Abstract Introduction Anatomic characterization of the upper airway remains important in directing and monitoring care of patients with obstructive sleep apnea (OSA). Nasopharyngoscopy is routine in clinical practice, but it is invasive, non-reproducible, and only allows subjective assessment. We used machine-learning enabled ultrasonography to correlate upper airway tissue characteristics with OSA severity. Methods Sixty-three subjects (14 female) with a mean age of 39.4±12.6 years, BMI of 26.4±4.6 kg/m2, and AHI of 19.0±16.1 were consented from Stanford Sleep Surgery (July 2020 to May 2021). Standardized ultrasound protocol was used to image the soft palate, oropharynx, and tongue-base. Via machine learning, an FDA-cleared backscattered ultrasound imaging (BUI) of the upper airway was performed. Combined with B-mode measurements of airway muscular cross-sections, a logistic regression model was built to correlate with OSA severity. Results BUI of subjects with mild OSA was different from moderate-severe (AHI≥15) OSA at the soft palate (p=0.0007). The axial-to-lateral ratio of upper airway length was reduced in the lower soft palate of the moderate-severe group (p =0.0207). The logistic regression model with BUI, axial-to-lateral ratio at the soft palate, and BMI showed an Area Under the Receiver Operating Characteristic (AUROC) curve of 0.84 (95% CI 0.726 to 0.920) in moderate-severe OSA. Conclusion A non-invasive yet replicable technique to visualize and phenotype the upper airway is critical in the management of patients with sleep-disordered breathing. Sonographic BUI combined with B-mode airway measurements analyzed by machine learning show promise in characterizing the upper airway in patients with moderate-severe OSA. Support (If Any)
Read full abstract