Deep Learning-Based Automated Versus Manual Snoring Scoring in Sleep Disordered Breathing

Snoring is a major acoustic manifestation of upper airway obstruction and an indicator of sleep-disordered breathing (SDB). However, objective snoring quantification during polysomnography (PSG) remains challenging because of inconsistent scoring criteria. We evaluated the performance of a deep learning-based automated PSG analysis system for snoring scoring by comparing automated and manual analyses.

This retrospective study included 100 adults who underwent Level 1 PSG for suspected obstructive sleep apnea between January 2020 and December 2024. Manual PSG scoring followed American Academy of Sleep Medicine criteria. Automated analysis was performed using the SOMNUM clinical decision support system. Agreement was evaluated using intraclass correlation coefficients (ICC), Bland-Altman analysis, and Pearson correlation analysis. Respiratory parameters and OSA severity classification were also assessed.

Automated snoring analysis showed good agreement with manual scoring for snoring time (ICC = 0.91; 95% CI, 0.84-0.94) and snoring percentage (ICC = 0.84; 95% CI, 0.76-0.89). Bland-Altman analysis demonstrated acceptable agreement, and Pearson correlation confirmed strong associations for snoring time (r = 0.86) and snoring percentage (r = 0.88) (both p < 0.001). Respiratory parameters also showed excellent agreement, including the apnea-hypopnea index (ICC = 0.97) and oxygen desaturation index (ICC = 0.97). No severe OSA cases were misclassified into lower-risk categories.

Automated PSG analysis demonstrated strong agreement with expert manual scoring for snoring, respiratory parameters, and OSA severity. These findings support its clinical utility for objective, reproducible snoring assessment and its potential as a digital biomarker for SDB.