Proceedings of Meetings on Acoustics

Transcription

1 Proceedings of Meetings on Acoustics Volume 19, ICA 2013 Montreal Montreal, Canada 2-7 June 2013 Speech Communication Session 4aSCb: Voice and F0 Across Tasks (Poster Session) 4aSCb19. Detection of obstructive sleep apnea by estimation of oral and nasal cavity cross-section areas from acoustic recordings of snore Hsu-Kang Huang, Yi-Wen Liu* and Rayleigh Ping-Ying Chiang *Corresponding author's address: Electrical Engineering, National Tsing Hua University, 101 KuangFu Rd. Sec. 2, Hsinchu, 30013, Taiwan, Taiwan, Obstructive sleep apnea (OSA) refers to the condition in which a person's breathing is paused while asleep, or the airflow is decreased, due to obstruction in the upper respiratory airway. In severe cases, OSA can cause complete arousal and deprive the patient from normal sleep. Surgical intervention is sometimes recommended, but accurate identification of the site of obstruction can be difficult. In the present study, we devised signal processing methods to estimate the site and the severity of airflow obstruction from recordings of sounds of snore. The vocal tract, the oral and the nasal cavity are modeled as three branches joining at the pharynx. Each branch consists of cylindrical segments whose cross-section areas can vary during snoring. Estimation of these cross-section areas consists of two steps: First, an auto-regressive movingaverage method is applied to find the linear coefficients of a pole-zero model that optimally accounts for the recorded sound. Then, the Levinson-Durbin algorithm is applied to convert the coefficients to ratios of cross-section areas between adjacent segments. The present method is applied to a set of recorded snore samples during clinically confirmed apnea episodes, and results are compared with those of simple snore. Effectiveness of the method is analyzed statistically. Published by the Acoustical Society of America through the American Institute of Physics 2013 Acoustical Society of America [DOI: / ] Received 20 Jan 2013; published 2 Jun 2013 Proceedings of Meetings on Acoustics, Vol. 19, (2013) Page 1

2 INTRODUCTION The present work aims to investigate whether acoustic features computed from the sounds of snoring can be effective for detecting OSA episodes and, if so, whether the site of obstruction can be estimated acoustically. Present standard protocols utilize polysomnography (PSG) to detect OSA episodes during sleep. A typical recording of PSG is shown in Fig. 1; many physiological signals are monitored simultaneously, including the electroencephalogram (EEG), electrocardiogram (ECG), thorax-abdomen movement (denoted as Thor and Abdo in Fig. 1) during breathing cycles, the oxygen saturation level in hemoglobin (SpO2), and so on. Although PSG provides comprehensive information for clinicians to examine the overall health of a patient during sleep, the act of multichannel recording typically requires placing various sensors and all the connecting wires on and around the patient s body. This arrangement inevitably makes it difficult for anybody to fall asleep. The hassle of setting up the PSG also hinders its deployment for home-based day-to-day sleep health monitoring. The hassle would be significantly reduced if there is a way to monitor OSA episodes based on acoustic recording of the sounds of breathing. Past studies suggested that the sounds of snoring can be classified according to the obstruction site along the respiratory airway; for instance, different sites of obstruction correspond to different resonance frequencies in the oral-nasal cavity (Quinn et al., 1995), and the frequency typically ranged from 0.1 to a 1 khz (Miyazaki et al., 1998). However, in these reports the site of upper airway obstruction was confirmed by sleep nasendoscopy (Croft and Pringle, 1991), a heavily invasive technique which may have caused the patients upperairway muscle relaxation status to deviate from its usual conditions during sleep. Further, the dosage of tranquilizer applied during sleep nasendoscopy also affects the sound frequency of snoring (Agrawal et al., 2002). To summarize, the invasive technique itself and the necessity of tranquilization both affect how snoring is produced. The validity and accuracy of these reports are thus questionable regarding the mapping between sound frequencies and site of obstruction. (a) (b) (c) Figure 1. Examples of polysomnography (PSG) recordings. (a) A typical page showing all the recording channels and the waveforms. (b) Two episodes of simple snore were detected. (c) An episode of OSA was detected. Note the long-lasting period with reduced breathing-related motion in the thorax and the abdomen and reduced flow from the nostril. Proceedings of Meetings on Acoustics, Vol. 19, (2013) Page 2

