Airway coach project: development of a machine learning-based model using clinical and ultrasound parameters to...

Videolaryngoscopy (VL) is recommended as a first-line technique for tracheal intubation; however, existing airway assessment tools-largely derived from direct laryngoscopy-provide limited guidance for videolaryngoscopy-specific decisions such as blade selection or anticipated adjunct use. Point-of-care airway ultrasound offers additional anatomical information that may complement conventional clinical assessment. Multimodal approaches integrating clinical and ultrasound-derived variables may improve pre-procedural videolaryngoscopy planning.

In this single-centre prospective observational study, 250 adults (ASA physical status I-III) undergoing elective surgery were assessed preoperatively using clinical variables and point-of-care airway ultrasound. Videolaryngoscopic intubation was initiated with a Macintosh-type blade and prospectively classified according to procedural performance: Grade 0 (Macintosh blade without adjuncts), Grade 1 (Macintosh blade requiring adjuncts), and Grade 2 (need to switch to a hyperangulated device). A supervised machine-learning approach was used to develop a multiclass classification model integrating clinical and ultrasound-derived variables, using stratified cross-validation on a training dataset and evaluation on an independent test set.

In the independent test set, a gradient boosting model demonstrated good discriminative performance for videolaryngoscopy strategy classification (AUC 0.95; accuracy 92%). Performance varied across outcome categories, with lower F1-scores in the less frequent classes (F1-score 0.75 for Grade 1 and 0.67 for Grade 2). Classification was driven by a combination of tongue-related ultrasound parameters, anterior neck soft-tissue distances, body-mass index, and age.

In this single-centre observational study, a machine learning-based model integrating clinical and airway ultrasound variables was developed to support videolaryngoscopy strategy planning. By targeting a functional, videolaryngoscopy-specific outcome that reflects procedural performance rather than glottic view alone, this multimodal, data-driven approach demonstrates feasibility and warrants further evaluation in independent populations before clinical implementation.

This study was registered at Clinical Trial.gov NCT06925009.