| Summary: | 碩士 === 國立臺北科技大學 === 電機工程研究所 === 104 === Sleep Apnea and Hypopnea Syndrome (SAHS) have become an increasingly important public-health problem in recent years. Clinical studies have shown that SAHS can adversely affect patient’s cardiovascular system and cause behavior disorder. Moreover, up to 85%-90% of these cases are of obstructive sleep apnea (OSA). Therefore, the study of how to diagnose, detect and deal with OSA is becoming a significant issue. Polysomnography (PSG) is practically considered as the standard means for SAHS diagnosis. However, PSG is expensive and time-consuming since it requires SAHS patients to spend one night in a sleep laboratory with professional technicians and doctors. Accordingly, with the purpose of improving such inconvenience, there exists a huge demand to develop much simplified methods for diagnosing the OSA.
In this paper, we exploit electrocardiogram (ECG) derived respiratory signals to determine whether the occurrence of OSA. At first, 95 kinds of statistics, time domain, and frequency domain features are extracted. Then 11 machine-learning algorithms are applied to categorize SAHS events. Among them, the best three algorithms are picked out for further decision combination: they are AdaBoost with Decision Stump, Bagging with REPTree, and Bagging with ADTree. The experimental results show that the Majority Vote performs good behavior of sensitivity and specificity. It is able to achieve sensitivity at 67.39%, specificity at 69.38%, and accuracy at 68.92%, respectively. In case of including additional 35 types of oxygen saturation features, the performance can be further improved. It achieves accuracy at 78.14%, sensitivity at 84.35%, and specificity at 77.57%, respectively.
|
|---|