3 Figure 2. Schnell and Lacroix s (2001) method for ARMA coefficients estimation. A Burg s lattice-based method was applied to the inverse discrete-fourier transform (IDFT) of ()/() and ()/(), respectively, in hope of that the conditionally optimal other set of parameters would reduce the mean square error step by step. (Here, the variable z is defined as =.) Hence, we seek to develop a computational framework which allows us to estimate the site and the severity of obstruction merely from acoustic recordings of snoring. To conduct this research, sounds of snoring have been collected from a lab member. The signals were matched by a time-varying auto-regressive moving-average (ARMA) model, and coefficients in the ARMA model were converted to a set of wave reflection coefficients which yielded an estimate of the shape of the vocal tract and the nasal cavity. Details of the methods are described below. METHODS Auto-Regressive Moving-Average (ARMA) Modeling of the Sounds of Snoring Assume that for a short period of time (e.g., 20 ms) the sounds of snoring [] can be regarded as a quasi-stationary signal, where n denotes the discrete time index. i Briefly speaking, ARMA modeling looks for a linear time-invariant (LTI) system that best explains how [] is generated. A discrete-time LTI system can be characterized by a set of Proceedings of Meetings on Acoustics, Vol. 19, (2013) Page 3

4 auto-regressive (AR) coefficients s and moving-average (MA) coefficients s; therefore comes the name ARMA modeling. To calculate the ARMA coefficients, the signal [] is regarded approximately as the result of filtering an excitation signal [] by an LTI system: [] [] = [ ] + [ ], (1) where the excitation signal e[n] is defined as the approximation error: [] =[] []. (2) Then, the goal of ARMA modeling is to search for s and s that jointly minimize the expectation of ([]). In the frequency domain, the mean square of = ([]) can be written as follows, ε= ( ) = ( ) (3) where uppercased variables denote the discrete-time Fourier transform of the corresponding lower-cased signals. Schnell and Lacroix (2001) showed that the mean-square error ε in Eq. (3) can be minimized by iteration. The minimization procedure is illustrated in Fig. 2; with some initial guess, this algorithm takes turns to find conditionally optimal s (AR coefficients) or s (MA coefficients) using a Burg s lattice-based method (Gibson and Haykin, 1980). The procedure continues until ε cannot be further reduced by conditional optimization. Estimation of Cross-Section Areas Along the Vocal Tract and in the Nasal Cavity Figure 3 illustrates how the upper respiratory airway can be modeled as three branches of tubes in which acoustic waves can propagate in both the forward (from glottis to lips) and the backward (if reflected) directions. The crosssection area (CSA) may vary along each tube so as to resemble the vocal tract and the nasal cavity in reality. From a sound-synthesis perspective, once the CSA along the vocal tract and the nasal cavity is known, their joint acoustic filtering effect can be modeled as an LTI system; the system is characterized by AR and MA coefficients that can be computed by a scattering matrix formulation (illustrated in Fig. 4). What is more difficult is the inverse problem given a set of the AR and MA coefficients, can the CSA along the vocal tract and the nasal cavity be determined? It turns out that the estimation of CSAs from the ARMA coefficients involves solving a set of nonlinear algebraic equations, and numerical solutions may not be physically realistic [for example, some of the CSAs might end up being negative (Huang, 2012)]. Here, we chose to estimate, separately, the CSA along the nasal cavity and that of the two other branches. First, the CSAs along the nasal cavity (in which the wave variables are denoted as ± () in Fig. 4) were estimated from the MA coefficients s while ignoring all the AR coefficients. We argue this is a reasonable approximation, since without the nasal branch the transfer function of the LTI would be an allpole function; in other words, s would all vanish. Therefore, it should be possible to estimate the nasal cavity CSAs from the MA coefficients s. Presently, this estimation was performed via Levinson-Durbin recursion (Levinson, 1947) which coverts s to a set of reflection coefficients,, shown in Fig. 4. Then, a similar procedure was applied to the AR coefficients s to obtain estimates of reflection coefficients,, for the main branch and,, for the oral branch. Finally, in each branch the CSA ratios between adjacent cylindrical segments are related to the reflection coefficients by the following equation: =, (4) where and are the CSAs of two adjacent segments. Therefore, the CSA of one segment can be calculated if the CSA of its immediate neighbor is known and the reflection coefficient between them is given; that is, =. (5) Proceedings of Meetings on Acoustics, Vol. 19, (2013) Page 4

