Early Prediction of Invasive Mechanical Ventilation Needs in Patients Admitted to Intensive Care Units: A Novel Ensemble Learning Approach

Accurately predicting the needs for invasive mechanical ventilation (IMV) in patients admitted to intensive care units (ICU) at an early stage is crucial as it helps improve patient outcomes and optimize medical resource use. However, this task is challenging because of diverse medical data types, complex data correlations, missing values, and imbalanced prediction classes. In this study, we propose a novel ensemble learning approach that combines traditional machine learning models and deep learning models to achieve superior prediction performance in predicting IMV needs. Applied to 30,313 valid ICU stays filtered from the publicly available MIMIC-IV database, our approach achieves state-of-the-art performance with an area under the receiver operating characteristic curve (AUC) of 0.94 in predicting IMV needs for ICU patients within six hours following their initial four-hour ICU stays. This surpasses the highest AUC of 0.935 reported in previous studies and sets a new benchmark for early prediction of IMV needs. This approach uses traditional machine learning models to capture relationships in static features and deep learning models to capture temporal dependencies in longitudinal features. Then their predictions are combined to produce the final predictions. Additionally, the feature importance analysis enhances interpretability by identifying the most influential features for predicting IMV needs, making the predictions transparent and actionable. By using static and longitudinal features routinely collected from ICU patients, our approach is both practical and highly extensible, readily adaptable to other electronic health record (EHR) databases and other clinical prediction tasks.