A novel hybrid ensemble learning approach for the prediction of ventilator weaning success using a bagged lightgbm classifier

Abstract

This study aims to develop the most accurate machine learning model to predict the success of ventilator weaning by classifying weaning trials and identifying the most influential predictive parameters. To classify weaning trials, a novel hybrid bagged light gradient boosting machine model (B-LGBM) was developed, trained, and validated using dataset collected from a cohort of 3,215 patients over a monitoring period of the last 1-hour record prior to the patient’s attempt to wean from the ventilator in the intensive care unit (ICU) employing a bagging approach. The dataset encompasses a total of 69 features, including demographic data, clinical indicators, ventilator settings, respiratory parameters, and patient outcomes. Model performance was evaluated based on accuracy, precision, recall, F1-score, specificity, and area under the receiver operating characteristic curve (AUC-ROC). Results revealed that the B-LGBM model outperformed other algorithms, achieving an AUC-ROC of 91.45%, accuracy of 89.08%, precision of 95.64%, and specificity of 83.33%. Key predictive parameters for successful weaning included pulse oximetry saturation, fraction of inspired oxygen, respiratory rate, expiratory minute volume, total ventilated hours, peak pressure, and partial pressure of arterial carbon dioxide. Accurate weaning predictions with a high AUC-ROC rate can be achieved for the clinical decision-making process with the B-LGBM model developed using ensemble techniques.