Novel multiclass classification for home-based diagnosis of sleep apnea hypopnea syndrome D. Sánchez-Morillo a,⇑ , M.A. López-Gordo a , A. León b a Biomedical Engineering and Telemedicine Lab, School of Engineering, University of Cádiz, C/Chile, 1, 11003 Cádiz, Spain b Pulmonology and Allergy Unit, Puerta del Mar University Hospital, 11009 Cádiz, Spain article info Keywords: SAHS Overnight pulse-oximetry Sleep apnea Binary hierarchical classifier SpO2 Multiclass Multivariate abstract Currently, Sleep Apnea-Hypopnea Syndrome (SAHS) is accurately diagnosed in Sleep Units. In the last decade, in order to reduce the burden for health systems and the consequent impact in patients, several home-located methods based on the binary classification of SAHS using overnight pulse oximetry (SpO 2 ) have been proposed. Binary classifiers give rise to higher accuracies, but the cost of misclassifying leads to high penalizations in terms of either health care costs or risks in patient’s health. This study presents a novel hierarchical classification scheme for the four-class SAHS diagnosis using a set of features extracted from SpO 2 and reported in specialized literature. An accuracy of 82.6% was achieved in the assessment of the four-class classification. The proposed method could be useful in the diagnosis of SAHS in an ambulatory home-based setting and could alleviate under-diagnosis rate and the waiting lists in sleep units. Ó 2013 Elsevier Ltd. All rights reserved. 1. Introduction The Sleep Apnea-Hypopnea Syndrome (SAHS) is a sleep-related breathing disorder characterized by repetitive reductions or cessa- tions of respiration. These episodes of reduction or absence of respiratory airflow induce the presence of characteristic modifica- tions in the oxygen saturation profile. The growing relevance of SAHS is a consequence of its raising prevalence, social repercussion and remarkable cardiovascular consequences (Leger, Bayon, Laaban, & Philip, 2012; Marin et al., 2012). The prevalence of SAHS in industrialized countries concerns 3–7% of adult men and 2–5% of adult women (Punjabi, 2008). The early diagnosis and treatment reduces the associated risk factors such as excessive daytime sleepiness, loss of concentration and performance and risk for cardio-vascular diseases (Leger et al., 2012). But it is assumed that more than 80% of women and 90% of men with moderate to severe obstructive sleep apnea may be still undiagnosed (Pang, Oto, Orl, & Terris, 2006). Diagnosis and assessment of the severity of SAHS require stud- ies beyond the clinical exploration. The American Sleep Disorders Association (ASDA) has established four levels for studies, accord- ing their complexity: level I: standard polysomnography (PSG); le- vel II: portable PSG; level III, respiratory polygraphy (PGR) and level IV: continuous monitoring of one or two parameters (Ferber et al., 1994). However, gold standard for SAHS diagnosis is still overnight attended PSG and the related apnea-hypopnea index (AHI). Among the limitations associated to PSG are the high costs, the dedicated medical personnel attention and the limitation of diagnostic facilities (Flemons, Douglas, Kuna, Rodenstein, & Wheatley, 2004). Additionally, long waiting lists for diagnosis and patient’s inhibitions to undertake PSG must not be neglected. Consequently, research focused on alternative diagnostic methods that overcome some of these limitations has remarkably increased. New approaches for simplified SAHS detection have been com- monly based on the analysis of a reduced set of data. The overnight analysis of arterial blood oxygen saturation (SpO 2 ) data recorded from non-invasive nocturnal pulse oximetry is widely used and ac- cepted for SAHS screening (Netzer, Eliasson, Netzer, & Kristo, 2001). Pulse oximetry is applied in SAHS for the detection of the desaturations in SpO 2 caused by apnea and hypopnea events. Sig- nal processing techniques in the time (Lin, Yeh, Yen, Hsu, & Hang, 2009; Magalang et al., 2003) and frequency domain (Hua & Yu, 2007; Morillo, Gross, León, & Crespo, 2012; Zamarrón, Gude, Bar- cala, Rodriguez, & Romero, 2003) and non-linear approaches (Alva- rez, Hornero, Abásolo, Del Campo, & Zamarrón, 2006; Hornero, Alvarez, Abásolo, Del Campo, & Zamarrón, 2007; Morillo, Rojas, Crespo, León, & Gross, 2009) have been proposed for SAHS screen- ing purposed using SpO 2 recordings. In the last years, multivariate diagnostic models have demon- strated proficiency in SAHS screening (Alvarez, Hornero, Marcos, & Del Campo, 2010; Marcos et al., 2008; Morillo & Gross, 2013). 0957-4174/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.08.062 ⇑ Corresponding author. Address: Escuela Superior de Ingeniería, Dpto. de Ingeniería Automática, C/Chile, 1, CP 11003 Cádiz, Spain. Tel.: +34 956015709; fax: +34 956015237. E-mail address: daniel.morillo@uca.es (D. Sánchez-Morillo). Expert Systems with Applications 41 (2014) 1654–1662 Contents lists available at ScienceDirect Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa