SPECTRAL ENVELOPE ANALYSIS OF SNORING SIGNALS

Transcription

1 SPECTRAL ENVELOPE ANALYSIS OF SNORING SIGNALS Mustafa Çavuşoğlu, Mustafa Kamaşak 2, Tolga Çiloğlu 3,Yeşim Serinağaoğlu 3,Osman Eroğul 4 Max Planck Instıtute for Biological Cybernetics, High Field MR Center, 72076, Tuebingen,Germany 2 İstanbul Technical University, Comuter Engineering Deartment, 34469, İstanbul,Turkey 3 Middle East Technical University, Electrical and Electronics Engineering Deartment, 06530, Ankara 4 Gülhane Military Medical Hosital, Biomedical and Clinical Engineering Center 0608, Ankara ABSTRACT In recent years, several studies have shown the relationshi between snoring and obstructive slee anea syndrome (OSAS). Besides time domain analysis of snoring signal, the sectral features and shaes of snores can be used to discriminate simle snorers and OSAS atients. In this study, we roose a method to classify simle snorers and OSAS atients based on sectral enveloe estimation of snoring signals. The formant frequencies and corresonding bandwidths are comuted for both grou, and the variation and consistency of the formant distributions are comuted. A total of 400 snoring eisodes from 7 simle snorer and 7 OSAS atients were analyzed. Significant differences are found in the formant frequencies of both grous. The results are discussed from the view oint of atho-hysiological asect. KEY WORDS Sectral analysis, formant frequency, OSAS, snore. Introduction Snores are breathing sounds roduced during the slee. Breathing triggers mechanical oscillations of the tissues such as soft alate or tongue around the constriction due to the relaxation of the tissues in the uer airway. The American Slee Disorders Association (ASDA) defined snoring as Loud uer airways breathing, without anea or hyoventilation, caused by vibrations of the haryngeal tissues []. In the last 5 years, snoring roblem has entered the realm of clinical medicine. It is a revalent symtom, and about 50% of the adult oulation snores frequently [2, 3]. It has been reorted as a risk factor for the develoment of diseases such as ischemic brain infraction, systemic arterial hyertension, and coronary artery disease and slee disturbance [4]. Slee anea is the cessation of airflow to the lungs during slee for 0 sec. or more [5, 6]. It may result from either uer airway collase (obstructive slee anea) or lack of neural inut from the central nervous system to the diahragm (central slee anea) [6]. OSA is the more common form of slee anea. Common symtoms of OSA are fatigue, daytime sleeiness, coroner roblems, and systemic hyertension [5, 6]. It is usually associated with loud, heavy snoring [6]. In recent years, several studies have shown the relationshi between snoring and obstructive slee anea syndrome (OSAS). These studies have shown that the frequency domain analyses of snores can be used to discriminate simle snorers and OSAS atients [7]. Analyzing snoring sounds is very useful and romising tool in slee related breathing disorders. Time domain arameters of snoring sounds such as the snore rate (number of snoring eisodes er hour), intensity of the snoring and the regularity of the snoring eisodes showed that simle snorers and OSAS atients can be classified by using snoring sound analysis instead of olysomnograhic studies which requires sending the whole night in the hosital. Hill et. al. [8] roosed the use of acoustic crest factor of snoring eisodes (the ratio of eak amlitude to root mean square value) to distinguish alatal from nonalatal snore. Abeyratne et. al. [9] roosed the measure intra-snoreitch-jum to diagnose OSAS. Sola-Soler et. al. [0] studied the differences in sectral enveloes of simle and OSAS snores and suggested the standard deviation of formant frequencies as a criterion for distinction among simle snores and OSAS snores. Miyazaki et. al. [] found significant differences in fundamental frequency deending on the snore roduction site. Sence et. al. found a significant correlation between median frequency of snore and anea/hyonea index (AHI). Cavusoglu et. al. [2] develoed an interface system based on time domain characteristics to differentiate the simle snorers and OSAS atients as well as to determine the