Obstructive Sleep Apnea syndrome (OSA) is a widespread serious disease in which upper airways (UA) are collapsed during sleep. In the USA, an estimated 9% of women and 24% of men were found to have at least mild OSA. Loud/disruptive snoring, gasping/choking are the classical nocturnal symptoms of OSA.
Polysomnography (PSG) is the current reference standard for OSA diagnosis, requires overnight hospital stay while physically connected to more than 15 channels of measurements. The recorded physiological signals are manually analysed to derive an OSA severity measure known as Apnea-hypopnea Index (AHI). PSG is expensive, inconvenient and requires the extensive involvement of a sleep technologist. As such, it is not suitable for community screening. Furthermore, at present, more than 80% of OSA suffers remain undiagnosed.
Recently, several alternatives to in-facility PSG targeting OSA diagnosis have been suggested to resolve this need. All these methods had at least one sensor requiring physical contact with the patient. Furthermore, the involvement of a sleep technician still needed to guarantee an acceptable level of diagnostic performance in home monitoring or community screening. Low cost, simple, unattended, convenient OSA screening technique is urgently required.
With the onset of sleep, the upper airway (UA) encounters a number of physiological changes. The UA constrictions and impaired muscle tone evoke soft tissues to vibrate causing snoring.
OSA has marked male predominance and also reported that gender increases the risk of OSA by a factor of 2-3. It is also well known that structural and functional properties of the UA are gender dependent. OSA, by definition, is closely associated with the UA patency in sleep. There exist gender-specific differences in vocal tract dimensions, elastic properties of tissues and prephonatory glottal shape. Considering the snore sound (SS) generation mechanism, it is highly likely that SS may capture these time, gender attributed UA properties and those could be embedded in the acoustic properties of SS.
The goal of this thesis is to develop a novel multi snore-feature class OSA screening tool by integrating snore features that capture functional, structural, spatio-temporal dependencies of SS. We also work under the hypothesis that gender specific analysis of snore sounds should lead to a higher OSA detection performance.
A clinical database was developed that contains the overnight snoring related sounds (SRS) simultaneously recorded during a hospital PSG. SRS were segmented and voiced snoring segments (VSS) were identified using a pattern recognition based algorithm. We also adopted a simple time domain based criterion to remove specific class of unwanted background sounds from analysis.
Number of snore feature time series (pitch, recurrence, formant, non - gaussianity, mel cepstrum coefficients and higher order statistics feature classes) were developed by computing features from VSS. Set of feature vectors containing statistical features of snore time series were derived for each feature class. Logistic regression (LR) modelling techniques were used to optimize the feature vectors for OSA diagnostic performance in individual as well as in integrated feature classes separately. LR models were developed and computed the OSA diagnostic sensitivity/specificity using mutually exclusive testing data sets. We also augmented neck circumference to snore features which is readily available at no extra costs.
The performance of the method was compared against the reference standard, PSG at two AHI decision thresholds (15 and 30) and cross validated. Our proposed techniques resulted in a mean sensitivity (specificity) of 93% (92%) for females and a mean sensitivity (specificity) of 92% (93% ) for males at AHI decision threshold = 15. At AHI decision threshold = 30, our method resulted a mean sensitivity (specificity) of 94% (93%) and 92% (93%) for females and males respectively.
Our results indicate that gender based analysis of snore sound improves OSA detection. It also illustrates that this method has the potential to develop an automatic, non-contact, unattended population screening OSA tool.