AI-based decision models for difficult airway assessment: from research innovation to clinical implementation—a narrative review

Difficult airway management causes significant anesthesia-related morbidity, yet traditional assessments lack sensitivity (30%−50%) and consistency. This review (2010–2025) examines artificial intelligence (AI) decision models for airway assessment, focusing on performance, limitations, and clinical translation. AI demonstrates significant statistical superiority: facial image analysis achieves 80%−90% sensitivity (vs. Mallampati's 39%), and deep learning models yield a pooled AUC of 0.84. Key techniques include convolutional neural networks, semi-supervised learning, and multimodal integration. Despite high predictive performance, widespread adoption faces fundamental barriers. Current studies are predominantly single-center and retrospective, lacking external validation, algorithmic fairness, standardized outcomes, and proven workflow integration. Furthermore, research heavily favors upper airway evaluation. Thoracic anesthesia, utilizing routine preoperative CTs, offers an immediate pathway for comprehensive whole-airway assessment. Ultimately, bridging the translational gap requires rigorous, prospective multicenter validation demonstrating tangible patient safety improvements, rather than relying solely on algorithmic sophistication.