5 Figure 3. A 3-branch, multitube model for acoustic wave scattering in the vocal tract and the oral-nasal cavity. The three branches are modeled as concatenated cylindrical segments of equal length l, and waves are reflected due to impedance (inversely proportional to the cross-section area) mismatch between adjacent segments. Figure 4. A signal flowchart for wave scattering in the 3-branch model (shown in Fig. 3). The main vocal tract, the oral tract, and the nasal tract, are joined at the pharynx depicted as the 3-port branch boundary in the middle. Variables s and s denote reflection coefficients, and signals ±, ±, ± s denote the forward and backward propagating waves in the main branch, the oral branch, and the nasal branch, respectively. MEASUREMENT AND RESULTS Simultaneous Recording the Sounds of Snoring and the PSG The afore-mentioned methods were applied to recorded sounds of snoring so as to estimate the CSA along the upper respiratory airway. A lab member who was a snorer volunteered to have his breathing sounds recorded for the whole night while he received PSG monitoring simultaneously. A condenser microphone (Superlux CM H8C, Longmont, Colorado, USA) was placed at about 30 cm above the subject s mouth while he slept. The directivity of the microphone was set to the highest among three possible patterns. The recording was carried out in a quiet bedroom at a local elderly-care center. The acoustic recording was instantly sampled and stored in a laptop computer using the Audition software (Adobe, San Jose, CA, USA). Data Analysis Analysis of PSG data after the recording revealed that 10 episodes of OSA occurred during that night. Near the time when these OSA episodes occurred, a total of 37 breathing cycles of the sounds of snoring were extracted. To compare against, we also cropped 39 cycles of simple snoring (non-osa) that occurred elsewhere in the overnight recording. For the audio clips of snoring, the signal was segmented into 50-ms frames so the CSAs were estimated frame by frame along the upper respiratory pathway. The estimation of CSAs was achieved via the previously mentioned methods (ARMA modeling followed by the Levinson-Durbin recursion). Other relevant parameters for signal processing are as follows: Sampling rate = 8 khz, and the order of AR coefficients are L = 4, M = 6, and N = 5 (see Fig. 4). Proceedings of Meetings on Acoustics, Vol. 19, (2013) Page 5

