Machine learning-based model for prediction of outcomes in palatal surgery for obstructive sleep apnoea.

Abstract

● Our retrospective cohort study aimed to elucidate which factors (nasal surgery: septoplasty or functional endoscopic sinus surgery (FESS), palate surgery: uvulopalatal flap (UPF) or palatal muscle resection (PMR)), demography, polysomnography, Friedman stage, drug-induced sleep endoscopy (DISE) results) influence the success of surgery in patients who had been operated to treat obstructive sleep apnea (OSA). ● We used seven machine learning (ML) models (logistic regression-based models (lasso, elastic net (EN), ridge), tree-based models (bagging, random forest (RF), support vector machine (SVM), and neural networks (NN)). ● Among the seven models, their performances were first ranked by the area under the receiver operating characteristic curve (AUC): lasso (0.8317), EN (0.7500), SVM (0.7163), bagging (0.7115), random forest (0.7115), NN (0.6923) and ridge (0.6538), which shows that lasso had the best performance. ● Tonsil grade and Nasopharynx Oropharynx Hypopharynx and Larynx (NOHL) classification of the epiglottis in DISE were selected as the most important variables in all models including lasso. ● Through the novel ML technique, the surgical success rate of OSA is ultimately dependent on tonsil grade and NOHL classification of the epiglottis in DISE regardless of the type of nasal or palate surgery.