A new method based on single channel ECG and hybrid machine learning for obstructive sleep apnea diagnosis

Obstructive Sleep Apnea (OSA) is a respiratory arrest condition that occurs only during sleep. This disorder not only causes sleep to be interrupted but also causes many other ailments, including the risk of death. According to the criteria published in the guidelines of the American Academy of Sleep Medicine (AASM), the diagnosis of the disorder is practiced in two stages, sleep staging and respiratory scoring with the help of the polysomnography (PSG) device. There is a cure for this disease after a diagnosis, but the fact that the diagnosis phase is quite difficult for the patient has led to the need for more comfortable and also reliable diagnostic methods. In this thesis, a new system that is easier to use and simpler than the PSG device has been developed for the reliable and comfortable diagnosis of OSA disease. Electrocardiography (ECG) signals of 10 patients were used in the study. The components of the ECG signal belonging to different frequency bands were derived, 225 features were extracted, the extracted features were selected and decreased, and classified by Fisher Feature Sorting Algorithm and Basic Component Analysis (PCA). In the study, four classifiers were used: Decision Tree (DT), Support Vector Machines (SVMs), k-Nearest Neighborhood Algorithm (kNN), and Ensemble Classifier. Classification performance was examined by the Leave-One-Out (LOO) method. In order to evaluate the classification performance, sensitivity, specificity and accuracy rate values were calculated and given. The features of each signal were selected and classified by the Fisher feature sorting algorithm. An increase of 10% in the success rate has occurred when the number of features selected were increased. Especially in the feature selection process of the ECG signal, the greatest success was achieved with an accuracy rate of 87.12% in sleep staging and 85.12% in respiratory scoring. At this stage, sensitivity and specificity for 10 features in sleep staging were 0.90 and 0.85, respectively, and sensitivity and specificity for 13 features in respiratory scoring were 0.85 and 0.86, respectively. These performance values are compatible with the literature and are quite high. The results obtained in the study showed that a system that can be used in practice, which will facilitate the diagnosis of OSA, can be developed.