Abstract
Objective: To evaluate the incidence of pleural effusion (PE) following coronary artery bypass grafting (CABG), identify associated risk factors, and develop a validated predictive model for early detection. Methods: A retrospective cohort of 1,979 patients who underwent CABG at Beijing Chaoyang Hospital (Capital Medical University) was randomly divided into training (70%) and validation (30%) sets. Risk factors for PE were identified through univariate analysis, LASSO regression, and multivariate logistic regression. Five machine learning models-nomogram, back-propagation neural network (BPNN), random forest, gradient boosting, and support vector machine-were developed. External validation was performed using data from 289 patients at the First Affiliated Hospital of Guangxi Medical University. Results: PE occurred in 71.0% of patients (1,405/1,979) within 3 days postoperatively. Independent risk factors included body mass index (BMI), carotid artery stenosis, postoperative pneumonia, duration of mechanical ventilation, intraoperative blood loss, operative time, and ejection fraction. Among the models, the BPNN demonstrated the best performance, with area under the curve (AUC) values of 0.828 in the training set and 0.751 in the internal validation set. The AUC for external validation was 0.737, outperforming the other models across all evaluation metrics. Conclusions: This study developed a predictive model for post-CABG pleural effusion with high discriminatory power, providing a useful tool for early risk stratification in clinical settings.
Keywords: Coronary artery bypass grafting, pleural effusion, prediction model, neural network
Introduction
The coronary arteries provide the primary blood supply to the heart. When these vessels become narrowed or occluded, myocardial cells receive insufficient blood and oxygen, resulting in myocardial ischemia and hypoxia. This process is a key driver of coronary artery disease (CAD), which includes coronary heart disease (CHD) and ischemic heart disease (IHD) [1]. Globally, cardiovascular diseases account for over 30% of annual deaths, affecting approximately 17 million people. CAD represents around 40% of these cases, underscoring its significant impact on global health [2].
Coronary artery bypass grafting (CABG) is a highly effective surgical intervention for patients with severe three-vessel disease or critical stenosis of the left main coronary artery [3,4]. However, postoperative pleural effusion (PE) remains a common complication [5-7]. PE not only prolongs hospital stays but also increases healthcare costs, imposing a substantial burden on patients and healthcare systems [8]. Moreover, it is a leading cause of readmission within one month following CABG [9-11].
Recent model-based studies have achieved promising results in predicting complications after CABG, such as septic shock, thrombocytopenia, liver dysfunction, and new-onset atrial fibrillation [12-14]. However, these studies exhibit several limitations. Many involve relatively small sample sizes, and some focus on a narrow set of clinical indicators, failing to comprehensively assess the factors contributing to complications. Notably, there is a significant lack of research specifically addressing postoperative PE. The few existing studies on PE offer limited and superficial analyses of associated risk factors, and effective predictive models for identifying high-risk patients are lacking.
To address this gap, our study conducted a comprehensive analysis of data collected from CABG patients at a medical center over the past decade. We aimed to identify risk factors associated with the development of postoperative PE and applied machine learning (ML) techniques to develop multiple predictive models, which were then successfully validated in clinical settings.
Materials and methods
Patient selection
We conducted a retrospective analysis of patient data from Beijing Chaoyang Hospital, Capital Medical University. The study included patients who underwent CABG between January 2010 and October 2020. The inclusion criteria were as follows: Patients aged 18 years or older, diagnosed with CHD and those who underwent CABG. Patients were excluded if they met any of the following criteria: Pre-existing PE prior to CABG, history of thoracic or cardiac surgery, other conditions potentially leading to PE (e.g., chylothorax, hemothorax, malignancies), liver or kidney dysfunction, emergency or secondary surgeries, and incomplete data.
To externally validate the model, we collected clinical data from patients who underwent CABG at the First Affiliated Hospital of Guangxi Medical University between 2010 and 2020, applying the same inclusion and exclusion criteria. This study was approved by the Ethics Committee of Beijing Chaoyang Hospital Affiliated to Capital Medical University (No: 2018-Ke-321).
Data extraction
Data were extracted from the medical records of eligible patients. The extracted data included variables relevant to postoperative PE prediction. The comprehensive dataset consisted of the following aspects:
Demographic information: gender, age, height, weight, body mass index (BMI), smoking index, and alcohol consumption.
PE-related indicators: presence or absence of PE, classification of effusion volume, location (unilateral/bilateral), and the first occurrence time of PE.
Disease assessment data: past medical history (hypertension, stroke, diabetes, chronic obstructive pulmonary disease [COPD], hypoalbuminemia, prior percutaneous coronary intervention [PCI], carotid artery stenosis [CAS], unstable angina, myocardial infarction), number of diseased coronary arteries, and ejection fraction (EF).
Surgery-related indicators: operative time, intraoperative blood loss, number of coronary bypass grafts, postoperative complications (e.g., pneumonia, pleural effusion), diagnostic methods for PE, and postoperative laboratory examination indicators (creatine kinase-myocardial band [CK-MB], albumin [ALB], total cholesterol [TC], triglycerides [TG], high-density lipoprotein cholesterol [HDL-C], low-density lipoprotein cholesterol [LDL-C], aspartate transaminase [AST], alanine transaminase [ALT], creatine kinase [CK], lactate dehydrogenase [LDH], gamma-glutamyl transpeptidase [GGT], blood urea nitrogen [BUN], creatinine [Cre], cardiac troponin I [cTnI], white blood cells [WBC], red blood cells [RBC], neutrophils [NEUT], lymphocytes [LYM], neutrophil/lymphocyte ratio [NLR], eosinophils [EOS], hemoglobin [HB], platelets [PLT], and platelets/lymphocyte ratio [PLR]).
Prognostic indicators: length of hospital stay, duration of mechanical ventilation, and time spent in the ICU.
Chest X-ray results obtained three days post-CABG were used to classify patients into two groups: those with PE and those without PE. The PE group was further stratified into three subgroups based on the volume of pleural effusion: (1) Small Volume Group: pleural fluid volume < 500 mL (characterized by a blunted costophrenic angle). (2) Medium Volume Group: pleural fluid volume 500-1,000 mL (pleural fluid plane near the lung hilum). (3) Large Volume Group: pleural fluid volume > 1,000 mL (pleural effusion plane exceeding the lung hilum) [15].
Outcome measures
The primary objective of this study was to develop and validate a predictive model for postoperative PE following CABG. The predictive performance of the models was evaluated using various metrics, including the area under the receiver operating characteristic (ROC) curve, calibration plots, and decision curve analysis.
Statistical methods
Statistical analysis was performed using SPSS version 27. Descriptive statistics were used to summarize the characteristics of the data. To enhance the clinical interpretation of continuous variables, optimal cut-off values were used to categorize them, with results presented as counts and percentages. Group comparisons were conducted using either the chi-square test or Fisher’s exact test, depending on the distribution of the data.
Univariate logistic regression analysis was initially performed to assess the impact of individual variables on the occurrence of postoperative PE. The least absolute shrinkage and selection operator (LASSO) method was then used to identify variables potentially influencing the outcome [16]. Finally, multivariate logistic regression analysis was performed to precisely identify variables for inclusion in the predictive model. A statistical significance level was set at a p-value of less than 0.05.
For model construction, the “caret” package in R software (version 4.1.3) was used, with random seeds set to 2023. The enrolled patients were divided into training (70%) and validation (30%) sets. The predictive model was developed using data from the training set. Additionally, 289 patients from the First Affiliated Hospital of Guangxi Medical University were included for external validation of the model. During model development, the “rms”, “neuralnet”, “randomForest”, “gbm”, and “e1071” packages in R were used to sequentially develop a nomogram, a back-propagation neural network (BPNN), a random forest (RF), a gradient boosting model (GBM), and a support vector machine (SVM).
Results
General characteristics of CABG patients
Data were collected from 2,398 patients who underwent CABG at Beijing Chaoyang Hospital, Capital Medical University. After excluding 419 participants who did not meet the inclusion criteria, 1,979 patients were ultimately included in the analysis (Figure 1). Of these, 1,487 (75.1%) were male and 492 (24.9%) were female (Figure 2A). The mean age of the cohort was 63.33 ± 8.80 years, with the highest proportion (42.0%) in the 61 to 70 years age group. A small percentage (0.8%) were aged 30 to 40 years, while 1.3% were over 81 years old (Figure 2B).
