Abstarct: Abstract Apnea has been considered one of the most challenging diseases for societies in recent years, affecting more than 939 million people worldwide, according to the World Health Organization. Apnea is a type of sleep disorder that occurs when a person's breathing stops during sleep. Diagnosing apnea in the early stages and treating it can have a direct impact on improving the quality of life of patients. So far, numerous studies have been conducted in this area, but most of these studies use classical classifiers for classification, which usually do not have good accuracy or are very sensitive to noise and other things that affect the electrocardiograph signal, which is not desirable and leads to reduced accuracy and poor performance. So the main challenge in this field is the appropriate approach to dealing with noise and sufficient accuracy in diagnosing apnea. To reduce errors and improve accuracy in the diagnosis of apnea, a new system is proposed to assist doctors in detecting obstructive sleep apnea using deep learning methods. In this innovation, the first step of this research is to preprocess the ECG signal using bandpass, wavelet, Savitzky-Golay and Hodrick-Proscott filters, which eliminate outliers, identify the position and extract the exact distance of the cardiac signal peaks. The use of efficient classifiers such as deep convolutional neural networks, k-nearest neighbors and perceptron neural networks in analyzing ECG signals related to apnea patients is expected to lead to good results. In this thesis, the proposed method is presented in the form of three approaches. In the first and second approaches, conventional filters and Savitzky-Golay filter are used for preprocessing, and in the third approach, a new filter called Hodrick-Proscott is used in addition. In the first and second approaches, a deep neural network with a simple structure is used, and in the third approach, a more complex structure is used. Implementation on the University College Dublin sleep apnea databaxse has resulted in an accuracy of 99.1% at best. In the publicly available Kegel databaxse, the best accuracy rate achieved is 95.6%. These results show an improvement of nearly 5% compared to previous research in this field, which is evidence of the claim that the new test has better performance in real-world problems.