Development of an Artificial Intelligence Model to Predict Endotracheal Intubation in Critically Ill Patients in Real...

Background/Objectives: In critically ill patients, endotracheal intubation (EI) is often performed to secure the airway or mechanical ventilation. Accurately predicting the timing of intubation significantly affects patient outcomes. We developed an artificial intelligence (AI) model designed for real-time risk stratification of patients requiring EI. Methods: We utilized the Medical Information Mart for Intensive Care-IV (MIMIC-IV) 2.2 dataset and performed model development using 15 clinical variables, including vital signs, Glasgow Coma Scale (GCS) score, and arterial blood gas analysis results. Patients intubated before or within 1 h of intensive care unit (ICU) admission were excluded. Clinical data from the ICU inherently consists of continuous time-series measurements. Traditional machine learning models often treat this information as static tabular data, neglecting vital temporal dynamics and patient history. Conversely, deep learning time-series approaches can capture these complex patterns over time. Thus, we applied the Gated Recurrent Unit with Decay++ (GRU-D++) model to predict the need for EI. GRU-D++ is an extension of the GRU and GRU-D. It builds upon the GRU-D to provide improved performance when handling datasets with exceptionally high rates of missing values. GRU-D++ is a time series deep learning model with an automatic mechanism for imputing missing values. This built-in capability eliminates the need for additional data preprocessing and has previously demonstrated high predictive performance. Using the 15 variables, we evaluated the optimal timing for EI in ICU-admitted patients by applying various AI models. Results: Among these, the GRU-D++ model demonstrated AUROC of 0.888, AUPR of 0.481, sensitivity of 0.474, specificity of 0.995, precision of 0.511, and F1 score of 0.491 on MIMIC-IV dataset. For KNUH dataset, the model demonstrated AUROC of 0.913, AUPR of 0.063, sensitivity of 0.162, specificity of 0.997, precision of 0.137, and F1 score of 0.147 within the 2 h in advance scenario. Furthermore, when compared with conventional scoring systems such as the Heart rate, Acidosis, Consciousness, Oxygenation, Respiratory rate (HACOR) score and Respiratory rate-Oxygenation (ROX) index, the GRU-D++ model also showed better performance predictive accuracy. Conclusions: The AI-based intubation prediction model developed in this study holds potential as a real-time risk stratification tool, providing timely risk assessments regarding the need EI. While operational threshold recalibration is essential prior to clinical deployment, further prospective multicenter studies are required to validate the clinical utility of this model in real-time practice.