To identify risk factors for postoperative respiratory failure (PORF) in cardiovascular surgery patients using machine learning algorithms and to construct a specific risk prediction model.

A retrospective cohort was conducted on 1 623 patients undergoing cardiovascular surgery between 2011 and 2020 from the INSPIRE database. Following data quality analysis, multiple imputation was employed to handle missing data, and the Boruta algorithm was used for feature selection. Eight machine learning models were constructed based on the selected features, including Gradient Boosting Machine (GBM), Generalized Linear Model (GLM), Extreme Gradient Boosting (XGBoost), K-Nearest Neighbors (KNN), Neural Network (NNET), Naive Bayes (NB), Support Vector Machine (SVM), and Random Forest (RF). Model performance was evaluated using metrics including the area under the curve (AUC), sensitivity, and specificity. Variables significantly influencing PORF were identified using the permutation importance algorithm.

The overall incidence of PORF was 27.05% (439/1 623). The in-hospital mortality rate was significantly higher in the PORF group than the non-PORF group (12.98% vs 1.60%, P<0.001). Among the developed models, the SVM model demonstrated the best performance, achieving an AUC of 0.705 in the testing set, with a sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of 0.481, 0.825, 0.504, and 0.812, respectively. Based on feature importance analysis, the top 10 variables most predictive of PORF were anesthesia duration, arterial partial pressure of carbon dioxide (PaCO₂), calcium level, lymphocyte percentage, cardiopulmonary bypass duration, intraoperative blood loss, age, creatinine level, aspartate aminotransferase (AST) level, and activated partial thromboplastin time (aPTT).

A predictieon model for PORF following cardiovascular surgery is successfully developed. This model can identify high-risk patients and estimate their probability of developing respiratory failure, thereby facilitating data-driven clinical decision-making.