Comparative analysis of popular predictors for difficult laryngoscopy using hybrid intelligent detection methods

Area of Science:

• Anesthesiology
• Artificial Intelligence in Medicine
• Medical Informatics

Background:

• Difficult laryngoscopy poses risks like airway injury and asphyxia, lacking standardized prediction guidelines.
• Existing predictors for difficult laryngoscopy lack comprehensive comparative analysis.
• The study addresses the need to evaluate and compare machine learning (ML) and deep learning (DL) models for predicting difficult laryngoscopy.

Purpose of the Study:

• To compare the efficacy and accuracy of ML-based and DL-based models in predicting difficult laryngoscopy.
• To identify the best performing model for difficult laryngoscopy prediction.
• To evaluate the utility of radiological variables in predicting difficult laryngoscopy.

Main Methods:

• A dataset of 671 patients was used to evaluate seven ML-based models and four DL-based approaches.
• Performance was assessed using single and integrated indicators, including radiological variables.
• Adaptive spatial interaction was applied to enhance prediction using preoperative cervical spine X-rays.

Main Results:

• The Naïve Bayes (ML) model achieved the highest accuracy (86.6%), F1 score (0.908), and average precision (0.837), outperforming DL models.
• Three radiological variables were valuable individually and combinedly for prediction.
• No significant performance difference was observed among individual radiological indicators or their combination.

Conclusions:

• Simple traditional machine learning models, like Naïve Bayes, are effective for predicting difficult laryngoscopy.
• Radiological indicators can be flexibly chosen by anesthesiologists for predicting difficult laryngoscopy.
• The study provides valuable insights for improving the prediction of difficult laryngoscopy using AI.