efficiency of alied snoring treatment objectively. Most of the studies are focused on the fundamental frequency of the snores but sectral enveloe estimation of the snoring signals also contains information about the ressure ulses and anatomical filtering involved in snore roduction [0]. Thus, the sectral analysis values, the formant tye structure and the shae of the sectrum hel to distinguish simle snorers and OSAS atients [3]. Polysomnograhy (PSG), erformed over a full night slee, is resently the gold standard for diagnosis of slee anea [4, 5]. Since PSG analysis requires sending the night at a hosital with many measurement electreodes connected to the body, it is time consuming

2 and uncomfortable for the atient. The motivation of this study was to use sectral enveloe analysis of snoring signals to differentiate the simle snorers and OSAS atients as an alternative to olysomnograhy. This analysis may also rovide information in localization of the snoring site such as alate or tongue. Our aroach in modeling snoring sounds is motivated by the well-known source-filter modeling framework of seech sounds. According to the source filter-model, seech is roduced by the excitation of the vocal tract (the acoustic filter) by an acoustic ressure wave due to the motion of vocal folds or due to a constriction created by the seech articulators. The resulting excitation signal has a mixed eriodic and noise content in varying roortions for different honemes. Source-filter model is also very suitable for modeling snoring sounds. The mechanism of snoring is similar to that of voiced seech. The main difference arises at the source art. The vibration of vocal folds is relaced by the vibration of relaxed haryngeal tissues and tongue. Snorer s relaxed tissues tend to come into contact when there is no airflow for examle, around the intake-outlet turnover instants of breathing. Following the start of an intake or outlet, the ressure difference across the contact causes the searation of contacting tissues. Once the tissues are searated, the ressure difference dros and the tissues come into contact again and the cycle continues in this manner. The location of the contact in the harynx, the tye of contacting tissues, the hysical roerties of the tissues deending on age and/or fat content affects the dynamics of this vibration. The excitation of the vocal and nasal tracts as a consequence of this vibration yields the li and nose radiation of the snoring sound. The formant frequencies are the resonant frequencies of the vocal tract filter its range deends on the shae of the resonant cavities. Since the source of the snoring sound differs in simle snorers and OSAS atients, it motivates to the investigation of how the formant frequencies and their bandwidths changes in simle snorers and OSAS atients. 2. Materials and Methods 2. Recording Set u and Database A Sennhiser ME 64 condenser microhone with a Hz ± 2.5 db frequency resonse was used for recording sounds. This microhone has a cardioid attern which hels to suress some of the echoes from the environment. It was laced 5 cm over the atient s head during slee. The signal was fed via a BNC cable to the Edirol UA-000 model multi-channel data acquisition system connected to a ersonal comuter via universal serial bus. The comuter was laced outside the sleeing room not to disturb the sleeing atients. The acquired signal was digitized at a samling frequency of 6 KHz with 6 bit resolution. A total of 400 snoring eisodes were analyzed obtained from 7 simle snorers and 7 OSAS atients equally. Snoring eisodes were selected by an automatic detector [5] and validated by a medical doctor. Anea/Hyonea Index (AHI) of simle snorers and OSAS atients are 4.29 and 39.2 resectively. 