Hidden Markov modelling of intra-snore episode behavior of acoustic characteristics of obstructive sleep apnea patients

Obstructive sleep apnea (OSA) is a breathing disorder that can cause serious medical consequences. It is caused by full (apnea) or partial (hypopnea) obstructions of the upper airway during sleep. The gold standard for diagnosis of OSA is the polysomnography (PSG). The main measure for OSA diagnosis is the apnea-hypopnea index (AHI). However, the AHI is a time averaged summary measure of vast amounts of information gathered in an overnight PSG study. It cannot capture the dynamic characteristics associated with apnea/hypopnea events and their overnight distribution. The dynamic characteristics of apnea/hypopnea events are affected by the structural and functional characteristics of the upper airway. The upper airway characteristics also affect the upper airway collapsibility. These effects are manifested in snoring sounds generated from the vibrations of upper airway structures which are then modified by the upper airway geometric and physical characteristics. Hence, it is highly likely that the acoustical behavior of snoring is affected by the upper airway structural and functional characteristics. In the current work, we propose a novel method to model the intra-snore episode behavior of the acoustic characteristics of snoring sounds which can indirectly describe the instantaneous and temporal dynamics of the upper airway. We model the intra-snore episode acoustical behavior by using hidden Markov models (HMMs) with Mel frequency cepstral coefficients. Assuming significant differences in the anatomical and physiological upper airway configurations between low-AHI and high-AHI subjects, we defined different snorer groups with respect to AHI thresholds 15 and 30 and also developed HMM-based classifiers to classify snore episodes into those groups. We also define a measure called instantaneous apneaness score (IAS) in terms of the log-likelihoods produced by respective HMMs. IAS indicates the degree of class membership of each episode to one of the predefined groups as well as the instantaneous OSA severity. We then assigned each patient to an overall AHI band based on the majority vote of each episode of snoring. The proposed method has a diagnostic sensitivity and specificity between 87-91%.