Figure 1.
Flow chart of patient inclusion. Abbreviations: PE, pleural effusion; CABG, Coronary artery bypass grafting.
Figure 2.
Basic information of patients. A. Gender; B. Age; C. Distribution map of the proportion of 1979 patients in each group. Abbreviation: PE, pleural effusion.
Among the 1,979 patients, 574 (29%) did not develop postoperative PE within 3 days, while 1,405 (71%) exhibited this complication. The majority (59.9%) presented with left-sided pleural effusion, with 93% showing mild effusion and only 4% having significant volume accumulation (Figure 2C). For model development, a random 70% of the dataset (n = 1,385) was allocated to the training set, with the remaining 30% (n = 594) designated for validation. The training cohort consisted of 1,385 patients, including 981 with postoperative PE and 404 controls without PE. These patients were stratified into the PE group and the without PE group.
Comparison of baseline characteristics and serum biochemical markers
A comparative analysis of baseline characteristics and serum biochemical markers was performed between the PE and without PE groups. The baseline demographics and clinical parameters are presented in Table 1. Significant differences were identified in multiple clinical parameters, including gender, BMI, hypertension, stroke, CAS, EF, duration in the ICU, mechanical ventilation time, surgical blood loss, and operation time (all P < 0.05). The analysis also revealed marked differences in serum levels of HDL-C, AST, ALT, CK, LDH, cTnI, NEUT, NLR, and EOS between the two groups (all P < 0.05, Table 2).
Table 1.
Comparison of baseline characteristics between the PE group and those without-PE
| Characteristic | Without PE n = 404 | With PE n = 981 | χ2 | P value |
|---|---|---|---|---|
| Age, years | 2.626 | 0.105 | ||
| < 60 | 145 (35.9) | 308 (31.4) | ||
| ≥ 60 | 259 (64.1) | 673 (68.6) | ||
| Gender | 3.915 | 0.048 | ||
| Male | 322 (79.7) | 733 (74.7) | ||
| Female | 82 (20.3) | 248 (25.3) | ||
| BMI, kg/m2 | 10.831 | < 0.001 | ||
| < 26 | 260 (64.4) | 537 (54.7) | ||
| ≥ 26 | 144 (35.6) | 444 (45.3) | ||
| Smoking index | 2.982 | 0.084 | ||
| < 290 | 189 (46.8) | 509 (51.9) | ||
| ≥ 290 | 215 (53.2) | 472 (48.1) | ||
| Alcohol consumption | 1.819 | 0.611 | ||
| Never | 166 (41.1) | 422 (43.0) | ||
| Sometimes | 156 (38.6) | 360 (36.7) | ||
| Former | 20 (5.0) | 62 (6.3) | ||
| Abuse | 62 (15.3) | 137 (14.0) | ||
| Medical history | ||||
| Hypertension | 263 (65.1) | 694 (70.7) | 4.271 | 0.039 |
| Dyslipidemia | 253 (62.6) | 594 (60.6) | 0.518 | 0.472 |
| Diabetes | 160 (39.6) | 392 (40.0) | 0.015 | 0.902 |
| COPD | 10 (2.5) | 27 (2.8) | 0.084 | 0.771 |
| Stroke | 56 (13.9) | 180 (18.3) | 4.076 | 0.043 |
| Hypoproteinemia | 194 (48.0) | 506 (51.6) | 1.451 | 0.228 |
| Prior PCI | 57 (14.1) | 139 (14.2) | 0.001 | 0.997 |
| CAS | 121 (30.0) | 458 (46.7) | 32.947 | < 0.001 |
| Unstable angina | 194 (48.0) | 485 (49.4) | 0.231 | 0.631 |
| Myocardial infarction | 106 (26.2) | 216 (22.0) | 2.855 | 0.091 |
| Postoperative pneumonia | 115 (28.5) | 340 (34.7) | 4.975 | 0.026 |
| No. of diseased coronary arteries | 2.743 | 0.254 | ||
| 1-2 | 57 (14.1) | 125 (12.7) | ||
| 3 | 318 (78.7) | 805 (82.1) | ||
| ≥ 4 | 29 (7.2) | 51 (5.2) | ||
| No. of coronary bypasses | 0.075 | 0.963 | ||
| 1-2 | 148 (36.6) | 357 (36.4) | ||
| 3 | 199 (49.3) | 480 (48.9) | ||
| ≥ 4 | 57 (14.1) | 144 (14.7) | ||
| EF, % | 12.006 | < 0.001 | ||
| < 50 | 43 (10.6) | 178803 (18.1) | ||
| ≥ 50 | 361 (89.4) | 803 (81.9) | ||
| Time in ICU, day | 12.575 | < 0.001 | ||
| < 4 | 226 (55.9) | 446 (45.5) | ||
| ≥ 4 | 178 (44.1) | 535 (54.5) | ||
| Use of respirator | 1.368 | 0.713 | ||
| NO | 50 (12.4) | 113 (11.5) | ||
| Non-invasive | 18 (4.5) | 36 (3.7) | ||
| Invasive | 308 (76.2) | 750 (76.5) | ||
| Invasive and Non-invasive | 28 (6.9) | 82 (8.3) | ||
| Mechanical ventilation time, day | 22.344 | < 0.001 | ||
| < 26 | 374 (92.6) | 812 (82.9) | ||
| ≥ 26 | 30 (7.4) | 169 (17.2) | ||
| Amount of bleeding, mL | 18.285 | < 0.001 | ||
| < 475 | 284 (70.3) | 569 (58.0) | ||
| ≥ 475 | 120 (29.7) | 412 (42.0) | ||
| Time of operation, hour | 11.431 | < 0.001 | ||
| < 4 | 183 (45.3) | 349 (35.6) | ||
| ≥ 4 | 221 (54.7) | 632 (64.4) |
Abbreviations: PE, Pleural effusion; M, median; BMI, Body Mass Index; COPD, Chronic Obstructive Pulmonary Disease; PCI, Percutaneous Transluminal Coronary Intervention; CAS, Carotid artery stenosis; EF, Preoperative Left Ventricular Ejection Fraction; ICU, Intensive Care Unit. Data are presented as number (%).
Table 2.
