| Summary: | 碩士 === 國立成功大學 === 電機工程學系 === 107 === This thesis aims to develop a home prescreening system based on a sleep questionnaire system and physiological signals of smartwatches for sleep apnea detection. The sleep questionnaire system is implemented by an application program (APP) running in portable devices, such as smartphones or pads. The APP contains five sleep-related questionnaires used for clinical evaluation, while the smartwatch contains a photoplethysmography (PPG) sensor, a blood oxygen saturation (SpO2) sensor and physiological signal analysis algorithms. Two types of home prescreening systems for sleep apnea have been developed: the first type is to use only the questionnaires filled out by subjects for the severity of sleep apnea evaluation, and the second type is to combine the questionnaires and the physiological signal analysis of the smartwatch to assess the severity of sleep apnea. In the first type of home prescreening systems, the total of 1,746 patients’ sleep questionnaires was collected from the sleep center of National Cheng Kung University Hospital, and the questionnaires had completed before the patient undertook a full night of polysomnography (PSG) sleep study. The values of AHI provided by the hospital were regarded as the gold standard of our system training. Before system training, we first removed the patients with missing data in the questionnaires, and then selected significant input features by a fast correlation-based filter. Using the selected features, three classifiers including support vector machines (SVM), gradient boosting decision trees (GBDT), and backpropagation neural networks (BPNN) were trained to detect the severity of apnea-hypopnea index (AHI). The results showed that the best classifier was BPNN which reached the average accuracy, sensitivity, and specificity were 74.4%, 77.3%, and 71.1%, respectively, with a 5-fold cross validation. The second type of home prescreening systems combined the sleep questionnaires with the results obtained by the physiological signal analysis algorithm of the smartwatch for sleep apnea detection. The physiological signal analysis algorithm first removed the noise or artifacts by a filter, and then detected the P waves from filtered PPG signals. Then, the algorithm calculated the intervals between two adjacent P waves, denoted as PPIs, and then we segmented PPIs combined with the blood oxygen saturation signals to form the training patterns of classifiers. In this study, three classifiers were implemented with performance comparisons: a SVM, a GBDT and a deep convolutional and long short-term memory network (DeepConvLSTM). The classifiers were trained to classify each segment whether it occurred sleep apnea event. The total of 38 patients was recruited in this study. They were arranged to stay at the sleep center for one night to collect the PSG data and the physiological signals of the smartwatches simultaneously during their sleep. The sleep events analyzed by registered sleep technologists were served the gold standard and a 5-fold cross validation was employed to validate the proposed classifiers. The results showed that the DeepConvLSTM classifier reached the best performance, and the average accuracy for sleep apnea event detection was 81.4%. The results have successfully validated the effectiveness of the proposed system as a home prescreening system for sleep apnea detection. In the future, we hope this system become a convenient and affordable tool that helps sleep apnea patients and provides valuable information to the doctor for arranging the hospitalization of severe patients with higher priority.
|
|---|
