Abstract
We applied machine learning to predicting difficult tracheal intubations and successful intubation at first intubation attempt (first-pass success) in the emergency department. While conventional methods (e.g., mLEMON) have been used to predict difficult intubations, their prediction ability is suboptimal. Additionally, there has been no clinically-meaningful model that predicts first-pass success. In the current study, we used prospective data (n=10,816) to develop prediction models using machine learning and examine their performance. We used patient characteristics and vital signs for predicting difficult airway, and airway management data for predicting first-pass success. The c-statistics of machine learning models for predicting difficult airway was higher compared to that of mLEMON as the reference (e.g., ensemble method, 0.73 [95%CI 0.67-0.79] vs. mLEMON, 0.62 [95%CI 0.58-0.65], p<0.01). Similarly, the machine learning models for predicting first-pass success had higher discriminative ability compared to the reference logistic regression model (e.g., gradient boosting, 0.82 [95%CI 0.80-0.84] vs. logistic regression, 0.60 [95%CI 0.58-0.63], p<0.01).