Comparison of laboratory test results between the PE group and patients without PE
| Characteristic | Without PE n = 404 | With PE n = 981 | χ2 | P value |
|---|---|---|---|---|
| CKMB, U/L | 0.907 | 0.341 | ||
| < 5.7 | 221 (54.7) | 564 (57.5) | ||
| ≥ 5.7 | 183 (45.3) | 417 (42.5) | ||
| ALB, g/L | 3.306 | 0.069 | ||
| < 20.9 | 96 (23.8) | 280 (28.5) | ||
| ≥ 20.9 | 308 (76.2) | 701 (71.5) | ||
| TC, mmol/L | 2.023 | 0.155 | ||
| < 1.60 | 184 (45.5) | 406 (41.4) | ||
| ≥ 1.60 | 220 (54.5) | 575 (58.6) | ||
| TG, mmol/L | 0.298 | 0.585 | ||
| < 0.59 | 130 (32.2) | 301 (30.7) | ||
| ≥ 0.59 | 274 (67.8) | 680 (69.3) | ||
| HDL-C, mmol/L | 5.630 | 0.018 | ||
| < 0.75 | 282 (69.8) | 745 (75.9) | ||
| ≥ 0.75 | 122 (30.2) | 236 (24.1) | ||
| LDL-C, mmol/L | 0.924 | 0.336 | ||
| < 0.45 | 99 (24.5) | 217 (22.1) | ||
| ≥ 0.45 | 305 (75.5) | 764 (77.9) | ||
| AST, U/L | 3.941 | 0.047 | ||
| < 24 | 66 (16.3) | 206 (21.0) | ||
| ≥ 24 | 338 (83.7) | 775 (79.0) | ||
| ALT, U/L | 5.327 | 0.021 | ||
| < 18 | 81 (20.0) | 254 (25.9) | ||
| ≥ 18 | 323 (80.0) | 727 (74.1) | ||
| CK, U/L | 4.604 | 0.032 | ||
| < 218 | 56 (13.9) | 183 (18.7) | ||
| ≥ 218 | 348 (86.1) | 798 (81.3) | ||
| LDH, U/L | 5.767 | 0.016 | ||
| < 320 | 368 (91.1) | 848 (86.4) | ||
| ≥ 320 | 36 (8.9) | 133 (13.6) | ||
| GGT, U/L | 1.741 | 0.187 | ||
| < 34 | 279 (69.1) | 712 (72.6) | ||
| ≥ 34 | 125 (30.9) | 269 (27.4) | ||
| BUN, mmol/L | 3.342 | 0.068 | ||
| < 5.37 | 263 (65.1) | 587 (69.8) | ||
| ≥ 5.37 | 141 (34.9) | 394 (40.2) | ||
| Cre, μmol/L | 3.703 | 0.054 | ||
| < 103.6 | 306 (75.7) | 693 (70.6) | ||
| ≥ 103.6 | 98 (24.3) | 288 (29.4) | ||
| cTnI, ng/L | 6.558 | 0.010 | ||
| < 24.75 | 387 (95.8) | 902 (91.9) | ||
| ≥ 24.75 | 17 (4.2) | 79 (8.1) | ||
| WBC, *109/L | 2.783 | 0.095 | ||
| < 10 | 211 (52.2) | 464 (47.3) | ||
| ≥ 10 | 193 (47.8) | 517 (52.7) | ||
| RBC, *102/L | 1.503 | 0.220 | ||
| < 4 | 372 (92.1) | 921 (93.9) | ||
| ≥ 4 | 32 (7.9) | 60 (6.1) | ||
| NEUT, *109/L | 5.696 | 0.017 | ||
| < 9.14 | 251 (62.1) | 541 (55.1) | ||
| ≥ 9.14 | 153 (37.9) | 440 (44.9) | ||
| LYM, *109/L | 0.938 | 0.333 | ||
| < 0.83 | 140 (34.7) | 367 (37.4) | ||
| ≥ 0.83 | 264 (65.3) | 614 (62.6) | ||
| NLR | 4.270 | 0.039 | ||
| < 7.63 | 200 (49.5) | 426 (43.4) | ||
| ≥ 7.63 | 204 (50.5) | 555 (56.6) | ||
| EOS, *109/L | 5.194 | 0.023 | ||
| < 0.04 | 143 (35.4) | 412 (42.0) | ||
| ≥ 0.04 | 261 (64.6) | 569 (58.0) | ||
| HB, g/L | 0.091 | 0.763 | ||
| < 100 | 235 (58.2) | 562 (57.3) | ||
| ≥ 100 | 169 (41.8) | 419 (42.7) | ||
| PLT, *109/L | 2.476 | 0.116 | ||
| < 160 | 254 (62.9) | 660 (67.3) | ||
| ≥ 160 | 150 (37.1) | 321 (32.7) | ||
| PLR | 0.134 | 0.714 | ||
| < 150.84 | 237 (58.7) | 565 (57.6) | ||
| ≥ 150.84 | 167 (41.3) | 416 (42.4) |
Abbreviations: PE, Pleural effusion; CKMB, creatine kinase myocardial band; ALB, albumin; TC, total cholesterol; TG, triglyceride; HDL-C, high density lipoprotein cholesterol; LDL-C, low density lipoprotein cholesterol; AST, Aspartate Transaminase; ALT, alanine transaminase; CK, Creatine Kinase; LDH, Lactate Dehydrogenas; GGT, Gamma Glutamyl Transpeptidase; BUN, Blood Urea Nitrogen; Cre, Creatinine; cTnI, Cardiac troponin I; WBC, White Blood Cells; RBC, Red Blood Cells; NEUT, Neutrophil; LYM, Lymphocyte; NLR, Neutrophil/Lymphocyte; EOS, Eosinophil; HB, Hemoglobin; PLT, Platelets; PLR, Platelets/Lymphocyte. Data are presented as number (%).
Influencing factors of PE post-CABG
Relying solely on variables identified as statistically significant in multivariate analysis may overlook other crucial factors that could influence the overall interpretation of the data [17]. To address this, we employed the least absolute shrinkage and selection operator (LASSO) regression method due to its advantages in variable selection and model simplification. LASSO regression is particularly effective in high-dimensional datasets, where it can simultaneously perform variable selection and parameter estimation by shrinking less important coefficients to zero. This helps identify a subset of meaningful variables, enhancing model interpretability and reducing overfitting [18].
Using this approach, 20 potentially meaningful variables were identified for further analysis. Due to the coefficient of LDH being zero, this variable was excluded (Table 3), narrowing the analysis to 19 variables that could influence postoperative PE (Figure 3). These variables included gender, BMI, hypertension, stroke, CAS, postoperative pneumonia, duration in ICU, mechanical ventilation time, surgical blood loss, operation duration, EF, HDL-C, AST, ALT, CK, cTnI, NEUT, NLR, and EOS.
Table 3.
LASSO regression analysis of the coefficients of each variable
| Variables | Coefficients |
|---|---|
| Gender | -0.042 |
| BMI | 0.070 |
| Hypertension | 0.035 |
| Stroke | 0.030 |
| CAS | 0.141 |
| Postoperative pneumonia | 0.099 |
| Time in ICU | 0.035 |
| Mechanical ventilation time | 0.096 |
| Amount of bleeding | 0.063 |
| Time of operation | 0.050 |
| EF | 0.073 |
| HDL-C | -0.038 |
| AST | -0.019 |
| ALT | -0.029 |
| CK | -0.033 |
| LDH | 0 |
| cTnI | 0.041 |
| NEUT | 0.019 |
| NLR | 0.022 |
| EOS | -0.013 |
Abbreviations: LASSO, least absolute shrinkage and selection operator; BMI, Body Mass Index; CAS, Carotid artery stenosis; ICU, Intensive Care Unit; EF, Preoperative Left Ventricular Ejection Fraction; HDL-C, high density lipoprotein cholesterol; AST, Aspartate Transaminase; ALT, alanine transaminase; CK, Creatine Kinase; LDH, Lactate Dehydrogenas; cTnI, Cardiac troponin I; NEUT, Neutrophil; NLR, Neutrophil/Lymphocyte; EOS, Eosinophil.
Figure 3.
Use LASSO regression to screen potential variables. A. The coefficient curves of 20 overall survival characteristics in the LASSO model. B. The 10-fold cross-validation of the optimal parameter (λ) in the LASSO model was selected by the minimum standard. The relationship curve between the partial likelihood deviation curve and the logarithm (λ). Abbreviation: LASSO, least absolute shrinkage and selection operator.
To further investigate their influence, a multivariate logistic regression analysis was conducted on these 19 variables. The analysis revealed seven independent predictors of postoperative PE in CABG patients: BMI, CAS, postoperative pneumonia, mechanical ventilation duration, surgical blood loss, operation time, and EF (Figure 4).
Figure 4.
Forest plot of multivariate Logistic regression analysis. Abbreviations: BMI, Body Mass Index; CAS, Carotid artery stenosis; EF, Preoperative Left Ventricular Ejection Fraction; HDL-C, high density lipoprotein cholesterol; ALT, alanine transaminase; CK, Creatine Kinase; LDH, low density lipoprotein cholesterol; cTnI, Cardiac troponin I; NEUT, Neutrophil; NLR, Neutrophil to Lymphocyte rate; EOS, Eosinophil.
Construction of prediction model for postoperative PE
Based on the seven selected variables, five models were constructed: a nomogram, BPNN, RF, gradient GBM, and SVM. The nomogram model (Figure 5A) incorporates seven predictive variables: BMI, CAS, EF, mechanical ventilation time, surgical bleeding volume, surgical duration, and postoperative pneumonia. The β (x - m) term line represents the product of the β coefficient, derived from linear regression using the least squares method, and the deviation of the independent variable x from its mean (m). The β coefficient represents the regression coefficient, x denotes the independent variable, and m signifies the mean value of that variable. The length of this line segment reflects the variable’s contribution to the prediction of postoperative PE. The scores of each variable are calculated and summed to determine the total score for the patient. The nomogram model estimates the probability of postoperative PE following CABG by plotting a vertical line from the total score to the probability scale.
