Sleep Stage Identification Using The Combination of ELM and PSO Based on ECG Signal and HRV Tri Fennia Lesmana Computer Science Department, Binus Graduate Program – Master of Computer Science, Bina Nusantara University Jakarta, Indonesia 11480 e-mail: tfennia@gmail.com Sani Muhamad Isa Computer Science Department, Binus Graduate Program – Master of Computer Science, Bina Nusantara University Jakarta, Indonesia 11480 e-mail: sani.m.isa@binus.ac.id Nico Surantha Computer Science Department, Binus Graduate Program – Master of Computer Science, Bina Nusantara University Jakarta, Indonesia 11480 e-mail: nsurantha@binus.edu Abstract—The aim of this research was to build a classification model with an optimal accuracy to identify human sleep stages using Heart Rate Variability (HRV) features based on Electrocardiogram (ECG) signal. The proposed method is the combination of Extreme Learning Machine (ELM) and Particle Swarm Optimization (PSO) for feature selection and hidden node number determination. The combination of ELM and PSO produces mean of testing accuracy of 82.1%, 76.77%, 71.52%, and 62.66% for 2, 3, 4, and 6 number of classes respectively. This paper also provides comparison to ELM and Support Vector Machine (SVM) methods whose testing accuracy is lower than the combination of ELM and PSO. Based on the results, can be concluded that the addition of PSO method is able to increase classification performance. Keywords-component; sleep stages; HRV; ECG I. INTRODUCTION A good sleep quality is one of the most important activity of human being. It can be maintained by taking a sufficient amount of sleep according to the sleep requirement of each person and age. Based on the sleep survey of 1000 participants conducted by National Sleep Foundation [1], 13% of them had sleep less than needed on non-workdays. On workdays, the percentage of participants had sleep less than needed is 30%. This survey implies that there are still many poor sleep quality people, especially on workdays. A poor sleep quality effects are tiredness, anxious, loss of energy, fatigue, weariness, depression, sleep disorders, and death risk [2]. Sleep disorders such as sleep apnea, insomnia, and shift work sleep disorder can be minimized by doing some preventive actions. Sleep disorder preventive actions include early detection of sleep disorder and sleep quality monitoring which those actions can be conducted by firstly identify the sleep stages. Sleep stages of human consists of awake, Non-Rapid Eye Movement (NREM), and Rapid Eye Movement (REM). NREM itself consists of light sleep (which is divided again into stage 1 and stage 2) and deep sleep (which is divided again into stage 3 and stage 4). Sleep stage identification usually can be done by going to the sleep specialist. The patient should come to sleep in the provided room with the associated devices. Common devices used are Electroencephalographic (EEG) for measuring brain wave and Electrocardiogram (ECG) for measuring heart rate. Device used in this paper is ECG because the installation to the patient is more comfortable than EEG. Those devices give the polysomnographic database as the result. After recording, sleep specialist then identifies the sleep stage manually, thoroughly, and carefully of each data generated. However, this kind of work is very exhausting. It will be very helpful if the sleep stage can be identified automatically as the initial diagnosis. Figure 1. ECG signal. Automatic sleep stage identification as initial diagnosis model is built up using machine learning technique. One of the most popular machine learning methods is neural network. Neural network is highly suitable for sleep stage analysis. Its fault tolerance is suitable in sleep research to overcome some undesirable events that blur the data [3]. Neural network is suitable in biomedical signal analysis because of its ability to learn complex and non-linear relation [4]. Despite of those advantages, neural network with backpropagation learning method is slow and can lead to local optima solution [5]. Extreme Learning Machine (ELM) comes to overcome the weakness of backpropagation neural network (BPNN). ELM is based on neural network concept which only have one input layer, one hidden layer, and one output layer. Previous research shows that ELM gives the best accuracy with processing time 1180 and 809 times faster than SVM and BPNN respectively [6]. ELM has a better generalization ability with faster processing time than BPNN. Another popular method is Support Vector Machine (SVM). Basic concept of SVM is to find an optimal 258 2018 3rd International Conference on Computer and Communication Systems 978-1-5386-6350-9/18/$31.00 ©2018 IEEE