Assessment of obstructive sleep apnea severity using audio-based snoring features
Jiali Xie, Pedro Fonseca, Johannes van Dijk, Sebastiaan Overeem, Xi Long
Abstract
Snoring is a prima symptom of obstructive sleep apnea (OSA). Here, we add audio-based snoring features to improve the non-obtrusive assessment of sleep apnea, by estimating the apnea-hypopnea index (AHI) and classifying OSA severity. We propose novel features to quantify temporal changes between snores (snore rate variability) and to describe trends in snore energy, based on the assessment of snore sounds from audio signals over the full night. We then combined those features with age, body mass index (BMI) and features described in literature. An extreme gradient boosting algorithm was trained with all these features on AHI estimation. The estimated AHI was then used to classify OSA severity. Audio-based estimated AHI showed a significant Spearman's correlation with the AHI based on gold-standard polysomnography (R = 0.786, P < 0.0001). Our results outperformed a model trained with solely previously described features in our dataset (R = 0.676, P < 0.0001) and a model trained with the combination of previously described features, age, and BMI (R = 0.731, P < 0.0001). The mean absolute error of AHI estimation was 7.26 events/h. Area under the receiver operating characteristic curve outcomes were 0.90, 0.87 and 0.93 for classifying patients with varying severity separated by the canonical thresholds of 5, 15 and 30 events/h respectively. The accuracy of classifying subjects to four classes (no, mild, moderate, and severe OSA) was 59.3 %. Additional audio-based snore features can improve the performance of non-obtrusive AHI estimation and OSA severity classification methods.