Figure 5.
Model construction of PE post-CABG. A. Nomogram; B. Random forest; C. BP neural network; D. Variable importance ranking in BP neural network; E. Variable importance ranking of gradient boosting model; F. The number of support vectors of support vector machine. Abbreviations: BMI, Body Mass Index; CAS, Carotid artery stenosis; EF, Preoperative Left Ventricular Ejection Fraction.
The RF model (Figure 5B) is an ensemble learning method that combines multiple decision trees to improve predictive performance [19]. Each tree in the forest is trained on a random subset of the data, and the final prediction is made by aggregating the results of all trees. This approach reduces overfitting and enhances the model’s ability to generalize to unseen data.
The BPNN model (Figure 5C and 5D) consists of seven input variables and two output results. This artificial neural network employs a supervised learning algorithm to map input data to output predictions through multiple hidden layers [20]. Each input variable contributes to the network’s ability to learn complex nonlinear relationships, while the two output results represent the predicted outcomes. The structure of the BPNN enables it to capture intricate patterns in the data, making it particularly effective for tasks requiring high-dimensional input data.
The GBM model (Figure 5E) is a powerful ensemble learning method that builds a strong predictive model by sequentially combining multiple weak prediction models, typically decision trees [21]. This approach is based on the principle of gradient descent, where each new tree is trained to correct the errors of the previous ensemble. Specifically, the GBM algorithm minimizes a loss function by iteratively adding trees, with each tree focusing on the residuals (i.e., the differences between predicted and actual values) of the previous iteration. This iterative process ensures that the model gradually improves its predictive accuracy. One of the key features of GBM is its ability to handle complex datasets with nonlinear relationships and interactions between variables. By combining the predictions of multiple trees, the GBM model can capture intricate patterns in the data that may not be apparent in individual models. Additionally, GBM provides feature importance scores, allowing researchers to identify the most influential variables in the dataset.
The SVM model (Figure 5F) utilizes a radial kernel function with 846 support vectors and a penalty factor of 1. The radial basis function kernel, commonly used in SVM, maps the input data into a higher-dimensional space to achieve better separation between classes [22]. The 846 support vectors represent the data points that lie closest to the decision boundary and play a crucial role in defining the hyperplane that separates the classes. The penalty factor, set to 1, controls the trade-off between achieving a low training error and maintaining a simple model. This configuration ensures that the SVM model achieves a balance between complexity and generalization, making it robust for classification tasks. Variable importance analysis for the RF, BPNN, and GBM demonstrates that CAS is the most influential variable, followed by the mechanical ventilation time.
Validation of the model
The training and validation set data were utilized to assess and compare the predictive performance of all models. In the training set, the sensitivity and specificity of the models developed to predict postoperative PE following CABG were as follows: Nomogram (0.624, 0.671), BPNN (0.691, 0.797), RF (0.780, 0.641), GBM (0.760, 0.690), and SVM (0.828, 0.616) (Table 4). Receiver operating characteristic (ROC) curves were constructed, with sensitivity plotted on the y-axis and (1 - specificity) on the x-axis, to evaluate the predictive accuracy of each model using the area under the curve (AUC). In the test set, only the BPNN achieved an AUC value exceeding 0.800, while the remaining models showed comparable classification capabilities with AUC values around 0.750 (Figure 6).
Table 4.
Performance comparison of models
| Group | Model | Sensitivity | Specificity | AUC | 95% CI |
|---|---|---|---|---|---|
| Training set | Nomogram | 0.624 | 0.671 | 0.702 | 0.673-0.732 |
| BPNN | 0.691 | 0.797 | 0.828 | 0.806-0.849 | |
| RF | 0.780 | 0.641 | 0.754 | 0.725-0.783 | |
| GBM | 0.760 | 0.690 | 0.790 | 0.765-0.814 | |
| SVM | 0.828 | 0.616 | 0.774 | 0.748-0.800 | |
| Validation set | Nomogram | 0.531 | 0.718 | 0.667 | 0.620-0.715 |
| BPNN | 0.693 | 0.729 | 0.751 | 0.071-0.792 | |
| RF | 0.818 | 0.642 | 0.762 | 0.719-0.804 | |
| GBM | 0.589 | 0.783 | 0.745 | 0.702-0.788 | |
| SVM | 0.809 | 0.553 | 0.708 | 0.661-0.755 |
Abbreviations: BPNN, back-propagation neural network; RF, random forest; GBM, gradient boosting model; SVM, support vector machine; AUC, area under the curv; CI, Confidence Interval.
Figure 6.
ROC curve, calibration curve and decision curve of the training set of the model. A-C. Nomogram; D-F. BPNN; G-I. RF; J-L. GBM; M-O. SVM. Abbreviations: BPNN, back-propagation neural network; RF, random forest; GBM, gradient boosting model; SVM, support vector machine.
In the validation set, the sensitivity of the nomogram, BPNN, RF, GBM, and SVM for predicting PE were 0.531, 0.693, 0.818, 0.589, and 0.809, respectively. The corresponding specificities were 0.718, 0.729, 0.642, 0.783, and 0.553 (Table 4). The AUC values for each model’s validation set were 0.667, 0.751, 0.762, 0.745, and 0.708, respectively. Using the calibration function in R software, the validation set was resampled via bootstrapping to obtain calibration curves for each model, effectively mitigating selection bias. Notably, the nomogram and calibration plot for the GBM closely resembled the ideal performance curve. Furthermore, the decision curves for all models surpassed the reference line, as depicted in Figure 7. No significant differences were observed between the training and validation sets in terms of data distribution (all P > 0.05, Table 5).
Figure 7.
ROC curve, calibration curve and decision curve of the validation set of the model. A-C. Nomogram; D-F. BPNN; G-I. RF; J-L. GBM; M-O. SVM. Abbreviations: BPNN, back-propagation neural network; RF, random forest; GBM, gradient boosting model; SVM, support vector machine.
Table 5.
