Deep learning-based automatic scoring of drug-induced sleep endoscopy in obstructive sleep apnea

Polysomnography is the standard tool for assessing obstructive sleep apnea (OSA) severity; however, it does not provide information regarding the anatomical site or extent of upper airway obstruction. Drug-induced sleep endoscopy (DISE) serves as a dynamic method to evaluate airway collapse under sleep-like conditions, thereby helping to bridge this gap. However, its clinical utility is limited by inter-observer variability and subjectivity in interpretation. We developed internally and externally validated deep learning models utilizing convolutional neural networks based on EfficientNet-B2 and Attention Multiple Instance Learning to predict the degree of airway obstruction (DISE-V-obs, DISE-OTE-obs) and the primary cause of obstruction (DISE-OTE-cause) using DISE videos from 1904 patients across five Korean hospitals. The F1 scores for DISE-V-obs, DISE-OTE-obs, and DISE-OTE-cause were 84.7%, 74.7%, and 88.2%, respectively. These objective predictions of obstruction degree and primary cause may enhance clinical decision-making and treatment planning for patients with OSA.