An ultrasound-based artificial intelligence framework for difficult airway prediction: A two-model, three-step decision framework

BACKGROUND: At present, the early warning of difficult airway remains fraught with challenges. Previous ultrasonic quantitative parameters have demonstrated favorable application potential in difficult airway assessment, and deep learning techniques have also exhibited satisfactory performance in the interpretation of this condition. Based on this, we aim to construct a "two-model, three-step" hierarchical strategy, develop an ultrasound image-based artificial intelligence (AI) framework for difficult airway prediction, and conduct its internal validation. METHODS: In this study, we included 903 patients who underwent elective general anesthesia surgery at the Affiliated Hospital of Qingdao University between May 2024 and April 2025. 752 cases were used for model training and validation, and 151 cases served as an internal test set. Four planes of neck ultrasound images were scanned for each patient and used to develop two artificial intelligence models (based on convolutional neural networks): CL-AI for initial screening and VIDIAC-AI for secondary risk stratification. Model performance was evaluated using five-fold cross-validation and internal testing. External validation was not performed. RESULTS: Among 903 patients, difficult laryngoscopy occurred in 189 cases (20.9%) under direct laryngoscopy and in 50 cases (5.5%) under video laryngoscopy. In the independent test set, the CL-AI model achieved an AUC of 0.86 (95% CI: 0.79-0.91), with an accuracy of 0.84, sensitivity of 0.84, specificity of 0.84, precision of 0.59, and an F1 score of 0.69. The VIDIAC-AI model achieved an AUC of 0.82 (95% CI: 0.75-0.88), with an accuracy of 0.81, sensitivity of 0.75, specificity of 0.81, precision of 0.18, and an F1 score of 0.29. CONCLUSIONS: This study proposes an ultrasound-based AI framework for risk stratification of difficult laryngoscopic exposure. The Two-Model, Three-Step decision framework is intended as a clinician decision-support tool, not an independent diagnostic method, and requires further validation in large multicenter cohorts.