Machine learning-based prediction of difficult laryngoscopy in infants with Pierre Robin sequence using quantitative 3D computed tomography parameters.

Authors

Abstract

Infants with Pierre Robin sequence (PRS) frequently present with difficult laryngoscopic exposure due to mandibular hypoplasia, glossoptosis, and upper airway narrowing. Quantitative three-dimensional computed tomography (3D-CT) enables objective characterization of airway anatomy and may improve preoperative airway assessment. This study aimed to identify key 3D-CT parameters associated with difficult laryngoscopic exposure in infants with PRS and to develop and externally validate machine learning-based predictive models. We retrospectively analyzed 214 infants with PRS who underwent mandibular distraction osteogenesis between 2023 and 2024. According to laryngoscopic view, patients were classified into an easy-exposure group (Cormack-Lehane grades I-II) and a difficult-exposure group (grades III-IV). Univariable and multivariable logistic regression analyses were performed to identify independent quantitative 3D-CT predictors. Seven machine learning models were constructed, including logistic regression, support vector machine, random forest, Extra Trees, XGBoost, LightGBM, and AdaBoost. Internal performance was evaluated using five-fold cross-validation, and generalizability was assessed in an independent temporal validation cohort from the same institution (<i>n</i> = 80). Model performance was assessed using discrimination, calibration, and decision curve analysis. Multivariable analysis identified four independent predictors: tongue length (D1) (odds ratio [OR] = 1.058, <i>p</i> = 0.005), tongue base-posterior pharyngeal wall distance (D4) (OR = 0.718, <i>p</i> < 0.001), sagittal oropharyngeal cross-sectional area (S2) (OR = 0.271, <i>p</i> = 0.001), and tongue base-epiglottic angle (A2) (OR = 0.952, <i>p</i> = 0.028). Among all models, XGBoost achieved the highest discrimination in the training cohort (AUC = 0.961). However, in the temporal validation cohort from the same institution, the Extra Trees model demonstrated superior generalizability (AUC = 0.876), with an accuracy of 0.812 and an F1-score of 0.805. Calibration analysis indicated excellent agreement between predicted and observed outcomes for the Extra Trees model (Hosmer-Lemeshow test, <i>p</i> > 0.999). Decision curve analysis showed a substantial net clinical benefit across threshold probabilities ranging from approximately 10 to 90%. Quantitative 3D-CT parameters reflecting tongue morphology and oropharyngeal airway dimensions are clinically relevant predictors of difficult laryngoscopic exposure in infants with PRS. The Extra Trees model showed promising performance in temporal validation within a single-center cohort.

Tags

Topics