Detection and Prediction of Obstructive Sleep Apnea Based on Traditional Machine Learning and Recent Deep Learning Architectures

• Main Authors: Lin, Yu-Zhe, 林裕哲
• Other Authors: Yu, Sung-Nien
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/f6kfmz

碩士 === 國立中正大學 === 電機工程研究所 === 106 === In this thesis, we proposed the identification and prediction methods of obstructive sleep apnea (OSA), using traditional machine learning and recent deep learning approaches. The human’s physiological signal Electrocardiogram (ECG) was used to identify and predict the occurrence of OSA. The differences of using different architectures were compared. This study is composed of three parts, the first part is the traditional machine learning (ML) identification. The architecture can be divided into signal processing, feature extraction, feature normalization and classification. The features include time domain, frequency domain, EDR (ECG-derived respiratory) waveform and other efficient features such as focusing on the spectra of the baseline wander. The support vector machine was employed in this study to be classifier. The effectiveness of different kinds of feature selector were discussed. The second part is to explore the use of deep learning (DL) architecture. Convolutional Neural Network (CNN) was employed in this study. Both ECG and its Fourier spectrum were used as inputs to and the effects were studied. Both architectures used the Leave-One-Person-Out (LOPO) cross-validation to testify the performance of the classifiers. The third part is about OSA prediction. The ECG signals within 1 to 4 minutes before the occurrence of OSA were analyzed the performance of OSA occurrence prediction by using different system architectures were compared. The Five-Fold cross-validation was employed to test the performance of the methods. The results show that the recognition rate of the first part (Traditional ML) can reach 87.27%, and the second part (Recent DL) which used the time and spectra signals achieved an accuracy of 90.14%. When the Traditional ML and the Recent DL were used in the third part, the best OSA occurrence prediction rate were 71.25% and 96.58% , respectively. Both traditional and deep learning architectures were published in the literatures to identify of sleep apnea and the recognition rate ranged from 82% to 89%. Very few papers were devoted to OSA prediction. Comparatively, this study proposed novel features and deep learning architectures to identify and predict the occurrence of OSA. More reliable cross validation methods were used to testify the performance of classifier. The results showed that the identification rate is significantly higher than the relevant papers and the prediction accuracy is high, which demonstrated the superiority of the proposed identification and prediction architectures in our study. With the impressive results in the identification and prediction of OSA, we ported the model trained by the deep learning frameworks into a real-time OSA detection system. The result demonstrate the possibility of using this approach for medical purpose which to assist the clinical professional in treating OSA an important auxiliary tool.