Comparison of patient data between training set and validation set, n (%)
| Characteristic | Total n = 1979 | Derivation set n = 1385 | Validation set n = 594 | χ2 | P value |
|---|---|---|---|---|---|
| Age, years | 0.009 | 0.925 | |||
| < 60 | 646 (32.6) | 453 (32.7) | 193 (32.5) | ||
| ≥ 60 | 1333 (67.4) | 932 (67.3) | 401 (67.5) | ||
| Gender | 2.643 | 0.104 | |||
| Male | 1487 (75.1) | 1055 (76.2) | 432 (72.7) | ||
| Female | 492 (24.9) | 330 (23.8) | 162 (27.3) | ||
| BMI, kg/m2 | 0.084 | 0.771 | |||
| < 26 | 1143 (57.8) | 797 (57.5) | 346 (58.2) | ||
| ≥ 26 | 836 (42.2) | 588 (42.5) | 248 (41.8) | ||
| Smoking index | 0.438 | 0.508 | |||
| < 290 | 1007 (50.9) | 698 (50.4) | 309 (52.0) | ||
| ≥ 290 | 972 (49.1) | 687 (49.6) | 285 (48.0) | ||
| Alcohol consumption | 4.066 | 0.254 | |||
| Never | 851 (43.0) | 588 (42.5) | 263 (44.3) | ||
| Sometimes | 748 (37.8) | 516 (37.3) | 232 (39.1) | ||
| Former | 115 (5.8) | 82 (5.9) | 33 (5.6) | ||
| Abuse | 265 (13.4) | 199 (14.3) | 66 (11.0) | ||
| Medical history | |||||
| Hypertension | 1364 (68.9) | 957 (69.1) | 407 (68.5) | 0.065 | 0.799 |
| Dyslipidemia | 1225 (61.9) | 847 (61.2) | 378 (63.6) | 1.085 | 0.298 |
| Diabetes | 778 (39.3) | 552 (39.3) | 226 (38.0) | 0.570 | 0.450 |
| COPD | 53 (2.7) | 37 (2.7) | 16 (2.7) | 0.001 | 0.978 |
| Stroke | 328 (16.6) | 236 (17.0) | 92 (15.5) | 0.724 | 0.395 |
| Hypoproteinemia | 985 (4.8) | 700 (50.5) | 285 (48.0) | 1.091 | 0.296 |
| Prior PCI | 300 (15.2) | 196 (14.2) | 104 (17.5) | 3.642 | 0.056 |
| CAS | 839 (42.4) | 579 (41.8) | 260 (43.8) | 0.658 | 0.417 |
| Unstable angina | 957 (48.4) | 679 (49.0) | 278 (46.8) | 0.823 | 0.364 |
| Myocardial infarction | 452 (22.8) | 322 (23.2) | 130 (21.9) | 0.439 | 0.508 |
| Postoperative pneumonia | 628 (31.7) | 455 (32.9) | 173 (29.1) | 2.666 | 0.103 |
| No. of diseased Coronary arteries | |||||
| 1-2 | 277 (14.0) | 182 (13.1) | 95 (16.0) | 3.801 | 0.150 |
| 3 | 1582 (79.9) | 1123 (81.1) | 459 (77.3) | ||
| ≥ 4 | 120 (6.1) | 80 (5.8) | 40 (6.7) | ||
| No. of coronary bypasses | 4.992 | 0.082 | |||
| 1-2 | 750 (37.9) | 505 (36.5) | 245 (41.2) | ||
| 3 | 939 (47.4) | 679 (49.0) | 260 (43.8) | ||
| ≥ 4 | 290 (14.7) | 201 (14.5) | 89 (15.0) | ||
| EF, % | 1.996 | 0.158 | |||
| < 50 | 301 (15.2) | 221 (16.0) | 80 (13.5) | ||
| ≥ 50 | 1678 (84.8) | 1164 (84.0) | 514 (86.5) | ||
| Time in ICU, day | 0.049 | 0.826 | |||
| < 4 | 957 (48.4) | 672 (48.5) | 285 (48.0) | ||
| ≥ 4 | 1022 (51.6) | 713 (51.5) | 309 (52.0) | ||
| Use of respirator | 0.819 | 0.845 | |||
| NO | 228 (11.5) | 163 (11.8) | 65 (10.9) | ||
| Non-invasive | 77 (3.9) | 54 (3.9) | 23 (3.9) | ||
| Invasive | 1522 (76.9) | 1058 (76.4) | 464 (78.1) | ||
| Invasive and Non-invasive | 152 (7.7) | 110 (7.9) | 42 (7.1) | ||
| Mechanical ventilation time, day | 1.276 | 0.259 | |||
| < 26 | 1706 (86.2) | 1186 (85.6) | 520 (87.5) | ||
| ≥ 26 | 273 (13.8) | 199 (14.4) | 74 (12.5) | ||
| Amount of bleeding, mL | 0.003 | 0.953 | |||
| < 475 | 1218 (61.5) | 853 (61.6) | 365 (61.4) | ||
| ≥ 475 | 761 (38.5) | 532 (38.4) | 229 (38.6) | ||
| Time of operation, hour | 2.798 | 0.094 | |||
| < 4 | 784 (39.6) | 532 (38.4) | 252 (42.4) | ||
| ≥ 4 | 1195 (604) | 853 (61.6) | 342 (57.6) | ||
| CKMB, U/L | 0.258 | 0.611 | |||
| < 5.7 | 1129 (57.0) | 785 (56.7) | 344 (57.9) | ||
| ≥ 5.7 | 850 (43.0) | 600 (43.3) | 250 (42.1) | ||
| ALB, g/L | 0.083 | 0.773 | |||
| < 20.9 | 541 (27.3) | 376 (27.1) | 165 (27.8) | ||
| ≥ 20.9 | 1438 (72.7) | 1009 (72.9) | 429 (72.2) | ||
| TC, mmol/L | 2.807 | 0.094 | |||
| < 1.60 | 819 (41.4) | 590 (42.6) | 229 (38.6) | ||
| ≥ 1.60 | 1160 (58.6) | 795 (57.4) | 365 (61.4) | ||
| TG, mmol/L | 0.803 | 0.370 | |||
| < 0.59 | 628 (31.7) | 431 (31.1) | 197 (33.2) | ||
| ≥ 0.59 | 1351 (68.3) | 954 (68.9) | 397 (66.8) | ||
| HDL-C, mmol/L | 1.632 | 0.201 | |||
| < 0.75 | 1451 (73.3) | 1027 (74.2) | 424 (71.4) | ||
| ≥ 0.75 | 528 (26.7) | 358 (25.8) | 170 (28.6) | ||
| LDL-C, mmol/L | 0.754 | 0.385 | |||
| < 0.45 | 441 (22.3) | 316 (22.8) | 125 (21.0) | ||
| ≥ 0.45 | 1538 (77.7) | 1069 (77.2) | 469 (79.0) | ||
| AST, U/L | 0.709 | 0.400 | |||
| < 24 | 379 (19.2) | 272 (19.6) | 107 (18.0) | ||
| ≥ 24 | 1600 (80.8) | 1113 (80.4) | 487 (82.0) | ||
| ALT, U/L | 0.119 | 0.730 | |||
| < 18 | 483 (24.4) | 335 (24.2) | 148 (24.9) | ||
| ≥ 18 | 1496 (75.6) | 1050 (75.8) | 446 (75.1) | ||
| CK, U/L | 1.329 | 0.249 | |||
| < 218 | 329 (16.6) | 239 (17.3) | 90 (15.2) | ||
| ≥ 218 | 1650 (83.4) | 1146 (82.7) | 504 (84.8) | ||
| LDH, U/L | 3.176 | 0.075 | |||
| < 320 | 1754 (88.6) | 1216 (87.8) | 538 (90.6) | ||
| ≥ 320 | 225 (11.4) | 169 (12.2) | 56 (9.4) | ||
| GGT, U/L | 0.053 | 0.819 | |||
| < 34 | 1413 (71.4) | 991 (71.6) | 422 (71.0) | ||
| ≥ 34 | 566 (28.6) | 394 (28.4) | 172 (29.0) | ||
| BUN, mmol/L | 3.009 | 0.083 | |||
| < 5.37 | 1239 (62.6) | 850 (61.4) | 389 (65.5) | ||
| ≥ 5.37 | 740 (37.4) | 535 (38.6) | 205 (34.5) | ||
| Cre, μmol/L | 3.058 | 0.080 | |||
| < 103.6 | 1450 (73.3) | 999 (72.1) | 451 (75.9) | ||
| ≥ 103.6 | 529 (26.7) | 386 (27.9) | 143 (24.1) | ||
| cTnI, ng/L | 2.467 | 0.116 | |||
| < 24.75 | 1853 (93.6) | 1289 (93.1) | 564 (94.9) | ||
| ≥ 24.75 | 126 (6.4) | 96 (6.9) | 30 (5.1) | ||
| WBC, *109/L | 0.340 | 0.560 | |||
| < 10 | 956 (48.3) | 675 (48.7) | 281 (47.3) | ||
| ≥ 10 | 1023 (51.7) | 710 (51.3) | 313 (52.7) | ||
| RBC, *102/L | 0.233 | 0.630 | |||
| < 4 | 1844 (93.2) | 1293 (93.4) | 551 (92.8) | ||
| ≥ 4 | 135 (6.8) | 92 (6.6) | 43 (7.2) | ||
| NEUT, *109/L | 1.626 | 0.202 | |||
| < 9.14 | 1150 (58.1) | 792 (57.2) | 358 (60.3) | ||
| ≥ 9.14 | 829 (41.9) | 593 (42.8) | 236 (39.7) | ||
| LYM, *109/L | 0.455 | 0.500 | |||
| < 0.83 | 715 (36.1) | 507 (36.6) | 208 (35.0) | ||
| ≥ 0.83 | 1264 (63.9) | 878 (63.4) | 386 (65.0) | ||
| NLR | 1.000 | 0.317 | |||
| < 7.63 | 909 (45.9) | 626 (45.2) | 283 (48.2) | ||
| ≥ 7.63 | 1070 (54.1) | 759 (54.8) | 311 (51.8) | ||
| EOS, *109/L | 5.433 | 0.020 | |||
| < 0.04 | 760 (38.4) | 555 (40.1) | 205 (34.5) | ||
| ≥ 0.04 | 1219 (61.6) | 830 (59.9) | 389 (65.5) | ||
| HB, g/L | 3.703 | 0.054 | |||
| < 100 | 1111 (56.1) | 797 (57.5) | 314 (52.9) | ||
| ≥ 100 | 868 (43.9) | 588 (42.5) | 280 (47.1) | ||
| PLT, *109/L | 2.070 | 0.150 | |||
| < 160 | 1286 (65.0) | 914 (66.0) | 372 (62.6) | ||
| ≥ 160 | 693 (35.0) | 471 (34.0) | 222 (37.4) | ||
| PLR | 0.703 | 0.402 | |||
| < 150.84 | 1158 (58.5) | 802 (57.9) | 356 (59.9) | ||
| ≥ 150.84 | 821 (41.5) | 583 (42.1) | 238 (40.1) |
Abbreviations: PE, Pleural effusion; M, median; BMI, Body Mass Index; COPD, Chronic Obstructive Pulmonary Disease; PCI, Percutaneous Transluminal Coronary Intervention; CAS, Carotid artery stenosis; EF, Preoperative Left Ventricular Ejection Fraction; ICU, Intensive Care Unit; CKMB, creatine kinase myocardial band; ALB, albumin; TC, total cholesterol; TG, triglyceride; HDL-C, high density lipoprotein cholesterol; LDL-C, low density lipoprotein cholesterol; AST, Aspartate Transaminase; ALT, alanine transaminase; CK, Creatine Kinase; LDH, Lactate Dehydrogenas; GGT, Gamma Glutamyl Transpeptidase; BUN, Blood Urea Nitrogen; Cre, Creatinine; cTnI, Cardiac troponin I; WBC, White Blood Cells; RBC, Red Blood Cells; NEUT, Neutrophil; LYM, Lymphocyte; NLR, Neutrophil/Lymphocyte; EOS, Eosinophil; HB, Hemoglobin; PLT, Platelets; PLR, Platelets/Lymphocyte. Data are presented as number (%).