2.2 The Formant Tye Sectrum Structure of Snoring Sounds Unlike the highly non-stationary character of seech because of the continuous motion of the articulators (jaw, lis, tongue, and velum) along the vocal tract, snoring signal bears non-stationarity at a much lower level. Articulators kee their ositions unchanged most of the time during snoring. Sectral enveloe estimations of snoring signals are shown in Figure. The surious eaks with very low amlitude are rejected by 3dB threshold. Insecting the snoring intervals clearly indicates the highly stationary behavior of the signal. This is much like the roduction of sustained vowels. Figure shows that the greater art of the energy content is below 2500 Hz and the main comonents lie in the low frequency range, at about 50 Hz. The enveloe eaks also at the frequencies about 500, 800 Hz, and 2200 Hz. The sectrum shows a fundamental frequency and a formants tye structure. Power (db) frequency (kh Figure. Sectral enveloe estimations of snoring signals In the context of seech rocessing, formants are the frequencies around which the eaks of the sectral enveloe are observed. They are the resonant frequencies of the vocal tract filter. Each formant is characterized by frequency, bandwidth and amlitude level, and its range deends on the shae of the resonant cavities. The different conditions in which the subjects and the atients who snore can affect the formants range in the frequency sectrum. The formants of a snoring signal including 5 snoring eisodes is shown in Figure

3 determined by minimizing the linear rediction coefficients. The formant frequencies are comuted by finding the roots of the linear rediction olynomial obtained from LPC analysis. The rediction olynomial can be exressed as follows: A / 2 * ( ak z ( ck z )( ck z ) (4) k k where is the order of the LPC filter. Resonance frequencies are found using only the comlex roots of th rediction olynomial. The k comlex root of the is exressed in equation (5). A( Figure 2. The formants of a snoring signal including 5 snoring eisodes Automatic sectral enveloe estimation of the snores in the database shows that formants stay almost unchanged within and across the snoring intervals and the formant frequencies of snoring eisodes have almost common formants in some frequency range. 2.3 Formant Frequency Analysis We comute the formant frequencies based on linear rediction coefficients (LPC) of the snoring eisodes. These coefficients are comuted using auto-correlation method with Levinson-Durbin recursions. x( n) e( n) + a x( n i) () i In equation (), x(n) reresents the snoring signal and is the degree of linear rediction. In addition, a i and e(n) corresonds to the linear rediction coefficients and rediction error resectively. We can exress the signal in frequency domain as follows: X i ( E( + aiz X ( (2) i we can modify (2) in order to obtain the transfer function that is exressed by (3). H ( X ( E( i i a z i i A( (3) jw c k k ck e (5) The th k formant frequency and the corresonding bandwidth can be written as: wk fk 2πTs (6) ln ck bk πt (7) T s where is the samling eriod. In this study, the order of the linear rediction filter is selected as Results We investigate the inter- and intra-atient variation of formant frequencies and their bandwidths for simle snorers and OSAS atients. To visualize these variations, we have randomly selected 0 snoring eisodes from each of the 7 simle snorers and 7 OSAS atients and comute the formant frequencies. Figure 3 and Figure 4 shows the inter- and intra- atient variation of formant frequencies in simle snorers and OSAS atients resectively. Average formant frequencies and corresonding bandwidths are calculated for the snoring eisodes in the database. Table and Table 2 show the average values of formant frequencies and corresonding bandwidths resectively for simle snorers and for OSAS atients. s The sound signal X ( is generated by stimulating the H (, location of the roduction (uvula, soft alate, E ( H ( etc), with.the transfer function, is 475

4 Table 2 Average values of formant bandwidths for Simle snorers and OSAS atients Subject BW BW2 BW3 BW4 Simle snorer OSAS atient Discussion and Conclusion Figure 3. Inter- and intra- atient variation of formant frequencies in simle snorers Figure 4. Inter- and intra- atient variation of formant frequencies in OSAS atients Table Average values of formant frequencies for Simle snorers and OSAS atients Subject F F2 F3 F4 Simle snorer OSAS atient The roduction mechanism of snoring sound is similar to that of voiced seech; hence we can aly the well-known source-filter model to analyze the snoring sounds. Sectral enveloe analysis of