin Respir J. 2020;14(12):1176–1181. DOI:10.1111/crj.13257
- Akpinar EE, Hoşgün D, Doğanay B, et al. Should the cut-off value of D-dimer be elevated to exclude pulmonary embolism in acute exacerbation of COPD. J Thorac Dis. 2013;5(4):430–434. DOI:10.3978/j.issn.2072-1439.2013.07.34
- Wang J, Wan Z, Liu Q, et al. Predictive value of red blood cell distribution width in chronic obstructive pulmonary disease patients with pulmonary embolism. Anal Cell Pathol (Amst). 2020;2020:1935742. DOI:10.1155/2020/1935742
- Białas AJ, Kornicki K, Ciebiada M, et al. Monocyte to large platelet ratio as a diagnostic tool for pulmonary embolism in patients with acute exacerbation of chronic obstructive pulmonary disease. Pol Arch Intern Med. 2018;128(1):15–23. DOI:10.20452/pamw.4141
- Yuan Z, Chen H, Xie P, et al. Attribute reduction methods in fuzzy rough set theory: an overview, comparative experiments, and new directions. Appl Soft Comput. 2021;107:107353. DOI:10.1016/j.asoc.2021.107353
- Hu R, Gan J, Zhu X, et al. Multi-task multi-modality SVM for early COVID-19 diagnosis using chest CT data. Inf Process Manage. 2022;59(1):102782. DOI:10.1016/j.ipm.2021.102782
- Gao N, Tian S, Li X, et al. Three-dimensional texture feature analysis of pulmonary nodules in CT images: lung cancer predictive models based on support vector machine classifier. J Digit Imaging. 2020;33(2):414–422. DOI:10.1007/s10278-019-00238-8
- Zhu Y, Tan Y, Hua Y, et al. Feature selection and performance evaluation of support vector machine (SVM)-based classifier for differentiating benign and malignant pulmonary nodules by computed tomography. J Digit Imaging. 2010;23(1):51–65. DOI:10.1007/s10278-009-9185-9
- Vogelmeier CF, Criner GJ, Martinez FJ, et al. Global strategy for the diagnosis, management, and prevention of chronic obstructive lung disease 2017 report. GOLD executive summary. Am J Respir Crit Care Med. 2017;195(5):557–582. DOI:10.1164/rccm.201701-0218PP
- Konstantinides SV, Meyer G, Becattini C, et al. 2019 ESC guidelines for the diagnosis and management of acute pulmonary embolism developed in collaboration with the european respiratory society (ERS): the task force for the diagnosis and management of acute pulmonary embolism of the european society of cardiology (ESC). Eur Respir J. 2019;54(3):1901647. DOI:10.1183/13993003.01647-2019
- Fernández C, Jiménez D, De Miguel J, et al. Enfermedad pulmonar obstructiva crónica en pacientes con tromboembolia de pulmón aguda sintomática. Arch Bronconeumol. 2009;45(6):286–290. DOI:10.1016/j.arbres.2008.10.008
- Miron MJ, Perrier A, Bounameaux H, et al. Contribution of noninvasive evaluation to the diagnosis of pulmonary embolism in hospitalized patients. Eur Respir J. 1999;13(6):1365–1370. DOI:10.1183/09031936.99.13613719
- Bounameaux H, de Moerloose P, Perrier A, et al. D-dimer testing in suspected venous thromboembolism: an update. QJM. 1997;90(7):437–442. DOI:10.1093/qjmed/90.7.437
- Jurisic V, Radenkovic S, Konjevic G. The actual role of LDH as tumor marker, biochemical and clinical aspects. Adv Exp Med Biol. 2015;867:115–124. DOI:10.1007/978-94-017-7215-0_8
- Ben SQ, Ni SS, Shen HH, et al. The dynamic changes of LDH isoenzyme 3 and D-dimer following pulmonary thromboembolism in canine. Thromb Res. 2007;120(4):575–583. DOI:10.1016/j.thromres.2006.12.015
- Babaoglu E, Hasanoglu HC, Senturk A, et al. Importance of biomarkers in risk stratification of pulmonary thromboembolism patients. J Investig Med. 2014;62(2):328–331. DOI:10.2310/JIM.0000000000000041
- Xue L, Tao L, Li X, et al. Plasma fibrinogen, D-dimer, and fibrin degradation product as biomarkers of rheumatoid arthritis. Sci Rep. 2021;11(1):16903. DOI:10.1038/s41598-021-96349-w
- Mac Gillavry MR, de Monyé W, Lijmer JG, et al. Clinical evaluation of a monoclonal antibody-based enzyme immunoassay for fibrin degradation products in patients with clinically suspected pulmonary embolism. ANTELOPE-Study group. Thromb Haemost. 2000;83(6):892–895.