Fig 1.
Schematic Overview of the Study.
(A) Flowchart illustrating patient inclusion/exclusion criteria and interim analysis procedures. After enrollment of the initial 300 patients, an interim analysis was performed to evaluate preliminary model performance. Due to insufficient generalizability observed at this stage, patient enrollment was continued according to the original study plan to expand the sample size. (B) Representative ultrasound scanning planes and corresponding images. This figure presents four ultrasound scanning planes (left, demonstrated by the operator) alongside representative images (right). (a) Midsagittal view of the mandible: A low-frequency convex probe (Mindray TE7, C5-2s) was placed with the marker aligned with the chin and the opposite end over the superior border of the thyroid cartilage. (b) Transverse view of the hyoid bone: A high-frequency linear probe (Mindray TE7, L9-3s) was positioned laterally at a 45° angle to the horizontal plane. (c) Transverse view of the thyroid cartilage: The same probe was placed perpendicularly over the anterior neck to visualize the thyroid cartilage. (d) Paramedian sagittal view of the larynx: The ultrasound probe was placed 1 cm lateral to the neck midline, with its marked end 1 cm superior to the hyoid upper border, its plane parallel to the midline and angled at 45° to the horizontal plane. Key anatomical landmarks related to airway assessment, such as the hyoid bone and epiglottis, are shown in red. (C) CL Classification. (D) VIDIAC scoring system used for video laryngoscopic classification. (Created by the authors, based on concepts described in Kohse EK, et al. Anaesthesia. 2022. This image is similar to but not identical with the original one, and is therefore intended solely for illustrative purposes.) The VIDIAC system comprises three components: (1) Epiglottis – contact between the blade tip and epiglottis; (2) Vocal cords – best achievable view of the glottis; (3) Arytenoids – presence of hypertrophy. Scores range from –1 to 5. In this study, VIDIAC ≥3 was defined as a difficult airway. CL, Cormack–Lehane grade; VIDIAC, Videolaryngoscopic Intubation and Difficult Airway Classification.
Fig 2.
Schematic flowchart of study methodology.
The diagram summarizes the overall study process: Top layer: perioperative data and image acquisition; Middle layer: training and validation of deep learning models; Bottom layer: DenseNet-BC architecture used for image-based prediction. GPU = Graphics Processing Unit.
Table 1.
Patients’background characteristics.
Table 2.
Detailed parameters of the artificial intelligence prediction model in the internal test set.
Fig 3.
ROC curves of different AI models.
(A) ROC curve comparison under the Cormack-Lehane (CL) classification framework, showing the performance of four ultrasound planes and the integrated model.(B) ROC curve comparison under the VIDIAC framework, showing the performance of four ultrasound planes and the integrated model. CL, Cormack–Lehane classification; VIDIAC, Videolaryngoscopic Intubation and Difficult Airway Classification; AI, artificial intelligence; DSH, Distance Skin to Hyoid; DSE, Distance Skin to Epiglottis; TMD, Thyromental Distance; others listed in figure.
Table 3.
Performance comparison between anesthesiologists and AI models using direct and video laryngoscopy.
Fig 4.
Schematic diagram of the artificial intelligence model workflow for difficult laryngoscopy prediction.
The workflow illustrates the process of two AI models (CL-Model and VIDIAC-Model) analyzing four ultrasound planes to predict difficult laryngoscopy and stratify risk. Ultrasound images from four planes are input into the CL-Model, which generates attention heatmaps for regions like the epiglottis, suprahyoid soft tissue, and thyroid cartilage. A multi-model ensemble integrates these outputs to classify cases as CL-Easy (low risk) or CL-Difficult. For cases classified as CL-grade difficult airway, the VIDIAC model was further employed to analyze the ultrasound planes, focusing on regions such as the submandibular region, hyoid bone, epiglottis, and thyroid cartilage. Its multi-model ensemble classifies cases into VIDIAC-Easy (medium risk) or VIDIAC-Difficult (high risk), completing the risk stratification. CL, Cormack–Lehane classification; VIDIAC, Videolaryngoscopic Intubation and Difficult Airway Classification; AI, artificial intelligence.
Fig 5.
Comparison of airway assessment workflows.
(A) Traditional workflow: Patient undergoes airway assessment by Attending and Chief Anesthesiologists, with difficult cases undergoing special preparation and easy cases receiving routine anesthesia. (B) AI-assisted workflow: Patient undergoes ultrasound assessment via CL-AI and VIDIAC-AI models, stratified into low-risk (direct laryngoscopy), medium-risk (video laryngoscopy), and high-risk (video laryngoscopy + emergency backup) groups. CL, Cormack–Lehane; VIDIAC, Videolaryngoscopic Intubation and Difficult Airway Classification; AI, artificial intelligence.