Validation of the clinical validity of the model
During the internal validation process, it was identified that the BPNN and GBM demonstrated superior predictive capabilities. To comprehensively evaluate their generalization capabilities, robustness, and credibility, data were collected from 289 patients who underwent CABG at the Department of Cardiovascular Surgery, the First Affiliated Hospital of Guangxi Medical University. In the external dataset, the AUC for the BPNN was 0.737 (95% CI: 0.655-0.818), while for the GBM, it was 0.710 (95% CI: 0.631-0.790). Based on Delong’s test, no statistically significant difference was found in the AUC between the BPNN and GBM (Z = 0.563, P = 0.573). Additionally, calibration plots revealed minimal discrepancies in the predictive performance of the models. Nevertheless, the BPNN demonstrated clinical benefits across a broader threshold range (Figure 8). It is concluded that the BPNN outperforms other prediction models in all respects, making it the optimal choice for prediction.
Figure 8.
External validation of BPNN and GBM. A and D. ROC curve; B and E. Calibration curve; C and F. Decision curve. Abbreviations: BPNN, back-propagation neural network; GBM, gradient boosting model.
Discussion
Pleural effusion following CABG significantly contributes to the rehospitalization rate of patients, as reported in [9-11]. This condition is influenced by a wide range of factors and mechanisms. Historically, efforts to predict and stratify risk based on individual factors have been insufficient to meet clinicians’ expectations. Consequently, a risk model that incorporates a composite of various indicators provides a more effective approach for accurate evaluation. In our study, BMI, coronary artery stenosis, postoperative pneumonia, mechanical ventilation time, surgical blood loss, operation duration, and ejection fraction were identified as significant factors associated with the development of postoperative PE following CABG. By focusing on these predictors, five models were developed to estimate the individualized risk of PE following CABG.
In this study, the occurrence of PE in patients three days following CABG was approximately 71% (1,405/1,979). Rezk et al. [23] reported a lower incidence of postoperative PE in patients undergoing CABG, with an incidence of 18% within one to two days post-surgery. Brookes et al. [9] found that approximately 14% (330/409) of patients required thoracic drainage following CABG. The differing incidence rates can be attributed to multiple factors. Firstly, they depend on the study’s critical value and hard endpoint, such as whether drainage is considered a key factor. Studies focusing on smaller amounts of effusion tend to report higher incidence rates due to sensitive detection methods. Pleural effusion volume varies significantly among individuals. A small volume may be absorbed spontaneously, while a large volume requires intervention. Studies focused on hard endpoints, like drainage, and only including severe cases may result in a lower incidence rate. Furthermore, different studies define meaningful effusion volume differently. Additionally, postoperative effusion changes over time, and the timing of evaluation can influence the results. The high proportion of patients with small-volume effusion in this study may contribute to the higher statistical incidence rate.
Among numerous factors, being overweight and obesity are recognized as common risk factors for coronary artery disease [23] and significant contributors to mortality following CABG [24]. A BMI greater than 40 kg/m2 is associated with an increased risk of deep and organ-space surgical site infections following CABG [25]. These infections, in turn, can elevate the probability of thoracic infections. Obesity can also impair the body’s specific and non-specific immune responses, exacerbating bacterial infections and predisposing individuals to chest infections [26], which can lead to postoperative PE. Pneumonia is another common complication following CABG [27]. Our findings demonstrate a significant association between postoperative pneumonia and the risk of developing postoperative PE. Furthermore, in patients admitted to the ICU, prolonged mechanical ventilation is associated with an increased risk of developing PE [28]. This finding is consistent with the research by Akhtar et al. [7], which indicated that patients requiring pleural effusion drainage tend to have longer intubation times. Prompt removal of ventilators in CABG patients can effectively improve cardiopulmonary function, minimize the risk of pulmonary embolism (PE), shorten ICU and hospital stays, and subsequently reduce healthcare costs [29].
Our observations demonstrate that the risk of developing postoperative PE increases by 1.349-fold when the surgical duration exceeds four hours, suggesting that prolonged operation time contributes to the occurrence of postoperative PE. During CABG, extended exposure to inflammatory cytokines in the lungs and alveoli has been reported to cause swelling, reduced lung compliance and expiratory volume, and lung injury [30]. The longer the surgical procedure lasts, the longer the chest is exposed to air, which may cause trauma or ischemia-reperfusion injury to the heart and lung tissue. This leads to an elevated inflammatory response and increased vascular permeability, thereby raising the likelihood of postoperative chest infection. The EF plays a crucial role in predicting the risk of postoperative PE. In the European System for Cardiac Operative Risk Evaluation (Euro SCORE II) [31], the EF is considered a vital factor for comprehensive surgical risk assessment.
Due to the substantial risks associated with perioperative cardiac surgery, several risk assessment systems have been developed to quantify the probability of mortality and postoperative complications following CABG. These include Euro SCORE [31], the Society of Cardiothoracic and Vascular Anesthesia Score (Sino SCORE) [32], and the SSII-CABG model [33]. Euro SCORE and Sino SCORE predict the likelihood of death within 30 days post-surgery, while the SSII-CABG model mainly focuses on long-term survival outcomes, particularly over a four-year period.
Although numerous predictive models exist that utilize various factors to assess the risk of postoperative PE, most studies focus primarily on predicting PE after tumor resection, such as liver cancer [34], and lung cancer [35]. Additionally, many studies are dedicated to developing diagnostic models to distinguish between benign and malignant pleural effusion [36-39]. However, the sample sizes of these models tend to be limited, and the variables they incorporate mainly consist of preoperative indicators. A significant gap remains in the development of a comprehensive risk assessment framework specifically designed for postoperative PE following cardiac surgery. Consequently, we believe it is crucial to develop a risk model capable of accurately assessing the probability of postoperative PE in patients undergoing CABG.
