Acoustic Analysis of Snoring Sounds as a Diagnostic Tool for Pediatric Obstructive Sleep Apnea

Introduction: Tonsillectomy is the first-line intervention for pediatric obstructive sleep apnea (OSA), but it is not always clear if the tonsils are the primary cause of obstruction. We hypothesized that snoring sounds may predict the presence of obstruction from adenoid or tonsillar hypertrophy.

Methods: In this prospective cohort study, snoring sounds were recorded in children using a digital stethoscope during drug-induced sleep endoscopy (DISE). The cohort was stratified according to obstruction pattern: tonsils, adenoids, both or neither. Audio files were standardized to remove excess noise, and 6,373 acoustic features were extracted using the OpenSMILE package in Python. A multivariate logistic regression model was developed using the top 100 acoustic features ranked by a mixed redundancy maximum relevance (mRMR) algorithm to predict the obstruction pattern observed during DISE. Patients were split into balanced 85% training and 15% testing groups and accuracy was measured.

Results: In our cohort (n = 63), most were male (63%) and white (83%) with a mean age of 8.3 years. Comorbidities included Down syndrome (25%) and cerebral palsy (8%). Most patients were surgically naïve (71%) while some had prior tonsil or adenoid surgery (17%). Obstruction from both adenoids and tonsils was most common (41%), followed by tonsils alone (31%), neither (19%), or adenoids alone (10%). The model achieved 90% accuracy in predicting obstructive locations.

Conclusions: Snoring sounds recorded with a digital stethoscope are a feasible source for identifying patterns of obstruction in children. These results should be interpreted with caution given the limited sample size. Data collection is on-going and further model refinement will include other sources of obstruction beyond the tonsils and adenoids.