Machine Learning‒Based Predictors of Prolonged Mechanical Ventilation Duration Following Coronary Artery Bypass Grafting.

OBJECTIVE: This study aims to develop a machine learning model for predicting whether patients will experience prolonged mechanical ventilation and to elucidate the roles of various factors in disease prediction. DESIGN: A retrospective study investigating influencing factors of prolonged mechanical ventilation in patients after coronary artery bypass grafting, PARTICIPANTS: The study cohort consisted of patients undergoing coronary artery bypass grafting selected from the MIMIC-IV 2.2 database. The study endpoint was prolonged mechanical ventilation, characterized by mechanical ventilation lasting more than 24 hours. Patients were stratified into normal duration and prolonged duration groups. INTERVENTIONS: Feature selection was performed using recursive partitioning and regression trees (RPART), random forest (RF), light gradient boosting machine (LightGBM), kernel k-nearest neighbors (K-KNN), and naive Bayes (NB) machine learning algorithms. Algorithm performance was assessed through multiple metrics including area under the receiver operating characteristic curve, area under the precision-recall curve, accuracy, misclassification rate, sensitivity, and specificity. The predictive importance of features in the top-performing model was quantified using SHapley Additive exPlanations (SHAP) values. MEASUREMENTS AND MAIN RESULTS: The study cohort comprised 2,356 patients who received coronary artery bypass grafting. The machine learning models incorporated 40 baseline features and 5 composite metrics. Assessment revealed that the random forest model achieved superior comprehensive performance, exhibiting an area under the receiver operating characteristic curve of 0.9976 and an area under the precision-recall curve of 0.9976. It maintained high accuracy and low classification error rate. The model preserved remarkably high sensitivity and specificity. SHAP analysis revealed that delirium, percutaneous oxygen saturation, glycemic variability, and Sequential Organ Failure Assessment (SOFA) score made predominant contributions. CONCLUSION: The random forest model exhibited optimal predictive capability for prolonged mechanical ventilation after coronary artery bypass grafting. Delirium and blood glucose fluctuations served as the key predictive factors in this model.