By utilizing machine learning algorithms, five predictive models for postoperative PE were successfully developed in our research. These models incorporated a diverse set of predictors, including surgery-related factors, preoperative considerations, and postoperative multidimensional variables. Specifically, BMI, coronary artery stenosis, postoperative pneumonia, duration of mechanical ventilation, surgical blood loss, operation duration, and EF were identified as critical predictors in our models. In terms of prediction performance, our analysis indicated that the BPNN significantly outperformed the other four models, namely the nomogram, RF, GBM, and SVM. Consequently, BPNN was chosen as the preferred predictive model. The ability of BPNN to effectively approximate complex nonlinear relationships makes it particularly suitable for handling problems with intricate internal mechanisms [40].
However, it is important to note the limitations of BPNN. One major drawback is its poor interpretability. Unlike a nomogram, which presents a visual and intuitive approach to understanding the relationship between variables and the predicted outcome, the internal workings of BPNN are often regarded as a “black box”. It is challenging to precisely explain how each input variable contributes to the final prediction, which could limit its application in fields where interpretability is crucial, such as clinical decision-making, where doctors need to understand the reasoning behind a prediction. Therefore, further research is needed to explore ways to improve its interpretability, such as developing novel methods for feature importance analysis or visualization techniques, to enhance its applicability in clinical settings.
Our study has some limitations that need to be acknowledged. The primary aim of this study was to analyze the risk factors for pleural effusion after CABG and construct a predictive model to reduce its incidence. However, most of the patients included in the study had a small amount of pleural effusion. Even if these patients developed PE after surgery, it often resolved spontaneously without requiring special intervention. Therefore, the clinical value of predicting this type of situation is limited. Future studies could focus on patients with severe pleural effusion for in-depth analysis. In addition, this study only included the available risk factors and lacked detailed information on surgical techniques used (such as open and closed pleural procedures) and postoperative care plans (such as extubation strategies or physical therapy programs). This may result in an incomplete assessment of the risk of PE.
In conclusion, our study has demonstrated that pleural effusion is a prevalent complication after coronary artery bypass grafting, with a relatively high incidence rate. To address this issue, we have effectively developed a prediction model for PE with excellent predictive performance. This model is invaluable for the early identification of high-risk patients for PE.
Disclosure of conflict of interest
None.
References
- 1.Forrest IS, Petrazzini BO, Duffy A, Park JK, Marquez-Luna C, Jordan DM, Rocheleau G, Cho JH, Rosenson RS, Narula J, Nadkarni GN, Do R. Machine learning-based marker for coronary artery disease: derivation and validation in two longitudinal cohorts. Lancet. 2023;401:215–225. doi: 10.1016/S0140-6736(22)02079-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Cheraghi M, Pooria A. A review: nanofibrous scaffold in possible prevention and treatment of coronary artery disease. Biotechnol Appl Biochem. 2019;66:478–483. doi: 10.1002/bab.1750. [DOI] [PubMed] [Google Scholar]
- 3.Kusu-Orkar TE, Kermali M, Oguamanam N, Bithas C, Harky A. Coronary artery bypass grafting: factors affecting outcomes. J Card Surg. 2020;35:3503–3511. doi: 10.1111/jocs.15013. [DOI] [PubMed] [Google Scholar]
- 4.Krittanawong C, Rizwan A, Khawaja M, Newman N, Escobar J, Virk HUH, Alam M, Al-Azzam F, Yong CM, Jneid H. The current state of coronary revascularization: coronary artery bypass graft surgery versus percutaneous coronary interventions. Curr Cardiol Rep. 2024;26:919–933. doi: 10.1007/s11886-024-02090-x. [DOI] [PubMed] [Google Scholar]
- 5.Lusquinhos J, Tavares M, Abelha F. Postoperative pulmonary complications and perioperative strategies: a systematic review. Cureus. 2023;15:e38786. doi: 10.7759/cureus.38786. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Montrief T, Koyfman A, Long B. Coronary artery bypass graft surgery complications: a review for emergency clinicians. Am J Emerg Med. 2018;36:2289–2297. doi: 10.1016/j.ajem.2018.09.014. [DOI] [PubMed] [Google Scholar]
- 7.Akhtar M, Razick DI, Saeed A, Baig O, Kamran R, Ansari U, Sajid Z, Rahman JE. Complications and outcomes of the nuss procedure in adult patients: a systematic review. Cureus. 2023;15:e35204. doi: 10.7759/cureus.35204. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Nakamura E, Sofue T, Higashitani M, Saiki K, Yamamoto T, Shiga T, Aoki Y, Shiraishi A, Kunisho Y, Onishi K, Kato A, Minamino T. A case of a peritoneal dialysis patient with left pleuroperitoneal communication caused by a pericardial defect after coronary artery bypass surgery. CEN Case Rep. 2024;13:457–462. doi: 10.1007/s13730-024-00867-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Brookes JDL, Williams M, Mathew M, Yan T, Bannon P. Pleural effusion post coronary artery bypass surgery: associations and complications. J Thorac Dis. 2021;13:1083–1089. doi: 10.21037/jtd-20-2082. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Jain A, Devarajan A, Assallum H, Malekan R, Lanier GM, Epelbaum O. Characteristics of early pleural effusions after orthotopic heart transplantation: comparison with coronary artery bypass graft surgery. Pleura Peritoneum. 2021;6:161–165. doi: 10.1515/pp-2021-0143. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Feng TR, White RS, Gaber-Baylis LK, Turnbull ZA, Rong LQ. Coronary artery bypass graft readmission rates and risk factors - a retrospective cohort study. Int J Surg. 2018;54:7–17. doi: 10.1016/j.ijsu.2018.04.022. [DOI] [PubMed] [Google Scholar]
- 12.Zhong Z, Yuan X, Liu S, Yang Y, Liu F. Machine learning prediction models for prognosis of critically ill patients after open-heart surgery. Sci Rep. 2021;11:3384. doi: 10.1038/s41598-021-83020-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Ma H, Chen D, Lv W, Liao Q, Li J, Zhu Q, Zhang Y, Deng L, Liu X, Wu Q, Liu X, Yang Q. Performance of an AI prediction tool for new-onset atrial fibrillation after coronary artery bypass grafting. EClinicalMedicine. 2025;81:103131. doi: 10.1016/j.eclinm.2025.103131. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Zhang H, Qiao H, Yang B, Lu Y, Bai T, Xue J, Liu Y. Development and validation of a diagnostic model based on left atrial diameter to predict postoperative atrial fibrillation after off-pump coronary artery bypass grafting. J Thorac Dis. 2023;15:3708–3725. doi: 10.21037/jtd-22-1706. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Roberts ME, Rahman NM, Maskell NA, Bibby AC, Blyth KG, Corcoran JP, Edey A, Evison M, de Fonseka D, Hallifax R, Harden S, Lawrie I, Lim E, McCracken DJ, Mercer R, Mishra EK, Nicholson AG, Noorzad F, Opstad K, Parsonage M, Stanton AE, Walker S BTS Pleural Guideline Development Group. British Thoracic Society Guideline for pleural disease. Thorax. 2023;78(Suppl 3):s1–s42. doi: 10.1136/thorax-2022-219784. [DOI] [PubMed] [Google Scholar]
- 16.Buch G, Schulz A, Schmidtmann I, Strauch K, Wild PS. A systematic review and evaluation of statistical methods for group variable selection. Stat Med. 2023;42:331–352. doi: 10.1002/sim.9620. [DOI] [PubMed] [Google Scholar]