snoring signals rovides an efficient way in simle snorer/osas classification which is very imortant both in the diagnostic and treatment or uvulo-alatoharyngolasty (UPPP) tye surgery rocess of atients. Filtering characteristics of uer airway and the location of ressure roduction are strongly related to the sectral enveloe. For examle, the shae of resonant cavities in the uer airway changes the range and value of formant frequencies [0]. In simle snorers (Figure 3), while the formant frequencies are consistent in different snoring eisodes of same individual, there may be some shifting in different atients. The magnitude of these shifting is less in the first formant frequency than those of other three formant frequencies. The intra-atient consistency of formant frequencies is highest in the first formant and lowest in the second formant. For each atient, the formant frequencies may vary. However, from figure 2, it can be seen that the formant variations are correlated among the atients. In OSAS atients (Figure 4), there is no considerable coincidence in formant frequency locations both interand intra atient snoring eisodes. The significantly lower variability of formant distributions in simle snorers than in OSAS atients can be exlained as follows: i) The source of the snoring sound (vibrating tissue in the uer airway) is in a multi-segmental structure in OSAS atients. ii) In OSAS atients, uer airway resistance is relatively lower comared to simle snorers. The reduction in the uer airway resistance is related to the site of snoring. In this study, we roose a method to discriminate the simle snorers and OSAS atients based on sectral enveloe estimation of snoring signals. The exerimental results are romising and show that the aroach given here can be used not only for source localization of 476

5 snoring but also for detection of the obstructive slee anea. References [] Thory.MJ The international classification of slee disorders: diagnostic and coding manual. Lawrence KS, ed. Allen Press Inc., USA 990; [2] Lugaresi.E, Cirignotta.F, Coccagna.C, PianaC., Some eidemio-logical data on snoring and cardioresiratory disturbances Slee 980; 3: [3] Norton.PG, Dunn.EV Snoring as a risk factor for disease: an eidemiological survey. Br. Med J 985; 29: [4] K. Wilson, R.A Stoohs, T.F. Mulrooney, L.J. Johnson, C. Guilleminault, Z. Huang, The snoring sectrum:acoustic assesment of snoring sound intensity in,39 individuals undergoing olysomnograhy Chest 999, vol 5: , [5] J.Lucas, J. Golish, G. Sleeer, and J.A.O Ryan, Home Resiratory Care. Englewood Cliffs, NJ: Prentice Hall, 988, ch. 6, [6] R.L. Wilkin, Clinical Assessment in Resiratory Care. St Louis, MO: Mosby, 985, [7] J.A. Fiz, J.Abad, R. Jane, M. Riera, M.A. Mananas, P. Caminal, D. Rodenstein, J. Morera, Acoustic Analayses of snoring sound in atients with simle snoring and OSA Euroean Resiration Journal, vol 9, , 996 [8] Hill P.D., Lee B.V, Osborne J.E., Osman E.Z., Palatal snoring identified acoustic crest factor analysis, Physiol. Meas , 999 [9] Abeyratne U R, Wakwella A S and Hukins C 2005 Pitch jum robability measueres for the analysis of snoring sounds in anea Physiol. Meas [0] J. Sola-Soler, R. Jane, J.A. Fiz, J. Morera, Sectral Enveloe Analysis in Snoring Signals from Simle Snorers and atients with Obstructive Slee Anea Proceeding of the 25 th International Conference of the IEEE EMBS, Cancun, Mexico, Setember [] Miyazaki S., Itasaka Y., Ishıkaea K., Togawa K, Acoustic analysis of snoring and the site of airway obstruction in slee related resiratory disorders, Acta Otolaryngol Sul(Stockh), 998, 537, 47-5 [2] Cavusoglu M, Kamasak M, Erogul O, Ciloglu T, Akcam T and Serinagaoglu Y SASA:A Software System for Slee and Snore Analysis Proc. of Int. Sym. on Health Informatics and informatics, Antalya2007 [3] Dalmasso F and Prota R 996 Snoring: analysis, measurement, clinical imlications and alications Eur. Resir. J [4]. C. Guilleminault, R. Stoohs, A. Clerk, M. Cetel, and P. Maistros, A cause of excessive daytime sleeiness-the The uer airway resistance syndrome, Chest, vol. 04,78-787,993 [5]. J.H. Peter, T. Podszus, P. von Wichert, Slee Related Disorders and Internal Diseases, New York: Sringer-Verlag 987,,