6 (a) (b) Figure 5: Cross-section area (CSA) estimation during snoring. (a) An example of snoring before the occurrence of an OSA episode. (b) An example of simple snoring. The panels on top show the radius of the segment that is closest to the pharynx in the nasal branch. Note that in (a) between 0.5 and 2.0 second, the radius is significantly shorter than elsewhere in time, indicating that this segment was obstructed. In contrast, obstruction is not as prominent in (b). DISCUSSION AND CONCLUSIONS Fig. 5 shows the variation of CSA in time at the segment in the nasal branch that is closest to the three-port joint in Fig. 4. More precisely, what is shown is the corresponding radius of a cylinder that would have the same CSA. Although the CSA of the entire upper respiratory airway could be estimated, the CSA is only shown for this segment nearest to the pharynx, for the location is also the most probable obstruction site based on clinical experiences (R.P.Y. Chiang). By inspection, it is distinguishable that the snoring before an OSA episode tends to contain of a short period [e.g., sec in Fig. 5(a), and possibly also near 5.0 sec] of obstruction during which the radius of that segment was consistently below 0.5 cm [Fig. 5(a)]. In contrast, this narrowing effect was not as obvious in Fig. 5(b) as in Fig. 5(a). Therefore, this figure demonstrates that it is possible to predict an OSA episode automatically via computer listening of snoring. However, at this point we found that the CSA of this segment in the nasal cavity of the model is not consistently smaller near an OSA episode than far away from any OSA episodes. If we define constriction as an instance of CSA dropping below (0.5) cm for more than 1.0 sec, constriction would be identified at the chosen nasalpharyngeal segment for 26 out of the 37 clips of OSA-related snoring. However, 16 out 39 simple-snore clips also contain similar instances of constriction. To sum up, the conditional probability for constriction to be detected at the nasal-pharyngeal segment was 70% for snoring near an OSA episode vs. 41% for simple snoring. Based on the results of CSA estimation, we have yet to examine whether constriction happens in other possible obstruction site. More comprehensive data analysis is warranted. We also noticed that our current assumption regarding the snore generation mechanism (as shown in Figs. 3 and 4) may need to be refined because the model only considers plane-wave propagation in the upper respiratory airway. While this plane-wave assumption has led to useful speech modeling methods in the past such as linear-prediction coding (LPC), the generation of snoring sounds involves more complicated fluid dynamics such as turbulence and interaction between waves and tissues. These research directions are worth pursuing in the future. Nevertheless, the present results demonstrate that it is computationally possible to obtain estimates of the shape of the respiratory airway merely from recordings of the sounds of snoring. Though the numerical methods involved in solving this sound-to-shape mapping problem are subject to further scrutiny, the idea certainly looks promising thus far. Proceedings of Meetings on Acoustics, Vol. 19, (2013) Page 6

7 ACKNOWLEDGMENTS This research is partially sponsored by National Tsing Hua University. The authors also thank Center of Innovation and Synergy for Intelligent Home and Living Technology (INSIGHT) at National Taiwan University for facilitating many other related research projects. REFERENCES Agrawal, S., et al. (2002). Sound frequency analysis and the site of snoring in natural and induced sleep, Clin. Otolaryngol. Allied Sci. 27(3): Croft C.B., and Pringle, M.B. (1991). Sleep nasendoscopy: a technique of assessment in snoring and obstructive sleep apnoea, Clin. Otolaryngol. 16: Gibson, C., and Haykin, S. (1980). Learning characteristics of adaptive lattice filtering algorithms, IEEE Trans. Acoustics, Speech and Signal Processing, 28(6), Huang, H.-K. (2012). Estimation of oral and nasal cavity cross-section area during snoring by linear prediction methods. M.S. thesis, Dept. Electrical Engineering, National Tsing Hua University. (in Chinese) Levinson, N. (1947). "The Wiener RMS error criterion in filter design and prediction." J. Math. Phys., 25, Miyazaki, S., et al. (1998). Acoustic analysis of snoring and the site of airway obstruction in sleep related respiratory disorders, Acta Otolaryngol. Suppl. 537: Quinn, S.J., Daly, N., and Ellis, P.D.M. (1995). Observation of the mechanism of snoring using sleep nasendoscopy, Clin. Otolaryngol. Allied Sci. 20(4): i Here, n is an integer-valued sample number, so [] denotes the n th sample of an underlying continuous-time signal (), and their relation is given by [] = (), where T is the sampling period. Proceedings of Meetings on Acoustics, Vol. 19, (2013) Page 7