- 17.Vasquez MM, Hu C, Roe DJ, Halonen M, Guerra S. Measurement error correction in the least absolute shrinkage and selection operator model when validation data are available. Stat Methods Med Res. 2019;28:670–680. doi: 10.1177/0962280217734241. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Sun T, Liu J, Yuan H, Li X, Yan H. Construction of a risk prediction model for lung infection after chemotherapy in lung cancer patients based on the machine learning algorithm. Front Oncol. 2024;14:1403392. doi: 10.3389/fonc.2024.1403392. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Yang L, Wu H, Jin X, Zheng P, Hu S, Xu X, Yu W, Yan J. Study of cardiovascular disease prediction model based on random forest in eastern China. Sci Rep. 2020;10:5245. doi: 10.1038/s41598-020-62133-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Jan Ben S, Dorner M, Gunther MP, von Kanel R, Euler S. Proof of concept: predicting distress in cancer patients using back propagation neural network (BPNN) Heliyon. 2023;9:e18328. doi: 10.1016/j.heliyon.2023.e18328. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Wang W, Sheng R, Liao S, Wu Z, Wang L, Liu C, Yang C, Jiang R. LightGBM is an effective predictive model for postoperative complications in gastric cancer: a study integrating radiomics with ensemble learning. J Imaging Inform Med. 2024;37:3034–3048. doi: 10.1007/s10278-024-01172-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Huang S, Cai N, Pacheco PP, Narrandes S, Wang Y, Xu W. Applications of support vector machine (SVM) learning in cancer genomics. Cancer Genomics Proteomics. 2018;15:41–51. doi: 10.21873/cgp.20063. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Rezk ME, Elgazzar MA, Abo Youssef SM, Emeraa AS, Elkafoury AE, Moussa HH. Open versus closed pleura internal mammary artery harvesting and early pulmonary function after coronary artery bypass grafting. Heart Lung Circ. 2020;29:1412–1417. doi: 10.1016/j.hlc.2019.09.014. [DOI] [PubMed] [Google Scholar]
- 24.Pencina MJ, Navar AM, Wojdyla D, Sanchez RJ, Khan I, Elassal J, D’Agostino RB Sr, Peterson ED, Sniderman AD. Quantifying importance of major risk factors for coronary heart disease. Circulation. 2019;139:1603–1611. doi: 10.1161/CIRCULATIONAHA.117.031855. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Lv M, Gao F, Liu B, Pandey P, Feng Y, Wang Y, Zhang Y, Li Z. The effects of obesity on mortality following coronary artery bypass graft surgery: a retrospective study from a single center in China. Med Sci Monit. 2021;27:e929912. doi: 10.12659/MSM.929912. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Bajpai G, Nahrendorf M. Infectious and lifestyle modifiers of immunity and host resilience. Immunity. 2021;54:1110–1122. doi: 10.1016/j.immuni.2021.05.011. [DOI] [PubMed] [Google Scholar]
- 27.Bartlett B, Sanfilippo FM, Lee S, Ludewick H, Waterer G, Rajwani A, Bharat C, Ihdayhid AR, Corrales-Medina V, Dwivedi G. The risk of adverse cardiac events following pneumonia in patients with coronary artery disease. Ann Am Thorac Soc. 2025 doi: 10.1513/AnnalsATS.202407-714OC. [Epub ahead of print] [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Luo Q, Zhao W, Su Z, Liu Y, Jia Y, Zhang L, Wang H, Li Y, Wu X, Li S, Yan F. Risk factors for prolonged pleural effusion following total cavopulmonary connection surgery: 9 years’ experience at Fuwai Hospital. Front Pediatr. 2019;7:456. doi: 10.3389/fped.2019.00456. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Pooria A, Pourya A, Gheini A. Postoperative complications associated with coronary artery bypass graft surgery and their therapeutic interventions. Future Cardiol. 2020;16:481–496. doi: 10.2217/fca-2019-0049. [DOI] [PubMed] [Google Scholar]
- 30.Guo MH, Toubar O, Issa H, Glineur D, Ponnambalam M, Vo TX, Rahmouni K, Chong AY, Ruel M. Long-term survival, cardiovascular, and functional outcomes after minimally invasive coronary artery bypass grafting in 566 patients. J Thorac Cardiovasc Surg. 2024;168:1080–1088. e2. doi: 10.1016/j.jtcvs.2023.07.047. [DOI] [PubMed] [Google Scholar]
- 31.Halvorsen S, Mehilli J, Cassese S, Hall TS, Abdelhamid M, Barbato E, De Hert S, de Laval I, Geisler T, Hinterbuchner L, Ibanez B, Lenarczyk R, Mansmann UR, McGreavy P, Mueller C, Muneretto C, Niessner A, Potpara TS, Ristic A, Sade LE, Schirmer H, Schupke S, Sillesen H, Skulstad H, Torracca L, Tutarel O, Van Der Meer P, Wojakowski W, Zacharowski K ESC Scientific Document Group. 2022 ESC Guidelines on cardiovascular assessment and management of patients undergoing non-cardiac surgery. Eur Heart J. 2022;43:3826–3924. doi: 10.1093/eurheartj/ehac270. [DOI] [PubMed] [Google Scholar]
- 32.Peng K, McIlroy DR, Bollen BA, Billings FT 4th, Zarbock A, Popescu WM, Fox AA, Shore-Lesserson L, Zhou S, Geube MA, Ji F, Bhatia M, Schwann NM, Shaw AD, Liu H. Society of cardiovascular anesthesiologists clinical practice update for management of acute kidney injury associated with cardiac surgery. Anesth Analg. 2022;135:744–756. doi: 10.1213/ANE.0000000000006068. [DOI] [PubMed] [Google Scholar]
- 33.Yuksel Y, Yildiz C, Kose S. Assessment of predictive value of syntax-ii score for adverse cardiac events and clinical outcomes in patients with acute coronary syndrome. Angiology. 2024;75:754–763. doi: 10.1177/00033197231181958. [DOI] [PubMed] [Google Scholar]
- 34.Nitta H, Mitsuura C, Shiraishi Y, Miyata T, Shimizu K, Harada K, Karashima R, Masuda T, Matsumoto K, Okino T, Yamashita YI, Baba H, Takamori H. Predictive model for postoperative pleural effusion after hepatectomy. Ann Gastroenterol Surg. 2020;5:373–380. doi: 10.1002/ags3.12417. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Li R, Qiu J, Qu C, Ma Z, Wang K, Zhang Y, Yue W, Tian H. Comparison of perioperative outcomes with or without routine chest tube drainage after video-assisted thoracoscopic pulmonary resection: a systematic review and meta-analysis. Front Oncol. 2022;12:915020. doi: 10.3389/fonc.2022.915020. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Wang S, Tian S, Li Y, Zhan N, Guo Y, Liu Y, Xu J, Ma Y, Zhang S, Song S, Geng W, Xia H, Ma P, Wang X, Liao T, Duan Y, Jin Y, Dong W. Development and validation of a novel scoring system developed from a nomogram to identify malignant pleural effusion. EBioMedicine. 2020;58:102924. doi: 10.1016/j.ebiom.2020.102924. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Wang S, Tan X, Li P, Fan Q, Xia H, Tian S, Pan F, Zhan N, Yu R, Zhang L, Duan Y, Xu J, Ma Y, Chen W, Li Y, Zhao Z, Liu C, Bao Q, Yang L, Jin Y. Differentiation of malignant from benign pleural effusions based on artificial intelligence. Thorax. 2023;78:376–382. doi: 10.1136/thorax-2021-218581. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Pan Y, Bai W, Chen J, Mao Y, Qian X, Xu K, Tang S, Zhang J, Chen C, Chen J, Hu X. Diagnosing malignant pleural effusion using clinical and analytical parameters. J Clin Lab Anal. 2019;33:e22689. doi: 10.1002/jcla.22689. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39.Ferreiro L, Gude F, Toubes ME, Lama A, Suarez-Antelo J, San-Jose E, Gonzalez-Barcala FJ, Golpe A, Alvarez-Dobano JM, Rabade C, Rodriguez-Nunez N, Diaz-Louzao C, Valdes L. Predictive models of malignant transudative pleural effusions. J Thorac Dis. 2017;9:106–116. doi: 10.21037/jtd.2017.01.12. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40.Akella A, Akella S. Machine learning algorithms for predicting coronary artery disease: efforts toward an open source solution. Future Sci OA. 2021;7:FSO698. doi: 10.2144/fsoa-2020-0206. [DOI] [PMC free article] [PubMed] [Google Scholar]