A Multimodal AI-Driven System for Preoperative Difficult Airway Assessment

Preoperative assessment of the difficult airway is essential to ensure anesthetic safety. However, traditional assessment methods rely on subjective clinical evaluations and demonstrate limited predictive accuracy. To address this challenge, we propose a multimodal AI-driven airway assessment system. The hardware architecture consists of an integrated camera module, computation module, and display module, while the software is built on a three-stage intelligent framework that comprises image feature extraction, feature-based parameter quantification, and data fusion-based risk prediction. In the feature extraction stage, the framework employs deep learning (DL) models to process input images. Subsequently, the framework automatically measures five critical airway parameters. In the prediction stage, these parameters are fused with personal basic information. Finally, an eXtreme gradient boosting (XGBoost) model processes the fused data to generate a final risk prediction for airway assessment. We evaluated the system performance using a dataset including 1,207 patients undergoing general anesthesia. On an independent test set, the DL models for feature extraction demonstrate high accuracy, with the AI-driven system achieving an area under the curve (AUC) of 0.831. The proposed system offers an objective and accurate solution for preoperative airway assessment, outperforming traditional methods.