Contents lists available at ScienceDirect Medical Hypotheses journal homepage: www.elsevier.com/locate/mehy A deep learning-based decision support system for diagnosis of OSAS using PTT signals Seda Arslan Tuncer a, ⁎ , Beyza Akılotu a , Suat Toraman b a Department of Software Engineering, Firat University, Elazig 23119, Turkey b Department of Informatics, Firat University, Elazig 23119, Turkey ARTICLE INFO Keywords: OSAS Convolutional Neural Networks Deep learning PTT ABSTRACT Sleep disorders, which negatively affect an individual’s daily quality of life, are a common problem for most of society. The most dangerous sleep disorder is obstructive sleep apnea syndrome (OSAS), which manifests itself during sleep and can cause the sudden death of patients. Many important parameters related to the diagnosis and treatments of such sleep disorders are simultaneously examined. This process is exhausting and time-consuming for experts and also requires experience; thus, it can cause difference of opinion among experts. Because of this, automatic sleep staging systems have been designed. In this study, a decision support system was developed to determine OSAS patients. In the developed decision support system, unlike in the available published literature, patient and healthy individual classification was performed using only the Pulse Transition Time (PTT) para- meter rather than other parameters obtained from polysomnographic data like ECG (Electrocardiogram), EEG (Electroencephalography), carbon dioxide measurement and EMG (Electromyography). The suggested method can perform feature extraction from PTT signals by means of a deep-learning method. AlexNet and VGG-16, which are two Convolutional Neural Network (CNN) models, have been used for feature extraction. With the features obtained, patients and healthy individuals were classified by the Support Vector Machine (SVM) and the k-nearest neighbors (k-NN) algorithms. When the performance of the study was compared with other studies in published literature, it was seen that satisfactory results were obtained. Introduction Sleep apnea is a significant disease defined as respiratory standstill during sleep, which is caused by snoring. It can occur many times during the night. During apnea, the muscles that allow the upper re- spiratory tract to open, relax. As a result of that, the base of the tongue, the palate or excessively enlarged tonsils block the airway; patients cannot breathe for at least 10 s. Fig. 1 shows the comparison of re- spiration between healthy individuals and patients with apnea [1]. There are three sleep apneas: Obstructive Sleep Apnea Syndrome (OSAS), Central Sleep Apnea (CSA) and Mixed Sleep Apnea (MSA). In terms of prevalence, the most common one is OSAS at approximately 84%. OSAS occurs when the muscles in the throat relax and the throat is blocked preventing air circulation [2–4]. The Polysomnography test (PSG) is used to identify OSAS. PSG records the brain activity and re- spiratory incidences of the patient throughout the night. It is based on the measurement of brain waves, eye movements, air flow from the mouth and nose, snoring, heart rate, leg movements and oxygen levels. One of the most important symptoms of sleep apnea is respiratory standstill during sleep. The diagnosis of the disease is clinically ex- amined using ECG, EEG and EMG signals and carbon dioxide mea- surements obtained from the patient [5]. OSAS can be determined as a result of examining the signals by identifying many parameters, such as how many times and how long respiration stops during sleep, how oxygen values and heart rate are affected and whether the patient falls into a deep sleep [6–17]. Motivation Individuals who experience sleep apnea symptoms can be faced with many serious problems during the day. These include sudden death during sleep, strokes, heart attacks and coronary failure, in- sufficient respiration in lung patients and uncontrollable diabetes. Thus, for the determination of OSAS disease, the systems supported with machine learning algorithms are necessary to help doctors during the diagnostic process. This study developed a decision support system that classifies individuals using the pulse transition time (PTT) obtained from the PSG data of OSAS patients and healthy individuals. AlexNet https://doi.org/10.1016/j.mehy.2019.03.026 Received 26 January 2019; Accepted 26 March 2019 ⁎ Corresponding author. E-mail addresses: satuncer@firat.edu.tr (S. Arslan Tuncer), storaman@firat.edu.tr (S. Toraman). Medical Hypotheses 127 (2019) 15–22 0306-9877/ © 2019 Elsevier Ltd. All rights reserved. T