Medical Engineering & Physics 34 (2012) 1213–1220 Contents lists available at SciVerse ScienceDirect Medical Engineering & Physics jou rnal h omepa g e: www.elsevier.com/locate/medengphy Multiclass classification of subjects with sleep apnoea–hypopnoea syndrome through snoring analysis Jordi Solà-Soler a,b,∗ , José Antonio Fiz c,b , José Morera c , Raimon Jané a,b a Department of ESAII, Universitat Politècnica de Catalunya, Barcelona, Spain b Institut de Bioenginyeria de Catalunya (IBEC), CIBER de Bioingeniería, Biomateriales y Nanomedicina (CIBER-BBN), Barcelona, Spain c Pulmonology Service, Hospital Universitari Germans Trias i Pujol, Badalona, Spain a r t i c l e i n f o Article history: Received 25 August 2011 Received in revised form 13 December 2011 Accepted 14 December 2011 Keywords: Snoring Sleep apnoea Bayes classifier Kernel density estimation a b s t r a c t The gold standard for diagnosing sleep apnoea–hypopnoea syndrome (SAHS) is polysomnography (PSG), an expensive, labour-intensive and time-consuming procedure. Accordingly, it would be very useful to have a screening method to allow early assessment of the severity of a subject, prior to his/her referral for PSG. Several differences have been reported between simple snorers and SAHS patients in the acoustic characteristics of snoring and its variability. In this paper, snores are fully characterised in the time domain, by their sound intensity and pitch, and in the frequency domain, by their formant frequencies and several shape and energy ratio measurements. We show that accurate multiclass classification of snoring subjects, with three levels of SAHS, can be achieved on the basis of acoustic analysis of snoring alone, without any requiring information on the duration or the number of apnoeas. Several classification methods are examined. The best of the approaches assessed is a Bayes model using a kernel density estimation method, although good results can also be obtained by a suitable combination of two binary logistic regression models. Multiclass snore-based classification allows early stratification of subjects according to their severity. This could be the basis of a single channel, snore-based screening procedure for SAHS. © 2011 IPEM. Published by Elsevier Ltd. All rights reserved. 1. Introduction Sleep apnoea–hypopnoea syndrome (SAHS) is a common dis- order the first symptom of which is usually heavy snoring. The impact of snoring ranges from no sleep disruption to continuously disrupted sleep [1]. The prevalence of SAHS is 3.2 times higher in snorers than in non-snorers [2]. Accordingly, a snoring analysis sys- tem could help to provide further indication of level of risk. The gold standard for diagnosing SAHS is polysomnography (PSG). This is a very expensive, labour-intensive and time-consuming procedure. It would be desirable to have a screening procedure that helped res- piratory physicians to rapidly determine the severity of a patient, in order to establish priority amongst candidates waiting for PSG. An ideal screening procedure should neither consider anyone with SAHS as healthy, nor send any healthy individual to the hospital for PSG. Recently, some authors have investigated the possibility of iden- tifying SAHS through the analysis of nocturnal oximetry [3] or oronasal airflow pressure [4]. Acoustic analysis of snoring reveals ∗ Corresponding author at: Institut de Bioenginyeria de Catalunya (IBEC), Baldiri Reixac, 4, Torre I, 9th Floor, 08028 Barcelona, Spain. Tel.: +34 934137358. E-mail addresses: jordi.sola@upc.edu (J. Solà-Soler), jafiz@msn.com (J.A. Fiz), josepmorera.germanstrias@gencat.cat (J. Morera), raimon.jane@upc.edu (R. Jané). information relating to the site and degree of obstruction of the upper airway [5]. Several studies have found statistically signifi- cant differences in the acoustic characteristics of snoring between patients with SAHS and simple snorers [6–10]. Most of these stud- ies have classified snoring individuals into two classes by means of an apnoea–hypopnoea index (AHI) threshold. However, no fur- ther information about the severity of the subject is provided. A recent publication of our group has described multiclass analysis of snoring subjects with SAHS [11]. Other authors have used a Bayes classifier with Gaussian density estimation to characterise individ- uals according to features of snoring and apnoea [12], but in general these variables do not follow a normal distribution. Our approach is based on a single channel, namely the sound signal, and in particular we exclusively use the acoustic informa- tion extracted from snores, without knowing the number or the duration of apnoeas. Good classification rates of snoring subjects with SAHS can be achieved with this tight constraint if (1) a deep analysis of snoring episodes is carried out, something that neces- sarily includes a wide range of snoring features and their variability, as we have shown in previous articles [13]; and (2) an automatic algorithm is used for the selection of the best set of features, for a given performance measure. In a preliminary study we analysed a Bayes classifier with a ker- nel density estimation method, using a range of snoring features. In this paper, we analyse the performance of this classifier when it 1350-4533/$ – see front matter © 2011 IPEM. Published by Elsevier Ltd. All rights reserved. doi:10.1016/j.medengphy.2011.12.008