Artificial Intelligence for Prediction of Difficult Airway in Anesthesia: A Systematic Review and Descriptive Synthesis

Introduction: Difficult airway remains an important cause of anesthesia-related morbidity because failure to identify high-risk airway before induction may lead to repeated laryngoscopy attempts, hypoxemia, aspiration, emergency front-of-neck access, cardiac arrest, or death. This review evaluated artificial intelligence (AI) models for difficult-airway prediction in anesthesia and airway management. Methods: A systematic literature review with descriptive synthesis was performed using a PICOTS framework and PRISMA-based reporting. PubMed/MEDLINE, Embase, and Scopus were used for the primary evidence map, with supplementary updating through PubMed/MEDLINE, Google Scholar, and reference checking to 13 October 2026. Eligible studies were original human clinical studies using AI, machine learning, deep learning, or computer vision to predict difficult intubation, difficult laryngoscopy, difficult video laryngoscopy, difficult mask ventilation, or composite difficult airway. Studies predicting respiratory-failure intubation or tube size/depth without difficult-airway assessment were excluded. Results: Nineteen eligible difficult-airway-related studies were included in the final descriptive synthesis. Sixteen studies reported extractable AUC or c-statistic values, while three studies were summarized narratively only. One estimate per study was summarized descriptively, with median discrimination of 0.847 (IQR 0.773-0.865; range 0.710-0.970). Calibration was reported in five studies and external validation in one study. Conclusion: AI models show promising discrimination for difficult-airway prediction, but evidence remains limited by heterogeneous outcome definitions, incomplete calibration reporting, sparse external validation, and uncertain clinical impact. AI should currently be considered an adjunct to anesthesiologist judgment rather than a stand-alone diagnostic method.