Received October 17, 2020, accepted November 1, 2020, date of publication November 5, 2020, date of current version November 16, 2020. Digital Object Identifier 10.1109/ACCESS.2020.3036024 Classifier Precision Analysis for Sleep Apnea Detection Using ECG Signals NUNO POMBO 1 , (Member, IEEE), BRUNO M. C. SILVA 1,2 , ANDRÉ MIGUEL PINHO 3 , AND NUNO GARCIA 1 1 Instituto de Telecomunicações, Universidade da Beira Interior, 6200-001 Covilhã, Portugal 2 IADE, Universidade Europeia, 1200-649 Lisboa, Portugal 3 Department of Informatics, Universidade da Beira Interior, 6200-001 Covilhã, Portugal Corresponding author: Nuno Pombo (ngpombo@di.ubi.pt) This work was supported by the FCT/MCTES through National Funds and When Applicable Co-Funded EU Funds under Project UIDB/EEA/50008/2020. ABSTRACT This article presents a study on the efficiency of implementing classifiers for the detection of sleep apnea moments based on a minute-to-minute Electrocardiogram (ECG) signal, detailing the comparison of the accuracy for different classifiers. At each ECG signal, a Sgolay filter was applied to extract the Heart Rate Variability (HRV) and the ECG-Derived Respiration (EDR) and they were used for the training, testing and validation of the classifiers. The same features were extended in a second phase in order to understand if all the classified features were important. According to the results obtained, the best accuracy was 82.12%, with a sensitivity and a specificity of 88.41% and 72.29%, respectively. This study shows the importance of choosing the right classifier for a specific problem as well as choosing and using the best features for a better accuracy. These promising early-stage results may lead to complementary studies to improve the classifiers for a possible real-world application. The performance of the proposed model was compared with other approaches used for the detection of sleep apnea. INDEX TERMS Sleep apnea, electrocardiogram, feature extraction, feature selection, artificial neural network, support vector machine. I. INTRODUCTION Sleep apnea is a clinical disorder characterized by cessation of breathing during sleep that can last seconds or even min- utes. Due to the fact that has direct effects on the cardiovas- cular system, such as systemic hypertension and sympathetic activity increment, it is considered an important cause for morbidity and mortality [1]. Since sleep is a key activity for each individual as it permits the human body to repair and maintain health [2], then is crucial to promote adequate clinical practices to mitigate its effects as evidenced when patients with sleep apnea who developed COVID-19 were considered at risk of great morbidity and mortality compared to other patients [3]. The gold-standard for sleep apnea diagnosis is the Polysomnography (PSG) that aggregates data collected from a myriad of body functions, such as: heart rhythm, eye movement, brain activity, and muscle activity, among others. However, this multi-parametric concurrent recording of The associate editor coordinating the review of this manuscript and approving it for publication was Sotirios Goudos . physiologic data, limits its adoption. Indeed, this is a com- plex, cumbersome, and time-consuming activity because it requires an exhaustive test in a controlled environment; like an hospital setting, to monitor the patient’s sleep, hence this diagnosis is both unfeasible for a large population and extremely expensive. So, is timely and promising the intro- duction of surrogate techniques that may be not only com- fortably applied to the patient but also a low-cost and simpler solution. The literature in alternative models to PSG are very abundant, namely related with proposals based on either a reduced set of signals [4]–[8] or a combination of signals [9]. Thus, in this study we demonstrate a comprehensive bench- mark of different classifiers and selected features based on a single signal, the Electrocardiogram (ECG). In line with this, four different classifiers to detect sleep apnea from ECG data were evaluated. In addition, these classifiers were tested on three different scenarios using distinct features (also extracted from the signal). The proposed methods could provide prac- titioners with a robust, simple and cost-efficient diagnosis tool compared with the classical screening schemes provided by PSG. VOLUME 8, 2020 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ 200477