Predictive model for difficult laryngoscopy using machine learning: retrospective cohort study

Abstract Background Both predictions and predictors of difficult laryngoscopy are controversial. Machine learning is an excellent alternative method for predicting difficult laryngoscopy. This study aimed to develop and validate practical predictive models for difficult laryngoscopy through machine learning. Methods Variables for the prediction of difficult laryngoscopy included age, Mallampati grade, body mass index, sternomental distance, and neck circumference. Difficult laryngoscopy was defined as grade 3 and 4 by the Cormack-Lehane classification. Pre-anesthesia and anesthesia data of 616 patients who had undergone anesthesia at a single center were included. The dataset was divided into a base training set (n = 492) and a base test set (n = 124), with equal distribution of difficult laryngoscopy. Training data sets were trained with six algorithms (multilayer perceptron, logistic regression, supportive vector machine, random forest, extreme gradient boosting, and light gradient boosting machine), and cross-validated. The model with the highest area under the receiver operating characteristic curve (AUROC) was chosen as the final model, which was validated with the test set. Results The results of cross-validation were best using the light gradient boosting machine algorithm with Mallampati score x age and sternomental distance as predictive model parameters. The predicted AUROC for the difficult laryngoscopy class was 0.71 (95% confidence interval, 0.59-0.83; p= 0.014), and the recall (sensitivity) was 0.85. Conclusion Predicting difficult laryngoscopy is possible with three parameters. Severe damage resulting from failure to predict difficult laryngoscopy with high recall is small with the reported model. The model's performance can be further enhanced by additional data training.

Predictive model for difficult laryngoscopy using machine learning: retrospective cohort study.

BACKGROUND: Both predictions and predictors of difficult laryngoscopy are controversial. Machine learning is an excellent alternative method for predicting difficult laryngoscopy. This study aimed to develop and validate practical predictive models for difficult laryngoscopy through machine learning. METHODS: Variables for the prediction of difficult laryngoscopy included age, Mallampati grade, body mass index, sternomental distance, and neck circumference. Difficult laryngoscopy was defined as grade 3 and 4 by the Cormack-Lehane classification. Pre-anesthesia and anesthesia data of 616 patients who had undergone anesthesia at a single center were included. The dataset was divided into a base training set (n = 492) and a base test set (n = 124), with equal distribution of difficult laryngoscopy. Training data sets were trained with six algorithms (multilayer perceptron, logistic regression, supportive vector machine, random forest, extreme gradient boosting, and light gradient boosting machine), and cross-validated. The model with the highest area under the receiver operating characteristic curve (AUROC) was chosen as the final model, which was validated with the test set. RESULTS: The results of cross-validation were best using the light gradient boosting machine algorithm with Mallampati score x age and sternomental distance as predictive model parameters. The predicted AUROC for the difficult laryngoscopy class was 0.71 (95% confidence interval, 0.59-0.83; p =  0.014), and the recall (sensitivity) was 0.85. CONCLUSION: Predicting difficult laryngoscopy is possible with three parameters. Severe damage resulting from failure to predict difficult laryngoscopy with high recall is small with the reported model. The model's performance can be further enhanced by additional data training.

Effects of different polymerases of avian influenza viruses on the growth and pathogenicity of A/Puerto Rico/8/1934 (H1N1)-derived reassorted viruses.

We generated reassorted PR8 viruses containing six different combinations of avian influenza virus (AIV) polymerase genes from A/chicken/Korea/01310/2001 (H9N2) (01310) and A/chicken/Korea/KBNP-0028/2000 (H9N2) (0028) to examine the effects of the AIV polymerase genes PB1, PB2, and PA on replication efficiency in different host cells and pathogenicity in mice. The virus titers of the reassorted viruses possessing 01310 [rPR8-PB2(01310)] and 0028 [rPR8-PB2(0028)] PB2 genes were significantly higher than those of the others except the rPR8 virus in embryonated chicken eggs at 37°C, and those of avian polymerase reassorted viruses were significantly less than rPR8 in MDCK cells at 32 and 37°C. rPR8-PB2(01310), rPR8-PB2(0028), and rPR8-PA(0028) caused no body weight loss in BALB/c mice but rPR8-PA(01310), rPR8-PB1(01310), and rPR8-PB1(0028) caused mortality and significantly different body weight loss compared to those in the mock treatment. In contrast to rPR8-PB2(0028) and rPR8-PA(0028), rPR8-PB2(01310) was not isolated from infected mice, and rPR8-PB1(0028) was less pathogenic than rPR8-PB1(01310). We determined the amino acid residues that were specific to the less pathogenic polymerases. A comparison with those of pandemic 2009 H1N1, human fatal H5N1 and H7N9, and pathogenic AIVs to mice without adaptation revealed that they possessed the mammalian pathogenic constellation of polymerases. Thus, the novel polymerase genes and amino acid residues may be useful to understand the host-barrier overcome of AIVs in mice and to develop safer and efficacious vaccines.