1. Ethical Statement
In this study, the Helsinki protocol was followed. The study was approved by our local ethical committee in Iran University of Medical Sciences (RHC.AC.IR.REC.1396.28) and patients’ written informed consent was waived by the ethical committee at the Iran University of Medical Sciences.
2. Patient and Study design
In this prospective cohort, we gathered the data of patients from their health records in a hospital center for cardiovascular diseases (Rajaei Cardiovascular Medical and Research Center). In this study, we included patients who were admitted to the hospital from September 2016 to September 2017 and were candidates for elective CABG, congenital heart disease, large vessel or valve replacement surgeries. Those who underwent emergency surgeries were excluded and a total of 918 patients were included in our study.
Before the operations, the demographic data of the patients and the history of underlying diseases such as diabetes mellitus, hypertension, renal dysfunction, and pulmonary diseases (e.g. COPD) were obtained through the medical records. History of smoking, home bakery, opium usage was directly asked from the patients who then underwent the measurements of BMI and neck circumference. Laboratory tests for the analysis of the collected blood samples were performed for blood urea nitrogen (BUN), creatinine, uric acid, fasting blood sugar (FBS), and hemoglobin (Hb). Glomerular filtration rate (GFR) was calculated using plasma creatinine with the Cockroft-Gault formula. Oxygen Saturation and PaO2 (partial pressure of oxygen) were added to the patient data 15 minutes prior to the operation. Besides, all of the patients were sent to undergo pulmonary function tests (PFT) to obtain data such as FEV1, FVC, FEV1 to FVC ratio, MMEF, and PEFR. Echocardiography was performed routinely and measurements of pulmonary artery pressure (SPAP), ejection fraction (EF), and diastolic dysfunction were assessed. Epworth Sleepiness Scale (ESS) sheets were filled up by an educated nurse on the day of admission. All of the patients were on the pump during the surgery and were performed median sternotomy regardless of the type of the operation. During the operation, the intraoperative feature set was completed by adding the type of the surgery, the time in which the patients were on the pump, the skin to skin time, and the clamp time. In the end, we included a positive blood transfusion if the patient received any of the blood products during the operation. The endpoint of this study was the development of pulmonary complications between 24 and 72 hours after the surgery. This can be specified as the diagnosis of one of the following: atelectasis; pleural effusion; pulmonary edema; pneumonia; and diaphragm paralysis, made by an expert pulmonologist by reviewing and commenting on chest X-rays and CT images.
3. Feature Selection
Feature selection is one of the core concepts in ML that hugely impacts the performance of the model. The features used to train ML models have a huge influence on the performance that can be achieved. Irrelevant or partially relevant features can negatively impact model performance. Details of the preoperative and intraoperative features are provided in Table 1.
We used two machine learning algorithms novel in the field of image processing, namely Multivariate Adaptive Regression Splines (MARS) for both the selection and prediction analysis, and Bayesian network classifier to assess the effects of selected features on the risk of PPC. The MARS method builds a model as described in equation (1) [18]:
As the equation above shows, the model sums the basis functions , each weighted by a (constant coefficient). The basis function can be either a constant 1, a hinge function, or a product of two or more of the latter. This will help us gain a more flexible and easier-to-understand model compared with other ML algorithms such as random forest or neural networks. MARS automatically selects the most important features, ranks them in order of their “importance value”, and builds the model on them. Therefore, it does not need any previous preparation of data and also suits large datasets and complex data structures. Finally, the most important advantage of MARS is that it can handle features of any type such as numerical, categorical, and nominal ones, either binary or not. We employed this easily accessible algorithm in the “earth” package of the software R version 4.0.2.
Bayesian Network (BN) method is annotated directed graphs that represent a set of variables (selected features) as nodes in a network, connected by edges representing the conditional probabilistic relations between them [19,20]. BN is based on the Bayes theorem [21] and it consists of two steps: 1) structural learning to find the global optimum global structure 2) parametric learning to estimate the conditional probability. The structure of the BN model was built with Bootstrap by repeating 100 times and the Tabu search algorithm merging data information and prior knowledge. In addition, parameters of the model are learned with Maximum Likelihood Estimate (MLE) and 10-fold cross-validation was used. BN method is used to select the most important features [20]. A Bayesian network with uniform (non-informative) prior distributions was used in this study. A constraint-based algorithm (PC algorithm) was used to determine the BN structure using conditional independence tests. Then, we performed the Expectation–maximization (EM) algorithm as an iterative approach for parameter estimation in our model.
4. Model performance and evaluation
After splitting the dataset into training and test, the MARS algorithm was employed to achieve the performance of the features in each of the training and testing datasets, assessed by the area under the receiver operating characteristic curve (AUC) along with its 95% confidence interval (CI). ROC curve analysis was performed by the “earth” R package. Additionally, the odds ratio (OR) with its 95% CI was achieved by utilizing the logistic regression (glm R function). Finally, statistical comparison of area under ROC curve was done by “pROC” R package [22]. Therefore, we could discover the possible significant prognostic performances of the features. The significance level of the P-value was considered <0.05.