Comparison of Machine Learning Models for Difficult Airway

Abstract

Background: The accurate prediction of difficult airway (DA) is important in ICU and general anaesthesia. Our hypothesis is that machine learning models can predict difficult tracheal intubation (DTI) and difficult laryngoscopy (DL). Methods: We performed a secondary analysis of two prospective observational DA research programmes. DTI and DL prediction models were established by machine learning based on multivariate data. Machine learning algorithms, such as logistic regression, support vector machine, and random forest, were used. Five times repeated 5-fold cross-validation were used to compare parameters such as the area under the receiver operating characteristic (ROC) curve (AUC), recall rate, accuracy and the F1 score. The feature importance of the indicators were analysed by the random forest and AdaBoost models. Results: 3958 tracheal intubation patients were included in this study. Among the five machine learning algorithms, the best AUCs were obtained by the Bayes model for DTI (0.956, 95% CI 0.950–0.961) and the random forest model for DL (0.903, 95% CI 0.895–0.911). The random forest model had the best accuracy for DTI (0.966, 95% CI 0.964–0.968) and DL (0.926, 95% CI 0.923–0.929). The random forest model also had the highest F1 score for DTI (0.361, 95% CI 0.327–0.394) and DL (0.530, 95% CI 0.492–0.569). The naïve Bayes model had the highest recall rate for DTI (0.902, 95% CI 0.877–0.927) and DL (0.809, 95% CI 0.787–0.830). Conclusions: Machine learning algorithms based on multivariate indicators were effective in predicting DA.